Method for the determination of a laundry weight in a laundry treatment appliance

Abstract

A method for the determination of a laundry weight in a laundry treatment appliance. The method includes selecting a laundry program in the laundry treatment appliance; starting the selected laundry program; sensing a plurality of parameters indicating operating conditions of the laundry treatment appliance during the laundry program; and predicating a weight of the laundry present within the laundry treatment appliance based on said plurality of parameters by means of a data-driven soft sensor.

Claims

1. A method for determining a laundry weight in a laundry treatment appliance, the method comprising: selecting a laundry program in the laundry treatment appliance; starting the selected laundry program; sensing a plurality of parameters indicating operating conditions of the laundry treatment appliance during the laundry program and one or more characteristics of the selected laundry program; and predicting a weight of the laundry present within the laundry treatment appliance based on the plurality of parameters by means of a data-driven soft sensor, wherein the predicting the weight includes using a supervised learning prediction.

2. The method according claim 1, wherein the laundry treatment appliance includes a rotatable drum and a motor to rotate the drum, and wherein the plurality of parameters includes one or more parameters indicative of operating conditions of the motor of the laundry treatment appliance.

3. The method according to claim 1, wherein the step of predicting a weight of the laundry by a supervised learning prediction includes predicting a weight of the laundry using a regression algorithm.

4. The method according to claim 3, wherein the step of predicting a weight of the laundry using a regression algorithm includes predicting a weight of the laundry using a regularized regression algorithm.

5. The method according to claim 4, wherein the step of predicting a weight of the laundry using a regularized regression algorithm includes predicting a weight of the laundry using a LASSO regression or a ridge regression.

6. The method according to claim 1, wherein the laundry treatment appliance is a horizontal axis or vertical axis laundry machine or a laundry washer-dryer including a rotatable drum and, after the step of starting the select laundry program, the method includes: performing a plurality of commutations of the drum, each commutation having an acceleration portion and a deceleration portion; and loading water into the drum; and the step of sensing a plurality of parameters includes: sensing a plurality of parameters indicating operating conditions of the laundry treatment appliance during one or more commutations.

7. The method according to claim 6, wherein the step of predicting a weight of the laundry present within the laundry treatment appliance based on the plurality of parameters by means of a data-driven soft sensor takes place before the step of loading water into the drum.

8. The method according to claim 6, wherein the step of sensing a plurality of parameters indicating operating conditions of the laundry treatment appliance during one or more commutations includes sensing the same parameters during the execution of each commutation of a plurality of commutations.

9. The method according to claim 6, wherein the prediction step based on the sensed parameters indicating operating conditions of the laundry during one or more commutations lasts less than a minute.

10. The method according to claim 1, wherein the laundry treatment appliance is a laundry dryer including a rotatable drum, and the step of predicting a weight of the laundry includes determining a weight of wet laundry and a weight of dry laundry.

11. The method according to claim 10, including, after the step of starting the laundry program, the step of blowing drying air into the drum, and the step of predicting a weight of the wet laundry is performed before the step of blowing dry air into the drum.

12. The method according to claim 1, including the step of: determining a duration of the selected laundry program or modifying a pre-set duration of the laundry program on the basis of the predicted weight of the laundry.

13. The method according to claim 1, including the step of: emitting a warning or a notice signal if the predicted weight of the laundry is above a pre-set threshold.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) The present invention will now be described with reference to the accompanying drawings that illustrate non-limiting embodiments thereof, wherein:

(2) FIG. 1 is a isometric view in section of a first embodiment of the laundry treatment appliance of the invention;

(3) FIG. 2 is a graph showing a parameter of the operating functions of the appliance of FIG. 1;

(4) FIG. 3 is a detail of the graph of FIG. 2;

(5) FIG. 4 is a schematic representation of one embodiment of the method of the invention;

(6) FIG. 5 is a schematic representation of a further embodiment of the method of the invention;

(7) FIG. 6 is a schematic representation of a further embodiment of the method of the invention;

(8) FIG. 7 is a confusion matrix of a Ridge Regression of the method of FIG. 6;

(9) FIG. 8 is a confusion matrix of a Lasso Regression of the method of FIG. 6;

(10) FIG. 9 is a classification rate histogram of the predictions of FIGS. 7 and 8;

(11) FIG. 10 is a isometric view in section of a second embodiment of the laundry treatment appliance of the invention;

(12) FIG. 11 is a graph showing a parameter of the operating functions of the appliance of FIG. 10;

(13) FIG. 12 is a schematic representation of a further embodiment of the method of the invention;

(14) FIG. 13 is a confusion matrix of an embodiment of the method of FIG. 11;

(15) FIG. 14 is a confusion matrix of another embodiment of the method of FIG. 12;

(16) FIGS. 15a and 15b are a schematic isometric view and lateral view in section of a third embodiment of the laundry treatment appliance of the invention;

(17) FIGS. 16a and 16b are a confusion Matrix and a prediction error histogram obtained in an embodiment of a method of the invention;

(18) FIGS. 17a and 17b are a confusion Matrix and a prediction error histogram obtained in an embodiment of a method of the invention;

(19) FIGS. 18a and 18b are a confusion Matrix and a prediction error histogram obtained in an embodiment of a method of the invention;

(20) FIGS. 19a and 19b are a confusion Matrix and a prediction error histogram obtained in an embodiment of a method of the invention;

(21) FIGS. 20a and 20b are a confusion Matrix and a prediction error histogram obtained in an embodiment of a method of the invention; and

(22) FIGS. 21a and 21b are a confusion Matrix and a prediction error histogram obtained in an embodiment of a method of the invention.

(23) As illustrated in FIG. 1, a vertical axis laundry washing machine 1 of the kind considered here comprises an outer casing 10 in the shape of substantially a parallelepiped. Most of a top wall of the casing 10 is occupied by a clothes loading aperture 11, wherein a lid (not shown) is hinged on along a rear side of the top wall. Further, on the top wall, a control panel 12 carrying the various control and display or indicating devices of the machine is preferably present. Within the casing 10, there are housed some compartments (not shown) intended for containing the chemical aids to be filled in for use in the washing phases of the operating cycles of the machine.

(24) The interior of the casing 10 accommodates an oscillating tub 18 which is essentially constituted by a cylindrical enclosure and two horizontal head pieces on the front and the rear side thereof, wherein said tub houses a single perforated drum 13 for loading and removing the clothes through the aperture 11. The drum 13, therefore, is of the type rotating about a vertical axis X, parallel to the height dimension of the casing 10, as driven by an electric motor 14 mounted below the tub 18.

(25) The mechanical action to wash the clothes inside the drum 13 is given by an agitator 15 which can move the laundry inside the drum in a clockwise or counter clockwise manner. Further, also the drum 13 can be driven in rotation in a clockwise or counter clockwise manner.

(26) Water can be loaded inside the drum 13 by means of tubes located in the upper portion of the casing 10 and the amount of water is regulated by the presence of one or more valves 16.

