Radiometric measurement device

Abstract

A radiometric measurement device includes a number n of sensors, wherein a respective sensor of the number n of sensors is configured to generate associated sensor data, such that overall a number n of sensor data is generated by means of the number n of sensors. A measurement variable calculation unit is configured to calculate a number m of measurement variable values depending on the number n of sensor data on the basis of values of a number d of parameters. A learning unit is configured to calculate the values of the number d of parameters on the basis of training data.

Claims

1. An apparatus, comprising: a process metrology radiometric measurement device, the radiometric measurement device comprising: a number n of sensors, wherein a respective sensor of the number n of sensors is configured to generate associated sensor data, such that overall a number n of sensor data is generated by way of the number n of sensors; a measurement variable calculation unit configured to calculate a number m of process measurement variable values depending on the number n of sensor data on the basis of values of a number d of parameters; and a learning unit, wherein the learning unit is configured to calculate the values of the number d of parameters on the basis of training data, wherein the radiometric measurement device self-calibrates via the learning unit, and at least one process measurement variable is selected from a set of process measurement variables comprising: filling level of a material, positions and/or thicknesses of individual material layers of a material, material density, conveyed mass of a material, mass flow rate of a material throughput.

2. The radiometric measurement device according to claim 1, wherein the measurement variable calculation unit comprises at least one feature extraction unit configured to extract feature data from the number n of sensor data, on the basis of the values of the number d of parameters, and the at least one measurement variable calculation unit is configured to calculate the number m of measurement variable values from the feature data on the basis of the values of the number d of parameters.

3. The radiometric measurement device according to claim 1, wherein the number n of sensors is selected from a set of sensors comprising: at least one radiometric sensor configured to generate sensor data in the form of a counting rate or radiation intensity data, and/or a radiometric sensor configured to generate sensor data in the form of information about radiometric spectra, at least one sensor configured to generate sensor data in the form of temperature data, at least one sensor configured to generate sensor data in the form of acceleration data, at least one sensor configured to generate sensor data in the form of velocity data, at least one sensor configured to generate sensor data in the form of position data, at least one sensor in the form of an ultrasonic sensor or laser sensor, configured to generate sensor data in the form of information about a loading height profile, and at least one sensor configured to generate sensor data in the form of moisture data.

4. The radiometric measurement device according to claim 1, wherein the learning unit is configured to extract a number n of training sensor data and a number m of associated setpoint values of the number m of measurement variable values from the training data, the measurement variable calculation unit is configured to calculate a number m of training values of the number m of measurement variable values depending on the number n of training sensor data, and the learning unit is configured to calculate the values of the number d of parameters on the basis of the number m of setpoint values of the number m of measurement variable values and the number m of training values of the number m of measurement variable values.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 is a highly schematic block diagram of a radiometric measurement device according to an embodiment of the invention.

(2) FIG. 2 is a highly schematic block diagram of an internal structure of one embodiment of a measurement variable calculation unit of the radiometric measurement device shown in FIG. 1.

(3) FIG. 3 is a highly schematic block diagram of an internal structure of a further embodiment of a measurement variable calculation unit of the radiometric measurement device shown in FIG. 1.

(4) FIG. 4 is a highly schematic block diagram of the radiometric measurement device shown in FIG. 1 in a learning mode of operation.

DETAILED DESCRIPTION OF THE DRAWINGS

(5) FIG. 1 shows highly schematically a block diagram of a radiometric measurement device 1.

(6) The radiometric measurement device 1 comprises a number n of sensors 2_1 to 2_n, wherein a respective sensor 2_i of the number n of sensors 2_1 to 2_n is configured to generate associated sensor data xi. The result is that overall a number n of sensor data x.sub.1, . . . ,x.sub.n is generated by means of the number n of sensors 2_1 to 2_n.

(7) The radiometric measurement device 1 furthermore comprises a measurement variable calculation unit 4 configured to calculate a number m of measurement variable values y.sub.1, . . . ,y.sub.m depending on the number n of sensor data x.sub.1, . . . ,x.sub.n on the basis of values of a number d of parameters θ.sub.1, . . . , θ.sub.d.

