Method and apparatus for detecting vibrational and/or acoustic transfers in a mechanical system

Abstract

Method and apparatus for detecting vibrational and/or acoustic transfers in a mechanical system A method and apparatus for detecting vibro-acoustic transfers in a mechanical system are provided. The method comprises: while operating the mechanical system, acquiring, at each of multiple input points, an input signal indicative of a mechanical load acting on the input point, and acquiring, at a response point, a response signal indicative of a mechanical response; training a neural network device using the input signals acquired at the input points and using the response signal acquired at the response point; and, for each of the input points: providing only the input signal acquired at the respective input point to the trained neural network device and obtaining, from the neural network device, a contribution signal indicative of a predicted contribution of the respective input signal to the response signal. Vibro-acoustic transfers may be detected solely based on operational data, thereby reducing time and a cost for performing a transfer analysis.

Claims

1. A method for detecting vibrational, acoustic, or vibrational and acoustic transfers in a mechanical system including a plurality of mechanically interacting elements, the method comprising: while operating the mechanical system according to an operating pattern, acquiring, at each input point of a plurality of input points of the mechanical system, an input signal indicative of a mechanical load acting on the respective input point during operation, and acquiring, at a response point of the mechanical system, a response signal indicative of a mechanical response of the mechanical system at the response point; training a neural network device using the input signals acquired at the plurality of input points as training input data and using the response signal acquired at the response point as training output data; and for each input point of the plurality of input points: providing only the input signal acquired at the respective input point as input data to the trained neural network device; and obtaining, as output data output from the trained neural network device in response to only the input signal acquired at the respective input point being provided thereto, a contribution signal indicative of a predicted contribution of the input signal acquired at the respective input point to the response signal acquired at the response point.

2. The method of claim 1, further comprising: testing the trained neural network device, the testing of the trained neural network device comprising: determining a sum of the plurality of contribution signals; subtracting the sum of the plurality of contribution signals from the response signal; determining that the neural network device is sufficiently trained when a result of the subtraction is less than a predetermined threshold; and otherwise, determining that the neural network device is not sufficiently trained.

3. The method of claim 1, wherein the training is supervised training comprising: providing the training input data to the neural network device, such that predicted output data is obtained; and adjusting weightings applied by respective neurons of the neural network device such as to reduce a deviation between the predicted output data and the training output data.

4. The method of claim 1, wherein the training of the neural network device comprises: splitting the input signals and the response signal into a plurality of batches of predetermined length; and for each batch of the plurality of batches, training the neural network device using respective portions of the input signals as the training input data and a respective portion of the response signal as the training output data.

5. The method of claim 1, further comprising: while operating the mechanical system according to a second operating pattern different from the first operating pattern, acquiring, at each input point of the plurality of input points, a second input signal, and acquiring, at the response point of the mechanical system, a second response signal; and performing the providing and the obtaining using the second input signals as the input signals and the second response signal as the response signal.

6. The method of claim 1, further comprising generating and displaying a decomposition of the response signal into the predicted contributions of the input signals acquired at each input point of the plurality of input points.

7. The method of claim 1, further comprising: comparing the predicted contributions of the input signals acquired at each input point of the plurality of input points; identifying an input signal the predicted contribution of which is greatest among the plurality of predicted contributions; and altering the mechanical system at the input point at which the identified input signal has been acquired.

8. The method of claim 1, wherein the neural network device is configured to implement a recurrent neural network comprising an input layer, an output layer, and at least one hidden layer, each layer of the input layer, the output layer, and the at least one hidden layer comprising at least one neuron.

9. The method of claim 8, wherein the neurons of the input layer and the output layer are each configured with a linear activation function, and the neurons of the at least one hidden layer are each configured with a sigmoid-shaped activation function.

10. The method of claim 8, wherein the recurrent neural network is a long short-term memory.

11. The method of claim 8, wherein a total number of neurons of the recurrent neural network is between 10 and 250.

12. A method for detecting vibrational, acoustic, or vibrational and acoustic transfers in a mechanical system including a plurality of mechanically interacting elements using a trained neural network device trained to perform detection of vibrational, acoustic, or vibrational and acoustic transfers in the mechanical system, the method comprising: while operating the mechanical system according to an operating pattern, acquiring, at each input point of a plurality of input points of the mechanical system, an input signal indicative of a mechanical load acting on the respective input point during operation, and acquiring, at a response point of the mechanical system, a response signal indicative of a mechanical response of the mechanical system at the response point; for each input point of the plurality of input points: providing only the input signal acquired at the respective input point as input data to the trained neural network device; and obtaining, as output data from the trained neural network device in response to only the input signal acquired at the respective input point being provided thereto, a contribution signal indicative of a predicted contribution of the input signal acquired at the input point to the response signal acquired at the response point, wherein training of the neural network device uses the input signals acquired at the plurality of input points as training input data and using the respese signal acquired at the response point as training output data, and for each input point of the plurality of input points.

