Method and control device for emulating noise components of a lossily recorded operating signal

Abstract

In order to imitate noise components of a lossily recorded operating signal of a technical system (TS), a neural network (NN, BNN) is trained for reproducing a recorded target operating signal (CST) and a statistical distribution of random components of the recorded target operating signal (CST) from a recorded input operating signal (CSI). A current input operating signal (SI) of the technical system (TS) is fed to the trained neural network (NN, BNN). From the fed current input operating signal (SI) and a noise signal (NS), an output signal (OST) is generated which has noise components imitating the statistical distribution. The output signal (OST) is then output as a current target operating signal for controlling the technical system (TS).

Description

Background Technology

[0001] To control complex technological systems, such as gas turbines, wind turbines, steam turbines, engines, robots, manufacturing equipment, or motor vehicles, data-driven approaches using machine learning are increasingly being employed. In this context, artificial neural networks are specifically trained to determine control-related operational signals for controlling the technological system based on detected operational signals, such as sensor values ​​from the system.

[0002] To successfully train a data-driven control device, a large amount of operational data from the technical system is typically required as training data. This training data should, as comprehensively as possible, cover the operating states and other operating conditions of the technical system. Such training data often exists in the form of a database storing a large number of operational signals recorded at the technical system.

[0003] However, due to bandwidth or storage limitations, these training operation signals are often lossily compressed before being recorded or have already been lossily detected at the technical system. This is particularly true for gas turbines, where operation signals available for training purposes are typically heavily compressed.

[0004] An example of this lossy compression or detection is the method often referred to as dead-banding, where a constant value is recorded for the corresponding running signal as long as the variation in the running signal remains below a pre-given threshold. Even when the corresponding running signal is discretized, comparable information is lost.

[0005] Such information loss, often irreversible, can significantly impact the training of data-driven control devices. With known training methods, less compression is often chosen, but this increases storage and / or bandwidth requirements. On the other hand, with heavily compressed training data, inaccuracies are tolerated in data-driven control. Summary of the Invention

[0006] The objective of this invention is to describe a method and control device that allow for more efficient training and / or stronger compression of training data.

[0007] This task is accomplished by a computer-based method for mimicking the noise component of a lossily recorded operating signal of a technical system, by a control device for controlling the technical system, by a computer program product, and by a computer-readable storage medium.

[0008] According to the present invention, the noise component of the operating signal of a lossily recorded technical system is simulated, wherein the input operating signal for the control device of the technical system and the target operating signal for controlling the technical system are recorded in a lossy manner. A neural network is trained based on the recorded operating signal to reproduce the recorded target operating signal and the statistical distribution of the random components of the recorded target operating signal based on the recorded input operating signal. Then, the current input operating signal of the technical system is fed to the trained neural network. According to the present invention, an output signal having a noise component that simulates the statistical distribution is generated based on the fed current input operating signal and the noise signal. The output signal is then output as the current target operating signal for controlling the technical system.

[0009] The present invention is based on the observation that, due to the lossy recording of operating signals, information, especially regarding random fluctuations in the operating signals, is also lost. Such fluctuations arise particularly from the indeterminate behavior of the technical system based on the available operating signals. However, such indeterminate or uncertain behavior constitutes a significant component of the behavior of many technical systems. Thus, for example, in the case of wind turbines, the exact wind conditions and consequently the available power cannot be determined in large quantities.

[0010] Correspondingly, simulators or other control devices of technical systems trained with lossy operating signals often exhibit more deterministic behavior than real systems. As long as actual statistical fluctuations are not accurately mapped, the reliability or operational risk of a technical system cannot usually be evaluated sufficiently accurately using such trained control devices.

[0011] In contrast, while the present invention typically cannot reconstruct the precise changes in uncertain behavior, it can mimic at least one noise component of the target running signal, which has a realistic statistical distribution. This allows for significantly more realistic predictions about the behavior of the technical system in many cases, even with lossy training data, and thus allows for more efficient training. Alternatively or additionally, the present invention also allows for more forceful compression of the training data.

[0012] Corresponding control devices, computer program products, and computer-readable, preferably non-volatile storage media are configured to implement the method according to the invention.

[0013] The method according to the invention and the control device according to the invention can be implemented or realized, for example, by means of one or more computers, processors, application-specific integrated circuits (ASICs), digital signal processors (DSPs) and / or so-called "field-programmable gate arrays" (FPGAs).

[0014] Advantageous embodiments and improvements of the present invention are described in other design embodiments.