(27) The washing machine 1 further includes a control unit 17, schematically depicted in FIG. 1, where a processor, such as a microcontroller, is present, to control the functioning of the washing machine and storing program cycles to be selected by a user from the control panel 12. The control unit 17 stores also a proper algorithm in order to predict a weight of the laundry introduced inside the drum 13.

(28) Further, the washing machine includes a plurality of sensors (not visible in the drawings) apt to sense values or parameters indicating the operating conditions of the washing machine 1 during its functioning. These parameters may depend on the type of program selected by means of the control panel 17.

(29) In order to select the prediction algorithm for the weight prediction in a data driven soft sensor stored in the control unit 17 and of the parameters to be sensed and to be given as inputs in the prediction algorithm, a first and a second dataset of test data where for a plurality of parameters sensed by sensors in the washing machine the corresponding measured value of the weight of the laundry was available have been used. This first and second dataset have been obtain in field tests on the vertical axis washing machine.

(30) In any of the program cycles stored in the control unit, the cycle of the washing machine can be divided in a warm up phase where a plurality of commutations take place, followed by a water loading phase, further followed by a stroke phase. In the first training dataset the loading of water is performed in a “step by step” manner, that is, water is added in the drum several times in order to reach the total amount of water needed for washing, and these water loading sub-steps takes place between a plurality of subsequent strokes. In the second training dataset, the water is loaded in a single step at the end of the warm up phase and before the strokes phase starts.

(31) The prediction of the weight of the laundry introduced in the washing machine is considered to be in this embodiment a regression problem, that is, the outcome of the prediction algorithm is a number indicating the weight of the laundry, such as a number of kilograms.

(32) Both first and second dataset include values relating to operating conditions of the appliance during the execution of the commutations, and/or water loading and/or strokes.

(33) In FIG. 2, the velocity of the agitator 13 with a 5 kg load of laundry is shown during commutations and strokes, to visualize the behaviour of the drum during these phases. In the commutations, the velocity of the agitator and the drum is the same. These depicted data are a part of the first dataset, that is, water loading takes place in a plurality of subsequent steps between different strokes.

(34) In the first dataset, all parameters indicative of the operating conditions during the first commutation in the warm up phase and of the water loading phases are not considered as inputs for the algorithm in the data driven soft sensor.

(35) In the first dataset, the phases of warm up and strokes are considered separately, that is, the warm up phase and the stroke phase and the parameters collected during these different phases are considered to be different events.

(36) Many parameters of the operating conditions of the appliance during the execution of the commutations and/or strokes are available. It is therefore evaluated which parameters of the operating conditions of the appliance which are sensed during the commutations and/or during the stroke are valid proper input for the data sensor, so that only a sub-set of all available sensed parameters is used as input to the selected algorithm.

(37) The sensed parameters of the appliance considered during the warm up phase are the following. Each commutation has been considered as being divided in an acceleration portion, where the agitator and the drum accelerates to a reference speed, a subsequent substantially constant “high” speed portion around the reference speed, in a deceleration portion where the agitator and the drum decelerates till a different reference speed is reached and a substantially constant “low” speed portion after deceleration. These portions, or regions, are identified in FIG. 3.

(38) A first class of parameters of operating conditions of the appliance which are sensed during execution of commutations and which has been considered as possible input in the algorithm is a class of parameters relative to the torque of the motor during a commutation: 1) WU Time: Duration of the commutation; 2) WU Max Peak Drum-Speed Maximum speed during commutation. 3) WU OverShoot DrumSpeed Acc Difference between the maximum velocity and the reference velocity, in the acceleration interval. 4) WU OverShoot DrumSpeed Acc Time Time interval needed to reach the maximum velocity from the beginning of the commutation. 5) WU Min Peak Drum-Speed Minimum value of the speed in the deceleration phase. 6) WU OverShoot DrumSpeed Dec Difference between the minimum speed and the reference speed in the deceleration region. 7) WU OverShoot DrumSpeed Dec Time Time interval needed to reach the minimum speed from the beginning of the deceleration region. 8) WU Crossing Drum-Speed Acc Time Time interval after which the speed exceeds for the first time the reference speed in the acceleration region. 9) WU Crossing Drum-Speed Dec Time Time interval after which the speed lowers for the first time, the value of the reference speed from the beginning of the deceleration region. 10) WU Settling In Period SteadyHigh Time to reach the steady state at 1% with respect to the reference speed, in the constant speed region after the acceleration region. 11) WU Settling In Period SteadyLow Time within which the steady state at 1% with respect to the reference speed is reached, in the constant speed region from the beginning of the constant region after the acceleration region. 12) WU Crossing To Settling SteadyHigh Time Time interval between the first time the reference speed has been reached and the steady state point at 1%, in the constant region after the acceleration region. 13) WU Crossing To Settling SteadyLow Time Time interval between the first time the reference speed has been reached and the steady state point at 1%, in the constant region after the deceleration region. 14) WU Integral Drum-Torque Acc Sum of the torque values, in absolute value, till the maximum speed values has been reached, in the acceleration region. 15) WU Integral Drum-Torque Dec Sum of the torque values, in absolute value, till the maximum speed value has been reached, in the acceleration region. 16) WU Max High DrumTorque Maximum value of the torque, in absolute value, in the time interval including the acceleration region and the high steady region after acceleration. 17) WU Max Low Drum-Torque Maximum value of the torque, in absolute value, in the time interval including the deceleration region and the low steady region after deceleration. 18) WU Mean Acc DrumTorque Mean value of the torque, in absolute value, till the minimum value subsequent to WU MaxHighDrumTorque has been reached. 19) WU Mean SteadyHigh DrumTorque Mean value of the torque, in absolute value, from the minimum value subsequent to WU MaxHighDrumTorque till the constant high steady region after acceleration has been reached. 20) WU Mean Dec DrumTorque Mean value of the torque, in absolute value, till the minimum value subsequent to WU MaxLowDrumTorque has been reached. 21) WU Mean SteadyLow DrumTorque Mean value of the torque, in absolute value, from the minimum value subsequent to WU MaxLowDrumTorque till the constant low steady region after deceleration has been reached.

(39) Regarding the strokes, two different classes of sensed parameters of the operating conditions of the appliance during the execution of a stroke have been considered: the first class is relative to the velocity of the agitator during each stroke and the second one is relative to the torque of the motor during a stroke. The first class includes (Drum Speed indicates the speed of the agitator and not of the drum): 1) S PeriodTime Duration of the stroke. 2) S Max DrumSpeed Maximum value of the speed (of the agitator). 3) S Max DrumSpeed Time Time interval within which the speed reaches its maximum value. 4) S Crossing DrumSpeed Time Time interval after which the speed exceeds for the first time the reference speed. 5) S NumOut High DrumSpeed number of times in which the speed has a value above 5% the reference speed. 6) S NumOut Low DrumSpeed number of times in which the speed has a value below 5% the reference speed. 7) S Settling In Period DrumSpeed Time interval after which the speed does not vary of more than ±5% with respect of the reference speed, in the time interval starting from the first moment in which the reference speed has been reached til the last moment the reference speed has been reached. 8) S Mean DrumSpeed Mean value of the speed. 9) S Variance DrumSpeed Value of the variance of the speed.