13. An apparatus for detecting vibrational, acoustic, or vibrational and acoustic transfers in a mechanical system including a plurality of mechanically interacting elements, the apparatus comprising: a neural network device; an acquisition unit configured to: while the mechanical system (5) is operated according to an operating pattern, acquire, at each input point of a plurality of input points of the mechanical system, an input signal indicative of a mechanical load acting on the respective input point during operation, and acquire, at a response point of the mechanical system, a response signal indicative of a mechanical response of the mechanical system at the response point; a training unit configured to train the neural network device using the input signals acquired at the plurality of input points as training input data and using the response signal acquired at the response point as training output data; and a prediction unit configured to, for each input point of the plurality of input points: provide only the input signal acquired at the respective input point as input data to the trained neural network device; and obtain, as output data from the trained neural network device in response to only the input signal acquired at the respective input point being provided thereto, a contribution signal indicative of a predicted contribution of the input signal acquired at the respective input point to the response signal acquired at the response point.

14. An apparatus for detecting vibrational, acoustic, or vibrational and acoustic transfers in a mechanical system including a plurality of mechanically interacting elements, the apparatus comprising a trained neural network device that is trained to perform detection of vibrational, acoustic, or vibrational and acoustic transfers in the mechanical system, the apparatus comprising: an acquisition unit configured to, while the mechanical system is operated according to an operating pattern, acquire, at each input point of a plurality of input points of the mechanical system, an input signal indicative of a mechanical load acting on the respective input point during operation, and acquire, at a response point of the mechanical system, a response signal indicative of a mechanical response of the mechanical system at the response point; a prediction unit configured to, for each input point of the plurality of input points: provide only the input signal acquired at the respective input point as input data to the trained neural network device; and obtain, as output data from the trained neural network device in response to only the input signal acquired at the respective input point being provided thereto, a contribution signal indicative of a predicted contribution of the input signal acquired at the respective input point to the response signal acquired at the response point, wherein training of the neural network device uses the input signals acquired at the plurality of input points as training input data and using the response signal acquired at the response point as training output data, and for each input point of the plurality of input points.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 shows a method for detecting vibro-acoustic, transfers in a mechanical system according to a first exemplary embodiment.

(2) FIG. 2 shows an apparatus for detecting vibro-acoustic transfers in a mechanical system according to the first exemplary embodiment.

(3) FIG. 3 is a block diagram illustrating conceptual details of act S20 of FIG. 1.

(4) FIGS. 4, 5, and 6 are block diagrams illustrating conceptual details of act S30 of FIG. 1.

(5) FIG. 7 shows a graphical decomposition of a response signal into a plurality of contribution signals according to a development of the first exemplary embodiment.

(6) FIG. 8 shows a method for detecting vibro-acoustic transfers in a mechanical system according to a second exemplary embodiment.

(7) FIG. 9 shows an apparatus for detecting vibro-acoustic transfers in a mechanical system according to the second exemplary embodiment.

(8) FIG. 10 shows a method for detecting vibro-acoustic transfers in a mechanical system according to a third exemplary embodiment.

(9) FIG. 11 shows an apparatus for detecting vibro-acoustic transfers in a mechanical system according to the third exemplary embodiment.

DETAILED DESCRIPTION

(10) In the figures, like reference numerals designate like or functionally equivalent elements, unless otherwise indicated. First, an example of a mechanical system including a plurality of interacting mechanical elements will be briefly described with reference to FIG. 2.

(11) FIG. 2 shows, inter alia, a prototype car 5 (e.g., a “car” hereinbelow) as one possible example of the mechanical system. For example, the car 5 includes a plurality of mechanically interacting elements, such as, for example, a motor 51, a front wheel 52, a rear wheel 53, a back door 54, and a seat 55.