[0015] According to an advantageous embodiment of the invention, a Bayesian neural network with latent parameters representing a statistical distribution can be used as the neural network. Here, the latent parameters can be inferred through training, particularly by means of variational inference methods and / or by means of Markov chain Monte Carlo methods. Noise signals can be fed into the input layer of the Bayesian neural network. An output signal can then be generated from the current input running signal and the noise signal by the Bayesian neural network trained using the inferred latent parameters. The Bayesian neural network can model random components, thereby realizing a probabilistic method of supervised learning. Efficient numerical training methods can be used to train such a Bayesian neural network with latent parameters.

[0016] According to another advantageous embodiment of the invention, a neural network can be trained to reproduce the statistical characteristic values ​​of a statistical distribution based on a recorded input running signal. Here, for training purposes, a likelihood function can be used as the error function to be minimized. For this purpose, a large number of numerical standard methods are available. The statistical characteristic values ​​can then be determined by the trained neural network for the current input running signal. Accordingly, a noise signal can be generated based on or according to the determined statistical characteristic values ​​and output as an output signal. In particular, the mean and variance of the statistical distribution can be used as characteristic values. Furthermore, the standard deviation, probability value, and / or distribution type of the statistical distribution can be used as characteristic values.

[0017] According to another advantageous embodiment of the invention, the current input operating signal can be continuously detected and fed to a neural network. Thus, by means of the neural network, a concurrent simulator, particularly a digital twin of the technical system, can preferably be run in real time. Since the random components of the current target operating signal largely conform to realistic statistics, the simulator can also simulate the uncertain behavior of the technical system in a sufficiently realistic manner. Attached Figure Description

[0018] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. Here, schematic diagrams are provided respectively: Figure 1 This illustrates the time-varying process of the operating signals of the technical system before and after its lossy compression. Figure 2 The first control device according to a first embodiment of the present invention is shown in the training phase. Figure 3 The first control device is shown in the application phase. Figure 4 A second control device according to a second embodiment of the present invention is shown in the training phase, and Figure 5 The second control device is shown in the application phase.

[0019] Where the same or corresponding reference numerals are used in the figures, these reference numerals denote the same or corresponding entities that can be implemented or designed, in particular as described in conjunction with the relevant figures. Detailed Implementation

[0020] Figure 1 This section elucidates the temporal changes in the operational data of a technical system (TS) before and after its lossy compression. A technical system (TS) can be, for example, a gas turbine, wind turbine, engine, robot, manufacturing equipment, motor vehicle, or other complex technical system.

[0021] Operating signals can be measured by sensing or otherwise detected at the technical system TS. These operating signals should be recorded for data-based training of the technical system TS or control devices for similar or analog technical systems.

[0022] The operating signal includes one or more input operating signals SI, which should be used as inputs to the control device for the technical system to be controlled. The input operating signal SI may, for example, include or represent sensor data, measurements, status data, control action data, or other data detected at or for the technical system TS for control purposes. In the case of a turbine, the input operating signal SI may, for example, include speed data, temperature data, or pressure data.

[0023] The operating signals also include one or more target operating signals ST, which are required for the efficient control of the technology system TS. The target operating signals ST may, in particular, include or represent optimized control actions, data regarding the effects of the control actions, data that is difficult to measure, predictive data, or other signals or data related to control. In the case of a turbine, the target operating signals ST may, for example, include data regarding combustion dynamics, vibration data, and / or data regarding temperature or pressure distribution at hard-to-reach locations on the turbine.

[0024] When operating complex technical systems, a large number of operating signals, referred to here as SI and ST, are typically generated. For this reason, these operating signals are usually compressed. To achieve sufficiently strong compression, lossy compression methods are often applied in this case. In the case of compression methods, also known as dead-zone compression, a constant value is recorded for the corresponding operating signal as long as the variation in the operating signal remains below a pre-given threshold. Alternatively or additionally, the operating signals can also be compressed by more or less coarse discretization.

[0025] For compression purposes, the operating signals SI and ST are transmitted from the technical system TS to the compression unit CPR. The compression unit CPR compresses the corresponding input operating signal SI into a compressed input operating signal CSI in a lossy manner, and compresses the corresponding target operating signal ST into a compressed target operating signal CST in a lossy manner. The compressed input operating signal CSI and the compressed target operating signal CST are stored in the database DB.

[0026] In this way, one or more databases can be created by one or more technical systems, containing a large number of compressed operating signals (CSI and CST). As mentioned above, the recorded operating signals CSI and CST should be used in conjunction with the present invention to train a learning-based control device for the technical system to be controlled in a data-driven manner. It is anticipated that the more similar the technical system to be controlled is to the one or more technical systems TS from which the training data (CSI and CST in this case) originates, the more efficient the training will be.