(40) The second class includes: 10) S Max DrumTorque Maximum value of the torque. 11) S Max DrumTorque Time Time interval within which the speed reaches its maximum value. 12) S Mean Power Mean value of the scalar product between speed and torque for the whole duration of the stroke. 13) S Integral Power Sum of the absolute values of the scalar products of speed and torque for the whole duration of the stroke. 14) S Mean DrumTorque Mean of the absolute value of the torque for the whole duration of the stroke. 15) S Var DrumTorque Variance of the mean absolute value of the torque for the whole duration of the stroke. 16) S Integral Torque Sum of the absolute value of the torque for the whole duration of the stroke. 17) S STD Integral Torque Standard deviation of the torque. 18) S DrumTorque PeakTime Time interval needed to reach the maximum value of the torque. 19) S Mean DrumTorqueBefore Max Mean value of the torque before reaching the maximum. 20) S Mean DrumTorqueAfter Max Mean value of the torque after having reached the maximum. 21) S Interrupted Stroke Logical variable indicating whether the series of strokes has been interrupted.

(41) In order to determine whether all the parameters listed above are equally relevant in the prediction of the weight of the laundry and/or whether some of them can be disregarded, a Pearson correlation index has been calculated. In addition to this correlation index, which allows to determine which parameters are strongly correlated to each other among those listed, an analysis of the parameters to determine which ones have a more predominant difference among different load classes has been performed.

(42) From these two analyses, in the present embodiment of the invention, the following parameters have been considered as a proper input for the soft sensor in order to obtain a good prediction of the weight of the laundry: WU OverShootDrumSpeedAcc mean WU OverShootDrumSpeedDec mean WU MeanDecDrumTorque mean WU MaxDecDrumTorque mean SET 1
in the warm up phase (i.e. relating to the commutations) and S MeanPower S MeanDrumTorque S IntegralDrumTorque S IntegralPower S MeanDrumTorqueAfterPeak SET 2
for each stroke.

(43) After having trained the algorithm using the first dataset above mentioned, which in the present case a linear algorithm has been selected, and more specifically both a LASSO and ridge regressions algorithms have been used, a Monte Carlo cross validation of the prediction of the weight of the laundry has been performed.

(44) Two types of weight predictions have been made. A first weight prediction is a “fast weight prediction” performed as soon as the appliance has been switched on and the laundry program has been selected. This fast prediction uses only parameters relative to the commutations, that is only parameters relative to the operating conditions during the warm up phase. It has been found that this prediction can give a “cross grained prediction” of the weight of the laundry rather accurate if an error of 1 or 2 kilograms in the total weight is accepted, in a very short time interval. The second weight prediction takes into account also the parameters relative to the operating conditions of the appliance during the strokes. The second weight prediction method according to the invention is depicted in FIG. 4.

(45) In the fast prediction method, first the parameters relative of the operating condition of the laundry machine 1 during the commutations are collected. The same parameters during each commutation are sensed. The parameters preferably used are the four parameters as listed above in SET 1. Then some statistical variables are calculated on the parameters sensed, that is, the mean value, the variance, the maximum and the minimum of the parameters are obtained in order to have a single value for the same parameters obtained for the plurality of commutations. For example, 3 different commutations are taken into account and an average of the same parameters obtained for the 3 commutations are obtained.

(46) These averages are then used as input of the prediction algorithm in order to obtain a fast prediction of the weight of the laundry.

(47) In the “slower” weight prediction, in addition to what has been calculated in the fast prediction method, additional parameters are taken into account. The parameters of the operating conditions of the appliance during each stroke of a plurality is calculated, for example 47 consecutive strokes; the selected parameters are those above listed in the selected sub-set SET 2. The parameters of each stroke are used in the selected prediction algorithm to have a preliminary prediction of the weight of the laundry, and the prediction is performed by the linear algorithm (LASSO or ridge regression) using as inputs the parameters sensed during a specific stroke and the average of the parameters of the operating conditions during the warm up phase. A plurality of weight predictions is thus obtained (in FIG. 4 they are called internal classification), equal to the number of strokes considered. As a final prediction of the weight of the laundry, the most represented value is used (in FIG. 4 this is called external classification).

(48) Indeed, in this slower prediction method, after having trained the algorithm with the results of the first dataset, the obtained predictions are more accurate that the prediction obtainable in the “fast weight prediction”.

(49) The second dataset used to train the prediction algorithm includes also values relative to the water loading phase which is a single phase between the warm up phase and the stroke phase. Further, in addition to the parameters relative to the torque and the speed, further parameters have been sensed, which are relative to the amount of water inside the drum, values of the selected program by the user and values of timers present in the appliance.

(50) The following parameters relative to the operating conditions of the washing machine 1 during water loading have been sensed according to the second dataset in addition to those in the first dataset: 1) WL Torque mean torque. 2) WL LoadingTime Time interval during which the level of the water changes from zero to its maximum value. 3) WL EvWashTime Time of water loading with reference to the open time of the valve. 4) WL IntegralWater Sum of the values of the level of water during the whole duration of the water loading step.

(51) The following parameters have been sensed in addition to those outlined above with reference to the first dataset with respect to the operating conditions of the washing machine during the strokes: 1) S IsCrossing Logical variable that indicate whether the real speed reaches the reference speed before the end of the stroke. 2) S IsSettling Logical variable which indicates whether the real speed is around 5% of the reference speed before the end of the stroke. 3) SS IntegralWater Sum of the value of the level of water in the drum during the whole series of strokes. 4) SS NumLoadWater Time interval of possible water addition during the strokes.

(52) The parameters to be used as input in the algorithm of the soft sensor are, with reference to the commutations in the warm up phase, the same as in the first dataset (SET 1). Regarding the strokes, in addition to the parameters indicated in the first dataset SET 2, many additional variables relative to the torque or speed can be considered, such as: (a) WU MeanDecDrumTorque mean. (b) WU MeanAccDrumTorque mean. (c) WU MaxAccDrumTorque mean. (d) WU MaxDecDrumTorque mean. (e) WU IntegralDrumTorqueAcc mean. (f) WU OverShootDrumSpeedDec mean. (g) WU OverShootDrumSpeedAcc mean. (h) SS MeanDrumSpeed. (i) SS MaxDrumSpeedTime. (j) SS VarianceDrumSpeed. (k) SS IntegralDrumTorque. (l) SS IntegralPower. (m) SS MeanDrumTorque. (n) SS MeanDrumTorqueAfterPeak. (o) SS MeanPower. SET 3

(53) Regarding water loading the most relevant parameter is WL EvWashTime.