(12) When the car 5 is driven on a road or a test stand (not shown), acoustic and/or vibrational mechanical loads may act upon the mechanical system 5 at a plurality of excitation sources. Specifically, vibrations of the motor 51 (e.g., primary excitation source), vibrations caused by rotation of the wheels 52, 53 and/or by mechanical interaction between the wheels 52, 53 and the road or test stand (not shown), or by a rattling door such as the back door 54 and the like (e.g., secondary excitation sources) may be input to the mechanical system 5.

(13) These acoustic and/or vibrational mechanical loads input to the mechanical system 5 may be transferred through the car 5 along a variety of structural and/or air-borne transfer paths and may ultimately reach a passenger sitting in the seat 55. The passenger may experience an unpleasant noise, vibrational, or harshness (NVH) sensation.

(14) During design of the prototype car 5, it may be desirable to know to which amount each of the excitations sources 51, 52, 53, 54 contribute to the NVH sensations. The proposed method and apparatus may be beneficially used to provide the required information to enable proper corrective action.

(15) FIG. 1 shows one embodiment of a method for detecting vitro-acoustic transfers in the mechanical system 5 shown in FIG. 2 according to a first exemplary embodiment. In FIG. 1, rectangles denote method acts, and arrows denote a sequence of execution of the method acts.

(16) FIG. 2 shows, next to the car 5, an apparatus 1 for detecting vitro-acoustic transfers in the mechanical system according to the first exemplary embodiment. In FIG. 2, rectangles illustrate respective entities, unit devices, and/or apparatuses, and arrows illustrate a signal flow.

(17) The first exemplary embodiment will be described with reference to both FIG. 1 and FIG. 2.

(18) The apparatus 1 of FIG. 2 includes an acquisition unit 10, a training unit 20, a prediction unit 30, and a neural network device 7.

(19) As the method starts (S), in act S10 of the method illustrated in FIG. 1, operational data is acquired from the car 5.

(20) More specifically, in act S11, the car 5 is operated according to an operating pattern. According to one variant, the car 5 may be driven on a road (not shown) and an operating parameter, such as a rotations per minute (rpm) value of the motor 51, may be recorded in predetermined time intervals and stored in association with time as the operating pattern. According to another variant, the car 5 may be operated on a test stand (not shown) according to a predetermined operating pattern.

(21) While the car 5 is being operated in this manner, in act S12, the acquisition unit 10 of the apparatus 1 acquires a plurality of input signals e± from a plurality of input points 61-64 located in the car 5.

(22) Herein, each of the input points 61-64 may be located at or near a respective one of the primary excitation source 51 and the plurality of secondary excitation sources 52-54. A respective input signal e± is a signal indicative of a mechanical load acting on the mechanical system 5 at the respective input point (e.g., one of 61-64).

(23) A respective input signal e± may be a force signal, an acceleration signal, an acoustic or vibrational signal, a noise signal, a strain signal and pray be acquired using a force sensor, an accelerometer, a microphone, a strain gauge, or the like (not shown) placed at a respective input point (e.g., one of 61-64).

(24) At the same time, while the car 5 is being operated according to act S11 and the input signals e± are acquired according to act S12, the acquisition unit 10 also acquires a response signal r from a response point 65 located in the car 5.

(25) Specifically, the response point may be located at or near a location where a passenger experiences an unpleasant NVH sensation, such as at a head rest of the seat 55.

(26) Also, the response signal r may be a force signal, an acceleration signal, an acoustic or vibrational signal, a noise signal, a strain signal and may be acquired using a force sensor, an accelerometer, a microphone, a strain gauge, or the like placed at the response point 65.

(27) The acquisition unit 10 provides the plurality of input signals e± and the response signal r to the training unit 20.

(28) In act S20 of the method illustrated in FIG. 1, the training unit 20 trains the neural network device 7 using the acquired operational data of the mechanical system 5 so as to cause the neural network device 7 to learn the behavior of the mechanical system 5.

(29) More specifically, the training unit S20 provides the plurality of input signals e± as training input data to the neural network device 7 and also provides the response signal r as training output, data to the neural network device 7, to train the neural network device 7.

(30) The training unit 20 provides the plurality of input signals e± and the response signal r to the prediction unit 30.

(31) In act S30, the prediction unit 30 uses the neural network device 7 to obtain a plurality of contribution signals c± indicative of a predicted contribution of the input signal e± acquired at the respective of the input points 61-64 to the response signal r acquired at the response point 65.