[0027] The control device should be trained based on compressed operating signals CSI and CST to determine or predict an optimized target operating signal ST based on the input operating signal SI of the technical system to be controlled, thereby controlling the technical system. In this way, a target operating signal that is difficult to measure or is only available at a later point in time can also be determined or reconstructed, which is useful for controlling the technical system.

[0028] However, as mentioned above, the information of the original operating signals SI and ST is lost due to lossy compression, which in particular affects the random noise components of the original operating signals SI and ST.

[0029] The change in the original target running signal ST caused by the dead zone to the compressed target running signal CST. Figure 1 This is explained in the following section. Here, the original target running signal ST is represented by a solid line, while the compressed target running signal CST is represented by a dashed line. The running signals are plotted relative to time in arbitrary units. It is shown that many of the finer structures of the original target running signal ST can no longer be found in the compressed target running signal CST. In particular, the randomly appearing fluctuations or noise components of the original target running signal ST can no longer be easily identified in the compressed target running signal CST.

[0030] The neural network should be trained using the training method according to the invention, particularly for reproducing the statistical distribution of the random components of the recorded target running signal CST. With the aid of the trained neural network and a noise generator, the noise components of the target running signal, which have a corresponding statistical distribution, can then be simulated.

[0031] In this context, training should generally be understood as optimizing the mapping from the input data of a machine learning model (here, a neural network) to its output data. This mapping is optimized during its training phase according to pre-given, learned, and / or to-be-learned criteria. For example, reproduction error, prediction error, the success of the output control action, or similarity with respect to statistical distributions can be used as criteria. Through training, for example, the network structure of the neurons in the neural network and / or the weights of the connections between neurons can be adjusted or optimized to satisfy the pre-given criteria as well as possible. Therefore, training can be interpreted as an optimization problem. A large number of efficient optimization methods are available for such optimization problems in the field of machine learning. Here, optimization should always be understood as an approximation of the optimal value.

[0032] Figure 2 The training of a first control device CTL1 according to a first embodiment of the present invention is explained. The first control device CTL1 includes one or more processors PROC for implementing the method steps of the control device CTL1 and one or more memories MEM for storing data to be processed.

[0033] The first control unit CTL1 further includes a Bayesian neural network (BNN) and a noise generator NSG. The noise generator NSG is used to generate a noise signal NS, for example, by means of a random number generator that generates pseudo-random numbers or other random data. Following common linguistic conventions, the term random data or random numbers should also include pseudo-random data or pseudo-random numbers. To train the Bayesian neural network (BNN), the compressed input running signal CSI and the compressed target running signal CST recorded therein, stored in the database DB, are read by the first control unit CTL1 and fed into the input layer of the Bayesian neural network (BNN) as training data. The generated noise signal NS, or its data values, are fed into the input layer of the Bayesian neural network (BNN) as other input data.

[0034] A Bayesian Neural Network (BNN) can be interpreted as a neural network with a stochastic component that models a statistical distribution. In particular, probabilistic methods for supervised learning can be implemented using BNNs.

[0035] According to the first embodiment, the Bayesian neural network (BNN) should be trained based on compressed training data CSI and CST, and a noise signal NS, to generate a target running signal OST with uncertain random components based on the input running signal and the noise signal. Here, the statistical distribution of the generated random components should correspond to the static distribution of the random components of the compressed target running signal CST. The target running signal OST should be output as an output signal or as output data through the output layer of the Bayesian neural network (BNN).

[0036] Given the same input operating signal CSI, the random component of the compressed target operating signal CST becomes particularly noticeable due to the statistical fluctuations of the CST. The statistical distribution of these fluctuations can be quantified, for example, by the mean, variance, or other statistical characteristic parameters of the fluctuations.

[0037] According to the first embodiment, the uncertainty properties of these fluctuations are modeled, represented, or represented by the so-called latent parameter (LV) of the Bayesian neural network (BNN). Therefore, the latent parameter LV represents, to some extent, the statistical distribution of the random components of the target running signal.

[0038] During the training of the Bayesian neural network (BNN), the latent parameter LV is preferably inferred from statistical fluctuations for each data point of the running signals CSI and CST. Here, the latent parameter LV is estimated, identified, or adjusted such that the uncertainty fluctuations induced by the fed noise signal NS reproduce the random components of the compressed target running signal CST with respect to its statistical distribution.