(54) The same correlation techniques described above to determine the preferred parameters to be sensed and used as input in the prediction algorithm have been used in order to reduce the parameters to be used as input to the algorithm to the lists above given.

(55) From now on, two diagrams will be specifically used in order to introduce the results: confusion matrix and prediction error histogram. In the field of machine learning, a confusion matrix, also known as a confusion, is a specific table layout that allows visualization of the performance of an algorithm, typically a supervised learning one.

(56) Each column of the matrix represents the instances in a predicted class while each row represents the instances in an actual (real) class (or vice-versa). The name stems from the fact that it makes it easy to see if the system is confusing two classes (i.e. commonly mislabelling one as another). In the main diagonal of the confusion matrix there is the exact match between predicted and real classes in percentage while other values represent the Classification Error (CE), i.e. the mismatch between predicted and real classes.

(57) In order to give a performance indicator on predictions realized, an index called Classification Rate CR will be provided with the corresponding confusion matrix; CR has been computed as
CR=[(total matches)*100]/(total comparisons)
where terms used are self-explaining.

(58) Two types of weight predictions have been made also using the training with the second dataset. A first weight prediction is a “fast weight prediction” performed as soon as the appliance has been switched on and the laundry program has been selected. The fast weight prediction preferably gives a prediction within 2 minutes from the beginning of the selected program in the machine. This fast prediction uses only parameters relative to the operative conditions of the appliance during commutations. Two types of fast prediction have been performed, as depicted in FIGS. 5 and 6.

(59) In FIG. 5, where the fast prediction is only a portion of the depicted method, the fast prediction takes into account only the results of parameters collected during the commutations. First, the parameters relative of the operating condition of the laundry machine 1 during the commutations are collected. The same parameters during each commutation are sensed. The parameters preferably used are the four of SET 1 as listed above. Then some statistical variables are calculated on the parameters sensed: the mean value, the variance, the maximum and the minimum of the parameters is obtained. For example, the same parameters in 3 different commutations are taken into account. At least one among: the mean value, the variance, the maximum and the minimum of each parameters is an input to the soft sensor algorithm.

(60) Alternatively, in an additional embodiment of the fast prediction, for the parameters sensed during each commutation, a calculation of the weight is obtained, as depicted in FIG. 6. Therefore, a number of weight predictions equal to the number of commutations considered is calculated (called in FIG. 6 internal classification). As a final prediction of the weight of the laundry, the most represented value is used (called in FIG. 6 external classification). It has been found that this prediction can give a “cross grained prediction” of the weight of the laundry rather accurate if an error of 1 or 2 kilograms in the total weight is accepted.

(61) The second weight prediction of this embodiment takes into account also the parameters relative to operating conditions of the appliance during the strokes and the water loading. This “slower” weight prediction preferably gives a prediction of the weight of the laundry within 5/6 minutes from the beginning of the selected program in the machine. The second weight prediction method according to the invention is depicted in FIG. 5. In this method, first the parameters relative of the operating condition of the laundry machine 1 during the commutations are collected. The same parameters during each commutation are sensed (SET 1). The parameters preferably used are 4 as listed above. Then some statistical variables are calculated on the parameters sensed: the mean value, the variance, the maximum and the minimum of the parameters is obtained. For example, the same parameters during 3 different commutations are taken into account.

(62) Further, the parameters of the operating conditions of the appliance during water loading (which is a single phase) are sensed.

(63) Further, the parameters of the operating conditions of the appliance during each stroke of a plurality is calculated, for example 47 consecutive strokes; the selected parameters are those above listed (SET 2 plus additional parameters of SET 3). The parameters of the series of strokes are then averaged and their mean values are obtained. Further, parameters relative to the overall series of strokes are determined as well. The parameters of the operative conditions of the appliance during water loading, the averaged parameters of the operative conditions of the appliance during the commutations and the averaged parameters of the operative conditions of the appliance during strokes are then used as inputs to a linear algorithm (LASSO or ridge regression).

(64) A very accurate prediction of the weight of the laundry is obtained with this model. Indeed, after having trained the algorithm with the results of the second dataset, the obtained predictions are more accurate that the prediction obtainable with the model trained on the first dataset.

(65) In FIGS. 7 and 8, confusions matrices where in the method of FIG. 6 the Ridge Regression and LASSO regression are used, respectively, are shown according to the above mentioned method which takes into account parameters of the operating conditions of the appliance during warm up, water loading and strokes. In the ridge regression, all classes below 6 kg are correctly predicted with a percentage above 65%, while this percentage decreases considering heavier loads. Good predictions are obtained with LASSO regression as well. In FIG. 9, a classification rate of the weight of the laundry previsions based on the above written method is shown, where—in addition to the LASSO regression and Ridge Regression—results obtained by a OLS (Ordinary least squares) regression are shown as well.

(66) In a second embodiment of the invention, the laundry treatment appliance is a horizontal axis washing machine depicted in FIG. 10.

(67) A horizontal axis laundry washing machine 2 comprises an outer casing 20 in the shape of substantially a parallelepiped. Most of a front wall of the casing 20 is occupied by a clothes loading aperture 21, which is openable and closable by a door 28 which is hinged on the front wall. A control panel 22 carrying the various control and display or indicating devices of the machine is preferably positioned in an upper portion of the front wall. Within the casing 20, there are housed some compartments (not shown) intended for containing the chemical aids to be filled in for use in the washing phases of the operating cycles of the machine.

(68) The interior of the casing 20 accommodates a tub 25 which is essentially constituted by a cylindrical enclosure and two vertical head pieces on the front and the rear side thereof, wherein said tub houses a single perforated drum 23 for loading and removing the clothes through the aperture 21. The drum 23, therefore, is of the type rotating about a horizontal axis Y, parallel to the width dimension of the casing 20, as driven by an electric motor 24 mounted below the tub 25.

(69) The mechanical action to wash the clothes inside the drum 23 is given the combined rotation of the drum and the effect of gravity. The drum 23 can rotate in a clockwise or counter clockwise manner.

(70) Water can be loaded inside the drum 23 by means of tubes and the amount of water is regulated by the presence of one or more valves 26.

(71) The washing machine 2 further includes a control unit 27, schematically depicted in FIG. 10, where a processor, such as a microcontroller, is present, to control the functioning of the washing machine and storing program cycles to be selected by a user from the control panel 22. The control unit 27 stores also a proper algorithm in order to predict a weight of the laundry introduced inside the drum 23.

(72) Further, the washing machine includes a plurality of sensors (not visible in the drawings) apt to sense values or parameters indicating the operating conditions of the washing machine 2 during its functioning.

(73) In order to select the prediction algorithm for the prediction in a data driven soft sensor stored in the control unit 27 and of the parameters to be sensed and to be given as inputs in the prediction algorithm, a third dataset of data available where a plurality of parameters that can be sensed by sensors in the washing machine and the corresponding measured value of the weight of the laundry have been used. This third dataset has been obtain in field tests on the washing machine.