(32) More specifically, acts S31 and S32 are executed once for each input point of the plurality of input points 61-64.

(33) In act S31, the prediction unit 30 provides only the input signal e± acquired at the respective input point (e.g., one of 61-64) as input data to the trained neural network device 7.

(34) In response to providing only the input signal e± acquired at the respective input point (e.g., one of 61-64) to the trained neural network device 7 in act S31, the prediction unit obtains, in act S32, output data from the trained neural network device 7. From the output data obtained in this way, the prediction unit 30 forms a contribution signal c±.

(35) A respective contribution signal c± obtained in this way is considered to be indicative of a predicted contribution of a respective input signal e± to the response signal r. A predicted contribution of a respective input signal e± to the response signal r may also be referred to as a portion or amount of the respective input, signal e± that is transferred through the mechanical system 5 and becomes part of the response signal r at the response point 65.

(36) The plurality of contribution signals c± is provided as an output of the method of FIG. 1 and the apparatus 1 of FIG. 2, and the method ends (E). Conceptual details of further developments of the first exemplary embodiment will now be described with reference to FIGS. 3-6.

(37) FIG. 3 is a block diagram illustrating conceptual details of act S20 of FIG. 1 and FIGS. 4, 5, and 6 are block diagrams illustrating conceptual details of act S30 of FIG. 1.

(38) Specifically, in act S20 (FIG. 1), as is shown in the block diagram in FIG. 3, the neural network device 7 is trained using the plurality of acquired input signals e1, e2, . . . , en, as training input data and using the acquired response signal r as training output data.

(39) The training may be carried out as supervised training including a forward pass and a reverse pass, as illustrated by the bijective arrows in FIG. 3.

(40) Specifically, each of input signals e1, e2, . . . en and the response signal r may be split into a plurality of corresponding portions of predetermined length. The predetermined length may be between 100 and 1000 readings. A respective set including a corresponding portion of the predetermined length of each of the input signals e1, e2, . . . en and the response signal may also be referred to as a batch.

(41) During a forward pass of the supervised training, for each of the batches, the corresponding portions of each of the input signals e1, e2, . . . en are provided to the neural network device 7 as training input data, and output data obtained from the neural network device 7 in response to the training input data is compared to the corresponding portion of the response signal of the batch. Weightings applied by respective neurons of the neural network device 7 are adjusted so as to reduce a deviation between the output data obtained from the neural network device 7 and the corresponding portion of the response signal r of the batch. During a backward pass of the supervised training, for each of the batches, the corresponding portion of the response signal r is provided to the neural network device 7 as training output data, the neural network device 7 is operated in a reverse manner, and training input data obtained from the neural network device 7 in response to the training output data is compared to the corresponding portions of the input signals e1, e2, . . . en of the batch. Weightings applied by the respective neurons of the neural network device 7 are adjusted so as to reduce a deviation between the input data obtained from the neural network device 7 and the corresponding portions of the input signals e1, e2, . . . en of the batch.

(42) After a forward pass and a backward pass have been completed for each of the batches, one epoch of training is completed.

(43) A respective signal e1, e2, . . . en, r may include a number of readings, and each reading may be associated with an operating parameter used at the time of acquisition of the respective reading for operating the mechanical system 5 according to the operating pattern in step S11.

(44) Along with each reading, a respective associated operating parameter may be provided to the neural network device 7 as part of the training input data and/or the training input data. Thereby, the neural network device 7 may learn the behavior of the mechanical system 5 under different operating conditions defined by the operating parameters.

(45) Act S20 may proceed to repeat the training acts described above to complete a predetermined number of epochs and/or until convergence of respective deviations between the training output data and the response signal r (e.g., forward pass) and between the training input data and the plurality of input signals e1, e2, . . . en is attained. In this manner, a trained neural network device 7 may be obtained.

(46) Specifically, in act S30 (FIG. 1), acts S31 and S31 are repeated for each of the plurality of input points for each of the plurality of input signals e1, e2, . . . en).

(47) Specifically, in a first iteration in the loop of S30 shown in FIG. 1, only the input signal e1 is provided to the trained neural network device 7 in act S31 (FIG. 1), as is shown in FIG. 4. Input signals e2, . . . en are not provided to the trained neural network device 7. From the output data obtained from the trained neural network device 7 in response to the input signal ei in act S32 (FIG. 1), a contribution signal Ci is formed. The contribution signal ci is considered to be a prediction of a contribution of the input signal ei acquired at a corresponding input point (61 in FIG. 2) to the response signal r acquired at the response point (65 in FIG. 2).