(8) Referring to FIG. 4, the radiometric measurement device 1 comprises a learning unit 5, wherein the learning unit 5 is configured to calculate the values of the number d of parameters θ.sub.1, . . . , θ.sub.d on the basis of training data xt.sub.1.sup.(i), . . . , xt.sub.n.sup.(i); ys.sub.1.sup.(i), . . . , ys.sub.m.sup.(i).

(9) The radiometric measurement device 1 converts input variables in the form of the sensor values x.sub.1, . . . ,x.sub.n, which can also be time-offset, into output variables in the form of the process measurement variables or measurement variable values y.sub.1, . . . ,y.sub.m.

(10) The conversion depends on the model parameters θ.sub.1, . . . , θ.sub.d, which are initially unknown and are learnt by way of the learning unit 5 via so-called machine learning. Recorded training data, also called learning data, are used here, which can be formed from real recorded data during operation and/or from simulation data.

(11) Machine learning means that the radiometric measurement device artificially generates knowledge from experience. The radiometric measurement device learns from examples and can generalize them after the end of the learning phase. That means that the examples are not simply learned by heart, rather the radiometric measurement device recognizes patterns and regularities in the training data. It can thus also assess unknown data (learning transfer).

(12) The radiometric measurement device preferably uses learning techniques from so-called supervised learning, wherein the radiometric measurement device learns a measurement function from given pairs of inputs and outputs. In this case, the correct measurement variable values with respect to a number n of sensor data are provided during learning for example on the basis of a reference measurement or a simulation.

(13) The radiometric measurement device thus formally approximates a measurement function ƒ:(x.sub.1, . . . , x.sub.n)(y.sub.1, . . . , y.sub.m),

(14) which maps n input variables or sensor data (x.sub.1, . . . , x.sub.n) onto m output variables or measurement variable values (y.sub.1, . . . , y.sub.m), by a suitable hypothesis function h.sub.θ:(x.sub.1, . . . , x.sub.n)(ŷ.sub.1, . . . , ŷ.sub.m),

(15) which maps the n sensor data (x.sub.1, . . . , x.sub.n) onto m estimated values (ŷ.sub.1, . . . , ŷ.sub.m) for the (y.sub.1, . . . , y.sub.m) and is dependent on the model parameters θ:=(θ.sub.1, . . . , θ.sub.d).

(16) In this case, each of the d individual model parameters θ.sub.i is understood to be one of the following three things:

(17) (1) a mathematical object, in particular a number, a vector, a function;

(18) (2) a parameterized piece of program logic or source code;

(19) (3) a piece of program logic or source code generated by a code generator.

(20) The model parameters (θ.sub.1, . . . , θ.sub.d) are learned from training data by the learning algorithm. More precisely, training data consist of l (wherein l is for example in a range of between 10.sup.5 and 10.sup.7, in particular l=10.sup.6) training pairs (xt.sup.(1), ys.sup.(1)), . . . , (xt.sup.(l), ys.sup.(l)), which e.g. each have the dimension n+m and for example each consist of a complete set of input data or training sensor data xt.sup.(i):=(xt.sub.1.sup.(i), . . . , xt.sub.n.sup.(i)) plus associated setpoint values ys.sup.(i):=(ys.sub.1.sup.(i), . . . , ys.sub.m.sup.(i)) of the number m of measurement variable values, wherein i=1, . . . , l. The setpoint values (ys.sub.1.sup.(i), . . . , ys.sub.m.sup.(i)) are also referred to as training labels.