(74) In any of the program cycles stored in the control unit, the cycle of the washing machine begins with a plurality of commutations before water loading. The commutations are, as detailed in the first embodiment of the vertical axis laundry machine, movements of the drum including an acceleration region and a deceleration region. The commutations may take place after a drain pump has been activated to remove remaining water in the tub. These commutations differ from the commutations in a vertical axis laundry machine because the rotations of the drum are along a horizontal axis and an agitator is not present. However, the definition of the commutation, having an acceleration region, a substantially constant region and a deceleration region remain the same and the visualization of such a commutation is as shown in FIG. 11. In this case, being no agitator present, the velocity of the drum is depicted in the figure.

(75) The prediction of the weight of the laundry introduced in the washing machine is considered to be a classification problem, that is, the outcome of the prediction algorithm is not a number, e.g., the weight of the laundry for example in kilograms, but it is a value which indicates to which class among a plurality of predefined classes, the weight of the laundry belong. For example, 6 load classes have been identified and preselected, the first class starts from a load equal to 0 kg and the last class reaches a maximum load of 10 kilograms. The classes are separated by a weight of 2 kilograms each (e.g. first class indicates a load of 0 kg, the second of 2 kg, and so on till the last class which indicates a load of 10 kg).

(76) Further, only a fast determination is considered in this embodiment, which is a determination of the weight of the laundry that takes less than a minute to be performed, preferably less than 30 seconds. The prediction of the weight of the laundry starts as soon as the selected program has been selected and begun, and a proper number of commutations has taken place.

(77) As in the previous embodiment, an algorithm has been selected, in this case a Logistic regression, which has been trained with the third dataset which has been provided. The third dataset includes field test of the horizontal axis washing machine.

(78) The cross validation of the model has been made using a Monte Carlo method.

(79) The parameters of the operating conditions of the washing machine considered to be input of the algorithm are parameters of the operating conditions of the appliance during the commutations of the drum only, so that a fast prediction can be made.

(80) The selected parameters of the operating conditions of the appliance 2 during the commutations which are selected to be the inputs of the algorithm are the following:

(81) Torque_Int: torque integral calculated during ramp up minus the torque average calculated at following highest reached speed (80 rpm)

(82) 0$\mathrm{Torque\_int}=\left[\left(\underset{i=1}{\overset{N}{.Math.}}{T}_i\right)-u\right].Math.\Delta t;$
where N is the number of torque samples during acceleration, T.sub.i is the torque value at instant i, u is the average torque calculated at following highest reached speed (80 rpm) step, Δ.sub.i is the sample time.

(83) Sample: number of samples used during ramp up for torque integral calculation (N)

(84) Energy: mechanical power integral (integral of torque minus the torque average calculated at first 80 rpm step*speed) calculated during ramp up

(85) $\mathrm{Energy}=\left[\underset{i=1}{\overset{N}{.Math.}}\left(T_i-u\right).Math.{\omega }_i\right].Math.\Delta t;$
where N is the number of torque samples during acceleration, T.sub.i is the torque value at instant i, u is the average torque calculated at first 80 rpm step, ω.sub.i is the drum speed at instant i, Δ.sub.i is the sample time.

(86) Torque_diff: absolute value of difference between Torque_Int of each commutation and the average Torque_Int value of the 4 commutations

(87) ${\mathrm{Torque\_diff}}_1={\mathrm{Torque\_int}}_1-\left(\underset{i=1}{\overset{4}{.Math.}}{\mathrm{Torque\_int}}_i\right);$
The parameters sensed during 4 subsequent commutations have preferably been used.

(88) The chosen algorithm is a Logistic Regression. It belongs to the class of Generalized Linear Models, a generalization of ordinary linear regression that allows predictions to have an error distribution models other than a normal distribution; this is achieved by exploiting a sigmoid function upon a linear combination of the input data x.

(89) In FIG. 12 a schematic representation of the method according to the present embodiment in order to predict the classification of the weight of the laundry in the horizontal axis washing machine is depicted.

(90) The problem in this embodiment is a multi-class classification problem, in which it is desired to separate K (in the specific embodiment K=6) different load levels (L1, L2, . . . , LK with L1<L2< . . . <LK). A framework where K−1 different logistic regression classifiers (f.sub.1, f.sub.2, . . . , f.sub.K-1) are constructed is set to assign a class for a new observation x, as follows:

(91) TABLE-US-00001 TABLE 1 Function Class 0 samples Class 1 samples f.sub.1 L2, L3, . . . , LK-1, LK L1 f.sub.2 L3, . . . , LK-1, LK L1, L2 . . . f.sub.K-2 LK-1, LK L1, L2, . . . , LK-2 f.sub.K-1 LK L1, L2, . . . , LK-2 , LK-1

(92) The K−1 classifiers are then exploited as follows to perform the classification: if f.sub.1(z)≥0.5 then the class L1 is assigned, otherwise f.sub.2 is tested, if f.sub.2(z)≥0.5 then the class L2 is assigned, otherwise f.sub.3 is tested, . . . if f.sub.K-2(z)≥0.5 then the class LK−2 is assigned, otherwise f.sub.K-2 is tested, if f.sub.K-1(z)≥0.5 then the class LK−1 is assigned, otherwise the class LK is tested.

(93) In the specific embodiment therefore, the input parameters are, considering the parameters taken during 4 different commutations:

(94) TABLE-US-00002 TABLE 2 Torque_Int = [x.sub.1 x.sub.2 x.sub.3 x.sub.4] Sample = [y.sub.l y.sub.2 y.sub.3 y.sub.4] Energy = [z.sub.1 z.sub.2 z.sub.3 z.sub.4] Torque_diff = [w.sub.1 w.sub.2 w.sub.3 w.sub.4]

(95) Considering that 6 classes to identify are present and 4 parameters, a matriz of 25 coefficients is to be built.

(96) $\hspace{1em}\left(\begin{array}{ccccc}{a}_1& a_2& a_3& a_4& a_5\\ b_1& b_2& b_3& b_4& b_5\\ c_1& c_2& c_3& c_4& c_5\\ d_1& d_2& d_3& d_4& d_5\\ e_1& e_2& e_3& e_4& e_5\end{array}\right)$

(97) Once defined the coefficients, the procedure described below, with reference to FIG. 12, is preferably to be followed.

(98) The first step is to define the class for each commutation.