(48) Next, as is shown in FIG. 5, only the input signal e2 is provided to the trained neural network device 7. From the output data obtained from the trained neural network device 7 in response to the input signal e2, a contribution signal c2 indicative of a predicted contribution of the input signal e2 to the response signal r is formed.

(49) Acts S31 and S32 (FIG. 1) are repeated in this manner for each of the input signals e±.

(50) As is shown in FIG. 6, only the input signal en is provided to the trained neural network device 7. From the output data obtained from the trained neural network device 7 in response to the input signal en, a contribution signal en indicative of a predicted contribution of the input signal en to the response signal r is formed. Merely in order to facilitate understanding, act S30 (FIG. 2) may also be described in the following way, where reference is made to FIGS. 2 to 6:

(51) The trained neural network device 7 may be used to predict, for each of the input points 61-64, a behavior of the mechanical system 5 in a hypothetical situation in which a mechanical load is acting only on a respective single input point (e.g., one of 61-64). Herein, the output data provided by the neural network device in each of FIGS. 4-6 may be considered to be a respective predicted hypothetical response signal predicted to be generated in the mechanical system 5 at the response point 65 when the mechanical system 5 is hypothetically excited only by the respective single input signal e1, e2, . . . en at the respective single input point (e.g. one of 61-64).

(52) By using the neural network device 7 in this manner to predict the behavior of the mechanical system 5 in a hypothetical situation in which a mechanical load is acting only on a respective single one of the input points 61-64, transfers of mechanical loads in the mechanical system 5 may be favorably detected without taking recourse to conventional transfer path analysis (TPA).

(53) Disadvantages of TPA such as complex calculations and having to partly disassemble the mechanical system 6 to be able to carry out an excitation test may be overcome.

(54) FIG. 7 shows an exemplary graphical decomposition 8 of the response signal r into a plurality of contribution signals c1, c2, c3 according to a development of the first exemplary embodiment.

(55) The development will be described with reference to FIG. 7 and FIG. 2.

(56) According to the development, the apparatus 1 may be configured to generate a digital representation of a graphical decomposition from the response signal r and the plurality of contribution signals c1, c2, c3. The apparatus 1 may include and/or be connected to a display device (not shown). In response to being supplied with the digital representation, the display device (not shown) may display a digital representation similar to the exemplary digital representation 8 shown in FIG. 7.

(57) In the exemplary graphical decomposition 8 shown in FIG. 7, an operation parameter such as a number of rotations per minute is plotted on the horizontal axis. Respective signals r, c1, c2, c3 are shown on top of each other along the vertical axis. An empty rectangle denotes low vibro-acoustic load (db0) of a respective signal in a respective operating parameter range. A hatched rectangle denotes medium vibro-acoustic load (db1). A cross-hatched rectangle denotes high vibro-acoustic load (db2). A triple-hatched rectangle denotes very high vibro-acoustic load (db3). The vibro-acoustic load may be in units of decibels in one example.

(58) As shown in FIG. 7, the response signal r acquired at the response point 65 indicates high or very high vibro-acoustic load over a wide operating parameter range from 0 to 3000 rpm. A counter-measure to improve an NVH experience at the response point 65 is to be provided.

(59) As also shown in FIG. 7, the contribution signals c1, c2, and c3 constitute a decomposition of the response signal r. In other words, a respective portion of the response signal r may be rep resented as a sum of corresponding (same position along the horizontal axis) portions of the contribution signals c1, c2, c3.

(60) The exemplary decomposition 8 includes information that, according to the prediction made by the neural network device 7, a highest contribution to the response signal r in the operating parameter range from 0 to 1000 rpm is the contribution ci of an input signal ei acquired at a first input point (such as near the first input point 61 at the motor 51). Accordingly, a countermeasure may be taken in the operating parameter range from 0 to 1000 rpm at the first input point 61.

(61) The exemplary decomposition 8 further includes information that the contribution c3 of an input signal e3 acquired at a third input point 63 (e.g., the rear axis next to the rear wheel 53), to the response signal r is high in the operating parameter range from 2000 to 3000 rpm. Accordingly, a countermeasure may be taken in the operating parameter range from 2000 to 3000 rpm at the third input point 63.