[0039] For this purpose, the statistical distribution of the compressed target running signal CST is compared with the statistical distribution of the output signal OST (CMP). Specifically, the mean and / or variance of the statistical distributions can be compared. The CMP comparison... Figure 2 The comparison CMP is shown outside of the Bayesian Neural Network (BNN). However, alternatively or additionally, the comparison CMP can also be implemented entirely or partially within the Bayesian Neural Network (BNN).

[0040] The deviation D between the statistical distributions being compared, determined when comparing CMPs—such as in Figure 2 The dashed arrow indicates a return to the Bayesian Neural Network (BNN). Therefore, the latent parameters LV and neural weights of the BNN are optimized to minimize the bias D. In this way, the BNN, along with its latent parameters LV, is trained to reproduce the target running signal and its random components with respect to its statistical distribution based on the input running signal.

[0041] Inference of latent parameters (LVs) can be carried out, in particular, using variational inference methods or Markov chain models. These and other efficient training and inference methods for Bayesian neural networks are described, for example, in Stefan Depeweg et al.'s "Learning and Policy Search in Stochastic Dynamical Systems with Bayesian Neural Networks" (ICLR 2017); Stefan Depeweg et al.'s "Decomposition of Uncertainty in Bayesian Deep Learning for Efficient and Risk-sensitive Learning" (International Conference on Machine Learning, 2018); and Christopher M. Bishop's "Pattern Recognition and Machine Learning" (Springer 2011).

[0042] In the first embodiment, the latent parameter LV and the adaptation of the latent parameter to the uncertainty fluctuations induced by the noise signal NS are important for the correct random behavior of the output target running signal OST. When the latent parameter LV is not implemented in the Bayesian neural network (BNN), the trained BNN will consistently provide the same output for the same input, i.e., deterministic behavior, unlike the real technical system TS. However, the target running signal OST output by the Bayesian neural network (BNN) with the latent parameter LV exhibits at least partial uncertainty and has a statistical distribution that largely conforms to reality. Here, the neural weights of the Bayesian neural network (BNN) represent or model, to some extent, the deterministic components of the target running signal, i.e., the deterministic functional relationship between the input and the target running signal. In contrast, the latent parameter LV models or represents, to some extent, the random components, noise structure, or noise components.

[0043] Figure 3 The first control device CTL1 with a trained Bayesian neural network (BNN) is explained in the application stage.

[0044] In this case, the first control device CTL1 receives the current input operation signal SI from the technical system TS and feeds the current input operation signal into the input layer of the trained Bayesian neural network BNN. Here, the technical system TS can be the technical system from which the training data is sourced, or a similar or analogous technical system.

[0045] A noise signal NS, or its data value, generated by a noise generator NSG, is fed as other input data into the input layer of a trained Bayesian neural network (BNN). From the current input running signal SI and the fed noise signal NS, the trained Bayesian neural network (BNN) generates an output signal OST, at least indirectly, based on the inferred latent parameter LV. This output signal is then output as the current target running signal through the output layer of the trained Bayesian neural network (BNN). The output signal OST has a simulated or reconstructed noise component that is statistically similar to the noise component of the actual target running signal ST. Due to the random-based excitation induced by the noise signal NS, the latent parameter LV enables the Bayesian neural network (BNN) to generate or simulate realistic uncertain noise components for the output signal OST.

[0046] The generated current target running signal OST is transmitted from the trained Bayesian neural network BNN to the digital twin DT implemented in the first control device CTL1 of the technology system TS. Additionally, the current input running signal SI is also fed into the digital twin DT.

[0047] The digital twin (DT) functions as a concurrent simulator of the technical system (TS). The digital twin itself performs a simulation of the technical system (TS), which runs concurrently with the operation of the TS and is continuously updated using operating signals, here referring to the SI and OST of the technical system (TS). This real-time, simultaneous simulation of the technical system (TS) allows for detailed monitoring of the TS and the determination of state data relevant to its efficient control in a simulated manner. With the aid of this data determined in a simulated manner by the digital twin (DT), the technical system (TS) can be controlled particularly effectively and predictively—as in… Figure 3 The dotted line and arrow in the middle indicate this.

[0048] Figure 4 The training of the second control device CTL2 according to the second embodiment of the present invention is explained.

[0049] The second control device CTL2 includes a neural network NN to be trained. To train the neural network NN, the compressed input running signal CSI and the compressed target running signal CST, which are recorded in the database DB, are read out by the second control device CTL2 and fed into the input layer of the neural network NN as training data.

[0050] The neural network NN should be trained on compressed training data CSI and CST to reproduce the statistical characteristics of the statistical distribution of the random components of the compressed target running signal CST based on the input running signal.