(99) 1.sup.st Commutation Calculation
ƒ.sub.1=a.sub.1+a.sub.2.Math.x.sub.1+a.sub.3.Math.y.sub.1+a.sub.4.Math.z.sub.1+a.sub.5.Math.w.sub.1 if ƒ.sub.1>0 then class=0 Analyze 2.sup.nd commutation if ƒ.sub.1≤0 Analyze 2.sup.nd classifier ƒ.sub.2
ƒ.sub.2=b.sub.1+b.sub.2.Math.x.sub.1+b.sub.3.Math.y.sub.1+b.sub.4.Math.z.sub.1+b.sub.5.Math.w.sub.1 if ƒ.sub.2>0 then class=2 Analyze 2.sup.nd commutation if ƒ.sub.2≤0 Analyze 3.sup.rd classifier ƒ.sub.3
ƒ.sub.3=c.sub.1+c.sub.2.Math.x.sub.1+c.sub.3.Math.y.sub.1+c.sub.4.Math.z.sub.1+c.sub.5.Math.w.sub.1 if ƒ.sub.3>0 then class=4 Analyze 2.sup.nd commutation if ƒ.sub.3≤0 Analyze 4.sup.th classifier ƒ.sub.4
ƒ.sub.4=d.sub.1+d.sub.2.Math.x.sub.1+d.sub.3.Math.y.sub.1+d.sub.4.Math.z.sub.1+d.sub.5.Math.w.sub.1 if ƒ.sub.4>0 then class=6 Analyze 2.sup.nd commutation if ƒ.sub.4≤0 Analyze 5.sup.th classifier ƒ.sub.5
ƒ.sub.5=e.sub.1+e.sub.2.Math.x.sub.1+e.sub.3.Math.y.sub.1+e.sub.4.Math.z.sub.1+e.sub.5.Math.w.sub.1 if ƒ.sub.5>0 then class=8 Analyze 2.sup.nd commutation if ƒ.sub.5≤0 then class=10

(100) The described approach for the 1.sup.st commutation has to be followed for the other 3 commutations.

(101) Once defined the class for each commutation, a decision policy based on majority is a majority is applied to choose the final load amount level.

(102) The final class is the mode of the all class commutations. This means the final class is the most represented one. Furthermore if two classes are equal “winners”, the larger one is chosen.

(103) Several predictions have been made changing the number of considered commutations.

(104) In the table below are reported the classification rate performances, i.e., the amount of loads classified correctly over the total amount. These results have been achieved on a dataset of 103 tests and are reported as a mean over K=1000 Monte Carlo simulations. The results depends on how many commutations have been taken into account, that is, the parameters of operating conditions of the washing machine during 4, 5, 6 . . . . Up to 10 commutations.

(105) TABLE-US-00003 TABLE 3 Commutations allowed: 4 5 6 7 8 9 10 Classification Rate 97.74% 97.91% 98.76% 98.67% 98.94% 99.41% 98.79%

(106) FIGS. 13 and 14 show the corresponding confusion matrices for the 4 and 9 commutation cases.

(107) As clear from FIGS. 13 and 14, the method of the invention is capable of predicting the correct weight class with a high accuracy.

(108) FIGS. 15a and 15b shows a laundry dryer 3 according to an embodiment of the invention. Preferably, laundry dryer 3 comprises an outer box or casing 30, preferably but not necessarily parallelepiped-shaped, and a treating chamber, such as a drum 33, for example having the shape of a hollow cylinder, for housing the laundry and in general the clothes and garments to be dried. The drum 33 is preferably rotatably fixed to the casing 30, so that it can rotate around a preferably horizontal axis R (in alternative embodiments, rotation axis may be tilted). Access to the drum 33 is achieved for example via a door 34, preferably hinged to casing 30, which can open and close an opening 34a realized on the casing itself.

(109) Laundry dryer 3 also preferably comprises an electrical motor assembly 35 for rotating, on command, revolving drum 33 along its axis inside casing 30.

(110) Further, laundry dryer 3 may include an electronic central control unit 37 which controls both the electrical motor assembly 35 and other components of the dryer 3 to perform, on command, one of the user-selectable drying cycles preferably stored in the same central control unit. The programs as well other parameters of the laundry dryer 3, or alarm and warning functions can be set and/or visualized in a control panel 31, preferably realized in a top portion of the dryer 3, such as above door 34.

(111) Dryer 3 additionally includes a process air circuit which comprises the drum 33 and an air process conduit 38, depicted as a plurality of arrows showing the path flow of a process air stream through the dryer 1 (see FIGS. 15a and 15b). Process air circuit also includes a fan or blower 39.

(112) A dedicated motor can be coupled to the fan 39, but in a possible simpler implementation the same motor can operate the fan 39 and the drum 33 (in other words only one of the two motors can be present, such as motor 35).

(113) The dryer 3 of the invention additionally comprises a process air generator, in the depicted embodiment a heat pump system 40 including a first heat exchanger (called also condenser) 41 and a second heat exchanger (called also evaporator) 42. Heat pump 40 also includes a refrigerant closed circuit (partly depicted) in which a refrigerant fluid flows, when the dryer 3 is in operation, cools off and may condense in correspondence of the condenser 41, releasing heat, and warms up, in correspondence of the second heat exchanger (evaporator) 42, absorbing heat. A compressor 43 receives refrigerant in a gaseous state from the evaporator 42 and supplies the condenser 41, thereby closing the refrigerant cycle. In the following the heat exchangers are named either condenser and evaporator or first and second heat exchanger, respectively. More in detail, the heat pump circuit connects via piping the second heat exchanger (evaporator) 42 via a compressor 43 to the condenser 31. The outlet of condenser 41 is connected to the inlet of the evaporator 42 via an expansion device, such as a choke, a valve or a capillary tube.

(114) Compressor 43 is driven by an electric motor (not visible in the figures), preferably integrated with the compressor in the same housing. Preferably, the compressor is a variable speed compressor so that the compressing velocity can be modified.

(115) Preferably, in correspondence of evaporator 42, the laundry dryer 3 of the invention may include a condensed-water canister (also not visible) which collects the condensed water produced, when the dryer 3 is in operation, inside evaporator 42 by condensation of the surplus moisture in the process air stream arriving from the drying chamber (i.e. drum) 3. The canister is located at the bottom of the evaporator 42. Preferably, through a connecting pipe and a pump (not shown in the drawings), the collected water is sent in a reservoir located in correspondence of the highest portion of the dryer 3 so as to facilitate a comfortable manual discharge of the water by the user of the dryer 3.

(116) The condenser 41 and the evaporator 42 of the heat pump 40 are located in correspondence of the process air conduit 38 formed in a basement 324 of the casing 30.

(117) The control unit 37 includes a processor, such as a microcontroller, is present, to control the functioning of the dryer and storing program cycles to be selected by a user from the control panel 32. The control unit 37 stores also a proper algorithm in order to predict a weight of the laundry introduced inside the drum 33.

(118) Further, the washing machine includes a plurality of sensors (not visible in the drawings) apt to sense values or parameters indicating the operating conditions of the dryer 3 during its functioning.

(119) In order to select the prediction algorithm for the prediction in a data driven soft sensor stored in the control unit 37 and of the parameters to be sensed and to be given as inputs in the prediction algorithm, a fourth and a fifth dataset of data available where a plurality of parameters that can be sensed by sensors in the dryer and the corresponding measured value of the weight of the laundry have been used. These fourth and fifth dataset have been obtain in field tests on the dryer.