(21) On the basis of the training sensor data (xt.sub.1.sup.(i), . . . , xt.sub.n.sup.(i)), the measurement variable calculation unit 4 calculates the training values (yt.sub.1.sup.(i), . . . , yt.sub.m.sup.(i)):=h.sub.θ(xt.sub.1.sup.(i), . . . , xt.sub.n.sup.(i)) of the number m of measurement variable values (y.sub.1, . . . , y.sub.m), said training values being dependent on the parameters (θ.sub.1, . . . , θ.sub.d). The learning unit 5 is configured to calculate the values of the parameters (θ.sub.1, . . . , θ.sub.d) for i=1, . . . ,l on the basis of setpoint values (ys.sub.1.sup.(i), . . . , ys.sub.m.sup.(i)) and the training values (yt.sub.1.sup.(i), . . . , yt.sub.m.sup.(i)).

(22) The calculation of the model parameters (θ.sub.1, . . . , θ.sub.d) can be carried out repeatedly iteratively. That is to say that random starting parameters (θ.sub.1, . . . , θ.sub.d) are used as a beginning. These are then improved iteratively by calculating repeatedly all (yt.sub.1.sup.(i), . . . , yt.sub.m.sup.(i)) on the basis of the respectively current (θ.sub.1, . . . , θ.sub.d) and then new, improved (θ.sub.1, . . . , θ.sub.d) therefrom, until a predefinable quality measure is attained (for example minimum of a cost function). Moreover, in each iteration step, it is possible to use only a subset of the total of 1 data sets (yt.sub.1.sup.(i), . . . , yt.sub.m.sup.(i)), a so-called mini-batch, in order to calculate new (θ.sub.1, . . . , θ.sub.d). That is to say that a plurality of iterations may be required in order to take account of the entire training data once, a so-called training epoch.

(23) The learning algorithm is carried out for example once upon the start-up of the radiometric measurement device or repeatedly in real time during the operation of the radiometric measurement device, e.g. by means of additional reference measurements.

(24) FIG. 2 shows highly schematically a block diagram of an internal structure of one embodiment of a measurement variable calculation unit 4 of the radiometric measurement device 1 shown in FIG. 1.

(25) The measurement variable calculation unit 4 comprises an optional feature extraction unit 3 configured to extract feature data FD from the number n of sensor data x1, . . . ,xn, in particular on the basis of the values of the number d of parameters θ.sub.1, . . . , θ.sub.d.

(26) The measurement variable calculation unit 4 furthermore comprises an artificial intelligence (AI) unit 6 configured to calculate the number m of measurement variable calculation unit values y1, . . . ,ym from the feature data FD on the basis of the values of the number d of parameters θ.sub.1, . . . , θ.sub.d.

(27) From the “raw” sensor data or measurement data x.sub.1, . . . ,x.sub.n, firstly suitable features FD are extracted and transformed in order to generate as meaningful input data as possible for the AI unit 6. In particular, one or more of the following techniques are used for this purpose: principal component analysis (PCA), discriminant analysis, statistical normalization, polynomial transformation, exponential transformation, logarithmic transformation.

(28) It goes without saying that the feature extraction can also be dispensed with, such that the AI unit 6 uses the non-preprocessed, raw sensor data x.sub.1, . . . ,x.sub.n.

(29) The AI unit 6 calculates a continuous output signal (regression method) or a discrete output signal (classification method), depending on the measurement application. It is realized by an AI model from one of the following four categories:

(30) Models which, by way of metrics or suitable similarity functions, in a single stage or a plurality of stages, compare the input values with the stored training data and then allocate to them the output values of those training data which are “nearby” or similar in a certain way.

(31) This can be for example one of the following two AI models: (1) k-nearest-neighbor classification, (2) k-nearest-neighbor regression;

(32) metrics or similarity functions used can be, in particular: p-norm, Minkowski distance, Kullback-Leibler divergence.

(33) Models which calculate threshold values from the training data, against which threshold values the given input values are then compared in a plurality of stages, usually recursively, in order to determine the associated output values.

(34) This can be for example one of the following two AI models: (1) decision tree classification, (2) decision tree regression.