[0051] In this embodiment, the mean AVG and variance V of the corresponding statistical distribution should be reproduced as statistical feature values ​​and output through the output layer of the neural network NN.

[0052] Within the training scope of the neural network NN, a comparison CMP is made between the compressed target running signal CST and the eigenvalues ​​AVG and VA output by the neural network NN. The comparison CMP examines whether, and to what extent, the statistical distribution of the random components of the compressed target running signal CST is compatible with or deviates from the output mean AVG and the output variance V. The corresponding deviation, or the corresponding reproduction error D, is determined as a quantitative measure of this and is incorporated back into the neural network NN for training the neural network NN—as in... Figure 4 This is indicated by the dashed arrow. Here, in particular, the so-called likelihood function can be used as the error function.

[0053] When training a neural network NN, the neural weights are adjusted to minimize the reproduction error D or the error function. Numerous efficient numerical standard methods are available for this purpose.

[0054] Figure 5 The diagram illustrates a second control device CTL2 with a trained neural network NN in the application phase of the control technology system TS. The technology system TS is preferably the same as, similar to, or a similar technology system from which the training data CSI and CST originate.

[0055] To control the technical system TS, the second control device CTL2 detects the current input operating signal SI of the technical system TS and feeds the current input operating signal SI into the input layer of a trained neural network NN. The trained neural network NN then generates statistical feature values ​​AVG and V for each dataset or each time series point, preferably for the current input operating signal SI.

[0056] Statistical feature values ​​AVG and V are output through the output layer of a trained neural network NN and fed to the noise generator NSG of the second control device CTL2. The noise generator NSG generates random data, or generally noise signals, wherein the mean and variance of the generated random data are adjustable. In this embodiment, the noise generator NSG generates one or more output values ​​for each pair of input statistical feature values ​​(here, AVG and V), the output values ​​having a predetermined mean AVG and a predetermined variance V on a statistical average. The output values ​​are output by the noise generator NSG as an output signal OST. Therefore, the output signal OST has an imitation of uncertain noise components with a realistic statistical distribution.

[0057] As in the first embodiment, the output signal OST, along with the current input signal SI, is transmitted to the digital twin DT of the technical system TS. As described above, the digital twin DT can be advantageously used to monitor or control the technical system TS.

[0058] Because the target operating signal OST realistically mimics the noise components, significantly more realistic predictions about the behavior of the technical system TS can be obtained even with lossy training data. In particular, predictions about the reliability or operational risks of the technical system TS benefit from this.

Claims

1. A computer-implemented method for mimicking the noise component of a lossily recorded operating signal of a technical system, wherein... a) Input operating signals for the control device of the technical system and target operating signals for controlling the technical system are recorded in a lossy manner. b) The neural network is trained based on the recorded running signals to reproduce the recorded target running signals and the statistical distribution of the random components of the recorded target running signals based on the recorded input running signals. c) The current input operating signal of the technical system is fed into the trained neural network. d) Generate an output signal having a noise component that mimics the statistical distribution based on the transmitted current input operating signal and noise signal, and e) The output signal is output as the current target operating signal to control the technical system.

2. The method according to claim 1, characterized in that, As the neural network, a Bayesian neural network with latent parameters representing the statistical distribution is used. The latent parameters are inferred through training. The noise signal is fed into the input layer of the Bayesian neural network, and The output signal is generated from the current input running signal and the noise signal by a Bayesian neural network trained using the inferred latent parameters.

3. The method according to claim 2, characterized in that, The inference is performed using variational inference methods and / or Markov chain Monte Carlo methods.

4. The method according to any one of claims 1 to 3, characterized in that, The neural network is trained to reproduce the statistical characteristics of the statistical distribution based on the recorded input running signals. The trained neural network determines statistical feature values ​​for the current input signal, and The noise signal is generated based on the determined statistical characteristic values ​​and is output as an output signal.

5. The method according to claim 4, characterized in that, For training purposes, the likelihood function is used as the error function to be minimized.

6. The method according to claim 4, characterized in that, The mean, variance, standard deviation, probability value, and / or distribution type of the statistical distribution are used as feature values.

7. The method according to claim 5, characterized in that, The mean, variance, standard deviation, probability value, and / or distribution type of the statistical distribution are used as feature values.

8. The method according to any one of claims 1 to 3, characterized in that, The current input signal is continuously detected and fed to the neural network, which then runs a concurrent simulator.

9. The method according to claim 8, characterized in that, A digital twin of the technology system operates with the aid of the neural network.

10. A control device for controlling a technology system, the control device being configured to implement the method according to any one of claims 1 to 9.

11. A computer program product configured to perform the method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program product according to claim 11.

Images