(120) The fourth dataset includes 304 tests and the fifth dataset includes 84 tests; they are different in terms of hardware setup and parameters sensed and used as inputs in the prediction algorithm. Another dataset has been studied separately in order to estimate only the initial moisture of the laundry introduced in the drum; it is called “Dataset synthetic” and the load “synthetic” for its tests is unique: 3.5 Kg.

(121) From data provided in the fourth and fifth datasets, Ydry (nominal weight, that is the weight of the wet laundry) is known, thus the equivalent wet load classes are determined using Ydry and rounding up the equality Ywet=Ydry+60% (Ydry). In this way, there are 7 classes (i-vii) and the 60% of initial moisture is taken into account to distinguish between Ywet classes according to common policy in dryers.

(122) The parameters used as inputs in the algorithm are in the fourth dataset relative to operating conditions of the motor, of the compressor of the heat pump and of the humidity sensor during the selected drying program. The fifth dataset includes only parameters indicating operating conditions of the humidity sensor.

(123) The Supervised Learning problem is solved using regression and in this perspective the essential input matrix is the Design matrix belonging to R.sup.N×p, where p is the number of predictors (features extracted from signals) and N is the total amount of tests available.

(124) Similarly, the output matrix Y is the concatenation of load or humidity observations; it will be indicated as Ywet, Ydry or Yhum depending on the output selected.

(125) The method of the invention implements estimation procedures that are deployable in firmware program, in this regard some simple features have been computed to summarize the entire information inherent in all signals provided. In the following Table the parameter chosen (115 in total) are described in details: practically speaking they are maximum or minimum values and relative positions, means or variances, slopes or integrals in different time intervals.

(126) The parameters sensed and used, either in the fourth dataset (where all parameters have been used) or in the fifth dataset (where only the parameters of the humidity sensor have been used) are the following: max1-power-noise-f-ha max value in [50,80]s ta-power-noise-f-ha descent time after 50 s tb-power-noise-f-ha rise time after 50 s initial-power-noise-f-ha value at tb final-power-noise-f-ha value at stop-time min2-power-noise-f-ha min value in [tb,stop-time]s max2-power-noise-f-ha max value in [tb,stop-time]s posmax2-power-noise-f-ha max value position in [tb,stop-time]s mean2-power-noise-f-ha mean value in [tb,stop-time]s var2-power-noise-f-ha variance in [tb,stop-time]s max1-power-noise-f-la max value in [1,80]s ta-power-noise-f-la descent time after 50 s tb-power-noise-f-la rise time after 50 s initial-power-noise-f-la value at tb final-power-noise-f-la value at stop-time min2-power-noise-f-la min value in [tb,stop-time]s max2-power-noise-f-la max value position in [tb,stop-time]s posmax2-power-noise-f-la mean value in [tb,stop-time]s mean2-power-noise-f-la variance in [tb,stop-time]s var2-power-noise-f-la variance in [tb,stop-time]s temp-iniziale-comp-cond initial value at start-time (temperature) min-comp-cond min value in [start-time, stop-time]s posmin-comp-cond min value position in [start-time, stop-time]s temp-finale-comp-cond final value at stop-time (temperature) var1-comp-cond variance in [start-time, tmin]s; tmin=mean value time var2-comp-cond variance in [tmin, stop-time]s; tmin=mean value time slope1-fcv-energy signal slope in [1, 70]s slope2-fcv-energy signal slope in [100, stop-time]s var1-fcv-motor-speed variance in [7, 65]s var2-fcv-motor-speed variance in [97, stop-time]s max1-fcv-motor-speed max value in [64, 82]s var1-fcv-st-power variance in [7, 65]s var2-fcv-st-power variance in [97, stop-time]s max1-fcv-st-power max value in [7, 65]s max2-fcv-st-power max value in [70, 87]s max3-fcv-st-power max value in [87, stop-time]s var1-fcv-st-torque variance in [7, 65]s var2-fcv-st-torque variance in [97,stop-time]s mean1-fcv-st-torque mean value in [7, 65]s mean2-fcv-st-torque mean value in [97,stop-time]s max1-fcv-st-torque max value in [70, 83]s max2-fcv-st-torque max value in [83, 100]s temp-iniziale-ntc-A initial value at start-time (temperature) temp-finale-ntc-A final value at stop-time (temperature) max1-ha-filt-ha max value in [50, stop-time]s max1-ha-filt-la max value in [200, stop-time]s max1-ha-filt-noise max value in [start-time, stop-time] mean1-ha-filt-noise mean value in [start-time, stop-time] var1-ha-filt-noise variance in [start-time, stop-time] min-mean-ha-min-signal min value if thsd is not reached, mean value otherwise outlier-ha-min-signal logic value; 1=thsd reached min-mean-ha-aver-data min value if thsd is not reached, mean value otherwise min-mean-ha-max-signal min value if thsd is not reached, mean value otherwise first-nonnull-ha-init-offset first non zero value max1-ha-noise max value in [start-time, 200] max-ha-min-signal-ha max value in [start-time, stop-time] mean-ha-min-signal-ha mean value in [start-time, stop-time] var-ha-min-signal-ha variance in [start-time, stop-time] max-hum-sens-shot-cnt max value in [start-time, stop-time] min-hum-sens-shot-cnt min value in [start-time, stop-time] diff-hum-sens-shot-cnt max-min in [start-time, stop-time] max-hum-sens-space max value in [start-time, stop-time] mean-hum-sens-space min value in [start-time, stop-time] mean-ha-signals-and-thsd mean value in [start-time, stop-time] var-ha-signals-und-thsd variance in [start-time, stop-time] var-hum-sens-space variance in [start-time, stop-time] final-value1-ha-filt-noise-LPF final value at stop-time slope1-ha-signals-and-thsd-LPF signal slope in [100, 120]s slope1-hum-peak-cnt-double-LPF signal slope in [100, 120]s slope1-hum-sens-shot-cnt-LPF signal slope in [100, 120]s final-value1-hum-sens-space-LPF final value at stop-time max-ntc-A max value in [start-time, time-end]s pos-max-ntc-A max value position in [start-time, time-end]s mean-ntc-A mean value in [1000, time-end]s var-ntc-A variance in [1000, time-end]s max-ha-aver-data max value in [1000, time-end]s pos-max-ha-aver-data max value position in [1000, time-end]s crossing-time-ha-main-signal time when signal value reaches thsd=2,5e04 crossing-time-ha-min-follow time when signal value reaches thsd=1e05 2 mean-ha-noise mean value in [1000, time-end]s var-ha-noise variance in [1000, time-end]s final-value-ha-noise final value in [start-time, time-end]s crossing-time-ha-min-signal-ha time when signal value reaches thsd=2e05 crossing-time-ha-filt-ha time when signal value reaches thsd=2e05 crossing-time-ha-filt-la time when signal value reaches thsd=3e05 max-le-power-noise-f-ha max value in [1000, time-end]s pos-max-le-power-noise-f-ha max value position in [1000, time-end]s mean-le-power-noise-f-ha mean value in [1000, time-end]s var-le-power-noise-f-ha variance in [1000, time-end]s max-le-power-noise-f-la max value in [1000, time-end]s pos-max-le-power-noise-f-la max value position in [1000, time-end]s mean-le-power-noise-f-la mean value in [1000, time-end]s var-le-power-noise-f-la variance in [1000, time-end]s mean-fcv-st-power mean value in [1000, time-end]s var-fcv-st-power variance in [1000, time-end]s mean-fcv-st-torque mean value in [1000, time-end]s var-fcv-st-torque variance in [1000, time-end]s mean-fcv-motor-speed mean value in [1000, time-end]s var-fcv-motor-speed variance in [1000, time-end]s speed-integral-fcv-motor-speed cumsum of values in [1000, time-end] slope-fcv-energy signal slope in [900, 1000]s mean-comp-cond-exit-temp mean value in [1000, time-end]s var-comp-cond-exit-temp variance in [1000, time-end]s crossing-time-ha-filt-noise-LPF time when signal value reaches thsd=2e04 max-hum-peak-normal-cnt-LPF max value in [start-time, time-end]s pos-max-hum-peak-normal-cnt-LPF max value position in [start-time, time-end]s crossing-time-ha-max-signal-LPF time when signal value reaches thsd=1e05 crossing-time-ha-min-signal-LPF time when signal value reaches thsd=6e03 max-ha-signals-und-thsd-LPF max value in [start-time, time-end]s pos-max-ha-signals-und-thsd-LPF max value position in [start-time, time-end]s max-hum-peak-cnt-double-LPF max value in [start-time, time-end]s pos-max-hum-peak-cnt-double-LPF max value position in [start-time, time-end]s max-hum-sens-shot-cnt-LPF max value in [start-time, time-end]s pos-max-hum-sens-shot-cnt-LPF max value position in [start-time, time-end]s crossing-time-hum-sens-space-LPF time when signal value reaches thsd=0,5