(62) A respective countermeasure may include altering the mechanical system 5 at the respective input point (e.g., one of 61-64). For example, a stiffness may be increased in an area of the input point 61, 63, a damper element may be arranged, or a design of the mechanical system 5 may be altered to alter a location of the respective input point 61, 63 and; or a transfer path between the respective input point 61, 63 and the response point 65.

(63) The respective countermeasure may be taken manually by an engineer or automatically, based on the contribution signals c± output by the apparatus 1, by a design assistance device (not shown).

(64) According to a further development, the apparatus 1 may output the contribution signals c± to a design assistance device or the like so as to cause altering the mechanical system 5 without generating the graphical decomposition 8.

(65) FIG. 8 shows a method for detecting vibro-acoustic transfers in a mechanical system 5 according to a second exemplary embodiment, and FIG. 9 shows a corresponding apparatus 2 according to the second exemplary embodiment.

(66) The second exemplary embodiment resembles the first exemplary embodiment, and like reference numerals designate like elements. Only differences between the second exemplary embodiment and the first exemplary embodiment are described below with reference to FIG. 8 and FIG. 9.

(67) The apparatus 2 according to the second exemplary embodiment also includes a testing unit 40. The testing unit 40 is configured to execute act S40 of the method of the second exemplary embodiment after the plurality of contribution signals c± have been obtained according to acts S10-S30 in a manner similar to the first exemplary embodiment.

(68) In act S40, the testing unit 40 determines a sum of the plurality of contribution signals c±. In other words, the testing unit 40 determines a sum signal, where each reading is obtained by adding corresponding readings of each of the contribution signals c±. The testing unit 40 then proceeds to subtract the sum signal from the response signal r to obtain a result signal. The testing unit 40 calculates a root mean square value of the result signal.

(69) If the root mean square value is below a predetermined threshold, the testing unit 40 determines that the neural network device 7 is sufficiently trained. Otherwise, the testing unit 40 determines that the neural network device 7 is not sufficiently trained.

(70) In other words, and merely to facilitate understanding, with reference to FIG. 7, the testing unit 40 may verify whether the contribution signals c1, c2, c3 correctly add up to the response signal r. If the contribution signals c1, c2, c3 do not correctly add up to the response signal r, the prediction made by the neural network device is not accurate, which may indicate the neural network device 7 is not sufficiently trained.

(71) When it is determined by the testing unit 40 in act S40 that the neural network device 7 is not sufficiently trained, the testing unit 40 notifies the training unit 20 that the neural network device 7 is not sufficiently trained.

(72) After that, according to one variant, the method returns to act 20 to repeat training using the same input signals e± and the same response signal r previously acquired in order to perform a more intense training of the neural network device 7 using the same training input data and training output data as in a previous execution of act S20.

(73) According to another variant, when it is determined that the neural network device 7 is not sufficiently trained, the testing unit 40 returns to act S10. In other words, operating of the mechanical system 5 is repeated in act S11 according to the same or a different operating pattern to acquire further operational data. After that, act S20 is repeated to perform more training of the neural network device 7 using the newly acquired operational data as training input data and training output data.

(74) When it is determined by the testing unit 40 in act S40 that the neural network device 7 is sufficiently trained, the testing unit 40 notifies the training unit 20 that the neural network device 7 is sufficiently trained, and the method ends E.

(75) According to the present exemplary embodiment, when the method according to FIG. 8 is repeated in entirety using the same mechanical system 5 and the same neural network device 7 using the same or a different operating pattern, act S20 may be obviated/omitted (e.g., no training according to act S20 may be performed) when the training unit 20 has been notified that the neural network device 7 is sufficiently trained.

(76) When further operational data (e.g., a second plurality of input signals e± and a second response signal r) is acquired while operating the car 5 using a second operating pattern that differs from the operating pattern used while training the neural network device 7, operating conditions that were not experienced during the previous training act S20 may be experienced. The neural network device 7 may be able to obtain contribution signals ci indicative of a predicted contribution of the respective of the second plurality of input signals e± during the operation condition that was not experienced during training, and advantageously without requiring further training.

(77) FIG. 10 shows a method for detecting vitro-acoustic, transfers in a mechanical system 5 according to a third exemplary embodiment, and FIG. 11 shows a corresponding apparatus 3 according to the third exemplary embodiment.