(35) Models which estimate transition probabilities from the training data and combine said transition probabilities with one another (possibly in a plurality of stages) additively and multiplicatively using the Bayes theorem in order to estimate for given input values a univariate or multivariate probability distribution on the output values. The output values having the highest probabilities are then allocated to the input values.

(36) This can be for example one of the following two AI models: (1) Bayes classifier, in particular naive Bayes classifier, (2) Bayesian network classifier.

(37) Models which, with the aid of methods of linear algebra, in a single stage or a plurality of stages, apply so-called activation functions to linear combinations and/or convolutions of the transformed or untransformed input values in order then to calculate the output values therefrom.

(38) This can be for example one of the following two AI models: (1) multiclass support vector machine (SVM) with one-vs-one or one-vs-all, kernel functions used can be, in particular: polynomial kernel, Gaussian RBF kernel, Laplace RBF kernel, sigmoid kernel, hyperbolic tangent kernel, Bessel kernel, Anova kernel, linear splines kernel; (2) artificial neural network (ANN) and/or deep neural network (DNN),

(39) activation functions used can be, in particular: identity, sigmoid, hyperbolic tangent, ReLu, softmax, Signum.

(40) Referring to FIG. 3, which illustrates highly schematically a block diagram of an internal structure of a further embodiment of a measurement variable calculation unit 4 of the radiometric measurement device 1 shown in FIG. 1, optionally a number from a plurality of individual AI units 6 of categories above can also be combined to form a more powerful overall model with the aid of ensemble learning. In FIG. 3, three feature extraction units 3 and three AI units 6 respectively connected downstream operate in parallel, wherein a so-called ensemble combiner 7 combines the respective data. By way of example, it is possible to use so-called bagging or so-called boosting as an ensemble learning technique.

(41) The model parameters (θ.sub.1, . . . , θ.sub.d) of the AI unit(s) 6 are determined from the training data by means of machine learning for example using one of the following techniques:

(42) (a) by one-off or repeated minimization of metrics or maximization of similarity functions. This can be in particular one or more of the following: Entropy, Gini impurity, variance, p-norm, Minkowski distance, Kullback-Leibler divergence;

(43) (b) by means of one-off or repeated minimization of a cost function which is dependent on the AI model chosen and the function arguments of which consist of the training data and the model parameters. The minimization may be subject to certain mathematical constraints (restriction of the search area), which may likewise be dependent on the training data and/or model parameters. The cost function is minimized with regard to the model parameters, using mathematical optimization methods and techniques, in particular one or more of the following: backpropagation, gradient descent based method, stochastic gradient descent based method (for example AdaGrad, RMSProp or Adam), Gauss-Newton method, quasi-Newton method, linear programming, quadratic programming.

(44) The minimization of a cost function can include the maximization of a merit function, in particular of a maximum likelihood function or of a maximum a posteriori probability function, in particular by changing the mathematical sign.

(45) In order to prevent overfitting, to improve the capability of learning transfer and thus to increase the performance of the AI unit(s) 6, in the learning method it is possible to use additional regularization techniques, such as, for example:

(46) p-norm penalty terms (L1, L2, etc.),

(47) dropout,

(48) batch normalization.

(49) Unless defined otherwise, all AI technical terms should be understood in accordance with the academic standard literature for AI and machine learning. See also, in particular:

(50) 1. Bishop, Christopher M.: “Pattern Recognition and Machine Learning”

(51) 2. Mitchell, Tom M.: “Machine Learning”

(52) 3. Russell, Stuart J. and Norvig, Peter: “Artificial Intelligence: A Modern Approach”

(53) 4. Richard O. Duda and Hart, Peter E. and David G. Stork: “Pattern Classification”

(54) 5. Aggarwal, Charu C.: “Neural Networks and Deep Learning: A Textbook”.

(55) The foregoing disclosure has been set forth merely to illustrate the invention and is not intended to be limiting. Since modifications of the disclosed embodiments incorporating the spirit and substance of the invention may occur to persons skilled in the art, the invention should be construed to include everything within the scope of the appended claims and equivalents thereof.