(127) A subset of these parameters can be used as well.

(128) Results are shown for two algorithms, LASSO and ridge regression, which have been trained using the fourth dataset where all parameters are present. The prediction of the initial wet load (Ywet), which is a fast prediction being at the beginning of the cycle, the prediction of the dry load (Ydry) which needs a good accuracy determining the end of the cycle, and a humidity prediction (Yhum) are given.

(129) As regards the initial humidity estimation, once obtained wet and dry load estimates it is also possible to extract Yhum simply using the water content (w.c.=Ywet−Ydry) and comparing it with dry load in percentage:
Yhum=(w.c.)*100/Ydry

(130) Alternatively a direct estimation of Yhum is always feasible in the same pattern available for load estimation (adjusting number and type of the classes).

(131) Result 1

(132) Dataset: all tests but no synthetic Estimated output: dry load (Ydry); Regularization technique: Ridge Regression; Type of parameters: all parameters in the list Design matrix: [388×101].

(133) These results are summarized in FIGS. 16a and 16b.

(134) This simulation is executed with Ridge Regression which uses all the parameters available for end-cycle estimation, thus, although this approach is not the most suitable for the implementation on dryers because of its complexity, it gives information on performances achievable with all parameters selected. The dry load estimates are evaluated precisely with a CR=82% as shown in FIG. 16a; when errors occur they remain contained around the main diagonal of the confusion matrix and therefore the lower classes are not “confused” with the upper ones.

(135) Result 2

(136) Dataset: all tests but no synthetic Estimated output: dry load (Ydry); Regularization technique: LASSO Regression; Type of parameters: all parameters in the list Design matrix: [388×101].

(137) These results are summarized in FIGS. 17a and 17b.

(138) LASSO regularization was employed here to compute dry load estimations using 15 predictors at most for each outer cross-validation.

(139) The experimental setting and final results are the same of the result 1 but the CR decreases in this case. In FIG. 17a prediction results are depicted while in FIG. 17b the histogram shows predictors selected by LASSO procedure; after a preliminary analysis the variables linked to motor speed and power together with NTC (temperature sensor) seem to be more influential to determine the dry load end-cycle estimation.

(140) Result 3

(141) Dataset: all tests no synthetic; Estimated output: wet load (Ywet); Regularization technique: Ridge Regression; Type of signals: all parameters in the list; Design matrix: [388×101].

(142) These results are summarized in FIGS. 18a and 18b.

(143) Wet load estimation result: CR=80%, mean error less than 0.4 [Kg].

(144) All predictors have been used.

(145) Now the goal is wet load estimation while the experimental settings and dataset remain the same of results 1 and 2. FIG. 18 reveals a loss of performance in terms of confusion matrix and mean error comparing to result 2; in particular the class 9 [Kg] are correctly evaluated under the 50% of the cases.

(146) Result 4

(147) Dataset: all tests no synthetic; Estimated output: wet load (Ywet); Regularization technique: LASSO15 Regression; Type of signals: all parameters in the list; Design matrix: [388×101].

(148) Wet load estimation result: CR=73%, mean error less than 0.5[Kg].

(149) 15 predictors have been used for each iteration.

(150) The results are summarized in FIGS. 19a and 19b.

(151) In this simulation LASSO was used to determine wet load estimations; here the performance increases respect to dry load estimation in the same conditions. Also in this case closer classes are hard to classify correctly.

(152) Result 5

(153) Dataset: all tests no synthetic; Estimated output: initial moisture (Yhum); Regularization technique: Ridge Regression; Type of signals: All parameters; Design matrix: [388×101].

(154) Results are summarized in FIGS. 20a and 20b.

(155) Initial moisture estimation result: CR=99%, mean error less than 0.5[%].

(156) All predictors have been used.

(157) Only direct estimations of initial humidity is proposed in the thesis and in this perspective results 5 and 6 provide outcomes for Yhum (fast) detection.

(158) Here the goal is to distinguish between two classes only: 50% and 60% of initial moisture respect to dry load. The experimental setting and dataset are exactly the same of previous cases and results for Ridge Regression are shown in FIG. 20; similarly to load estimation case when estimates exceed the maximum feasible value, they are evaluated with 60% anyway.

(159) Result 6

(160) Dataset: all tests no synthetic; Estimated output: initial moisture (Yhum); Regularization technique: LASSO15 Regression; Type of signals: All parameters; Design matrix: [388×101].

(161) Results are summarized in FIGS. 21a and 21b.

(162) Initial moisture estimation result: CR=98%, mean error less than 0.5[%]. 15 predictors have been used for each iteration.

(163) In this simulation LASSO regularization was used to estimate the initial moisture. Results achieved are similar to the ones of Ridge Regression (result 5) but in this case only 15 predictors are employed at most for each iteration with considerable advantages practical implementation aspects.

(164) Parameters derived from compressor signal gain further relevance for the initial moisture estimation.