(78) The third exemplary embodiment resembles the first exemplary embodiment, and like reference numerals designate like elements.

(79) With reference to FIG. 10 and FIG. 11, the third exemplary embodiment differs from the first exemplary embodiment in that the apparatus 3 includes a previously trained neural network device 71 and does not include a training unit (e.g., 20 in FIG. 2). Correspondingly, the method shown in FIG. 10 does not include a training act (e.g., S20 in FIG. 2).

(80) The trained neural network device 71 may be a neural network device that has been previously trained using acts such as acts S10 and S20 (FIG. 2) of the method according to the first or second exemplary embodiment.

(81) The trained neural network device 71 may be installed in the car 5. As has been discussed for the first and second exemplary embodiments, the trained neural network device 71 may have artificial knowledge about mechanical transfers of the mechanical system of the car 5. This knowledge may be used to enable various functionality of the car 5. For example, the trained neural network device 71 may be used to control an audio system, a noise cancelling system, and the like of the car 5 also in unforeseen operating conditions that were not encountered during a design phase of the car 5.

(82) A practical exemplary use case of the exemplary embodiments will be briefly discussed.

(83) An apparatus 1, 2 for detecting vibrational and/or acoustic transfers in a prototype car 5 was built.

(84) The neural network device 7 of the apparatus 1, 2 was configured to implement a recurrent neural network. The recurrent neural network included several layers of artificial neurons, and, for example, an input layer, an output layer, and five hidden layers. For example, the recurrent neural network was configured as a long short-term memory. The neurons of the input layer and of the output layer were configured with a linear activation function. The neurons of the five hidden layers were configured with a sigmoid-shaped, hyperbolic tangent activation function. A dropout of 0.2 was implemented between each layer. A total number of neurons of the LSTM was 100.

(85) The prototype car was operated according to an operating pattern including 5 run-up operations runs, all on 4th gear at 50% throttle. Respective input and response signals were acquired, each including 27136 readings.

(86) The neural network device 7 was trained in hatches of 1000 readings. Convergence was obtained in less than 1000 epochs.

(87) Contribution signals c± of individual input signals e± and a predicted response signal for the plurality of input signals e± were determined using both a method according to the embodiments and using conventional transfer path analysis (TPA). A useful level of agreement between the two approaches was confirmed.

(88) Although the present invention has been described in accordance with exemplary embodiments, it is obvious for the person skilled in the art that modifications are possible in all exemplary embodiments.

(89) In the exemplary embodiments, supervised training has been described as a method of training the neural network device 7. However, an unsupervised training method may be used instead.

(90) A car 5 has been described as an example of the mechanical system 5. However, the present invention may also be useful in applications in aviation engineering, machinery engineering, and the like.

(91) The graphical decomposition 8 shown in FIG. 7 is merely an example. The graphical decomposition 8 may include a different, such as a larger, quantity of operating ranges and a different (e.g., larger) quantity of contribution signals c±. Other ways of visualizing the contribution signals c± may be used for the graphical decomposition 8, such as a plurality of graphs, plots, pie charts, and the like.

(92) It is understood that a respective neural network device is implemented to receive a plurality of input signals and to generate a plurality of output signals after being trained, where the output signals are generated employing artificial intelligence acquired through the before-mentioned training processes. The input and output signals may include encoded data referring to or representing physically observable quantities. In embodiments, such quantity is a mechanical or vibrational load, a frequency, or another mechanical stimulus to the mechanical system and/or a measure for an NVH contribution. In embodiments, input and or output signals are generated and transmitted through a network, and the content of the signals are stored at least temporarily by respective memory devices.

(93) The disclosed embodiments of methods and devices allow for an efficient transfer path analysis in car or vehicle design automation.

(94) It is to be understood that the elements and features recited in the appended claims may be combined in different ways to produce new claims that likewise fall within the scope of the present invention. Thus, whereas the dependent claims appended below depend from only a single independent or dependent claim, it is to be understood that these dependent claims can, alternatively, be made to depend in the alternative from any preceding or following claim, whether independent or dependent, and that such new combinations are to be understood as forming a part of the present specification.

(95) While the present invention has been described above by reference to various embodiments, it should be understood that many changes and modifications can be made to the described embodiments. It is therefore intended that the foregoing description be regarded as illustrative rather than limiting, and that it be understood that all equivalents and/or combinations of embodiments are intended to be included in this description.