Mechanical anomaly explainable intelligent detection method based on algorithm unfolding for adversarial learning

Abstract

A mechanical abnormality explainable intelligent detection method based on algorithm unfolding confrontation learning is disclosed, in the method, the vibration signal y of the normal operation of the mechanical equipment is collected, and is cut into a training data set according to a fixed signal length; a deep sparse coding model is established, the vibration signal y is deeply coded, and then is mapped to a d-dimensional hidden code z; the deep sparse coding model is solved by using a proximal gradient mapping algorithm, and the optimization solving algorithm is unfolded into a sparse coding network, the reconstruction from the hidden code z to the original signal is a decoding network, and the two together constitute a generator network G; a discriminator network D is constructed to constrain the distribution of the hidden code Z and match the prior distribution p(z); the generator network G and the discriminator network D are trained by using the training data set, the network parameters are updated by using a back propagation technology, and the reconstruction error of the network to the vibration signal and the gap between the coding distribution p(z|y) and the prior distribution p(z) are reduced; the test signal is input into the trained generator network G, and the authenticity of the signal is output by the discriminator network D.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical anomaly diagnosis, and particularly relates to a mechanical anomaly explainable intelligent detection method based on algorithm unfolding adversarial learning. BACKGROUND

[0002] Anomaly detection is a technique for finding patterns of unexpected behavior in data, which has a very wide application in engineering problems. Anomaly detection in equipment prognostics and health management (PHM) helps people find the abnormal state of equipment in time and take key measures before major accidents occur, thereby ensuring the safety of people and property. Therefore, accurate and reliable mechanical anomaly detection is of great significance to ensure the safe operation of key equipment.

[0003] Traditional anomaly detection techniques, such as principal component analysis (PCA), one-class support vector machine (One-Class SVM), K-nearest neighbor algorithm (K-NN), random forest, etc., have clear theoretical framework, and the detection results have clear explainability. However, in the face of high-dimensional complex signals, their detection accuracy may decrease due to limited learning ability. With the great development of intelligent algorithms based on artificial neural networks (ANN) (such as deep learning, adversarial learning, etc.), it is possible for anomaly detection algorithms to learn low-dimensional nonlinear representation from high-dimensional data. Typical neural network-based anomaly detection algorithms, such as variational autoencoder (VAE), adversarial autoencoder (AAE) and Ganomaly, mostly use networks for encoding and decoding. However, the mechanical anomaly detection task usually takes vibration signals containing a large amount of noise as input, and in this case, the results of many network-based algorithms may be suspicious. Because these algorithms are usually constructed as black boxes, users only have access to input and output, and their explainability is not clear, which cannot guarantee that they have encoded the noise signals meaningfully. Therefore, the research on the explainability of the algorithm is crucial to improve the credibility of anomaly detection.

[0004] The explainability of machine learning algorithms can be generally divided into two categories: one is pre-explainability modeling, which means using inherently explainable techniques to model a specific problem, so that a model with explainable structure can be obtained; the other is post-explainability analysis, which refers to developing techniques to explain a trained model, usually involving many visualization methods. Both types of explainability can improve the credibility of the algorithm. Algorithm unfolding is an effective method of constructing networks, which has developed rapidly in theoretical research and practical application. Usually we unfold the solution algorithm of the model in the iteration scale, generate a network, and allow the hyperparameters in it to update in the network. Algorithm unfolding is based on a specific model, which has pre-explainability, while the partial model itself has a reconstruction mechanism, which provides post-explainability, so algorithm unfolding has great potential in building networks with better explainability. At present, an explainable anomaly detection method based on algorithm unfolding is still needed.

[0005] The above information disclosed in the background section is only intended to enhance the understanding of the background of the present application, and therefore can contain information that does not constitute prior art known to those of ordinary skill in the art in the country. SUMMARY

[0006] In view of the problems in the prior art, the present application proposes a mechanical anomaly explainable intelligent detection method based on algorithm unfolding adversarial learning, which converts the problem of solving feature signal distribution into encoding from data distribution to hidden encoding distribution and then decoding to feature distribution based on probability inference; the encoding process uses deep sparse coding, and the solution algorithm of the model is unfolded into a network, which is constrained by the adversarial training strategy to the hidden encoding distribution; the discriminator detects the state of the input signal by judging whether it is subject to normal distribution; since the network main body is unfolded by the optimization solution algorithm of a specific model, its structure has explainability, i.e., pre-explainability. The visualization in the result analysis is based on the reconfigurability of the sparse representation model, and the result has explainability, i.e., post-explainability, by reconstructing the input signal and viewing the atomic features.

[0007] The purpose of the present application is achieved by the following technical scheme, a mechanical anomaly explainable intelligent detection method based on algorithm unfolding adversarial learning comprises:

[0008] In the first step, the vibration signal y of the normal operation of the mechanical equipment is collected, which is cut into a training data set according to a fixed signal length;

[0009] In the second step, a deep sparse coding model is established, the vibration signal y is deeply coded, and then mapped to a d-dimensional hidden code z;

[0010] In the third step, the deep sparse coding model is solved by using a proximal gradient mapping algorithm, and the optimization solving algorithm is unfolded into a sparse coding network, and the reconstruction of the original signal from the hidden coding z is a decoding network, both of which constitute a generator network G together, and a discriminator network D is constructed to constrain the distribution of the hidden coding z to match the prior distribution p(z);

[0011] In the fourth step, the generator network G and the discriminator network D are trained using the training data set by setting a fixed number of iterations, and the network parameters are updated using the back propagation technique to reduce the reconstruction error of the network on the vibration signal and the gap between the coding distribution p(z|y) and the prior distribution p(z);

[0012] In the fifth step, the test signal is input into the trained generator network G, and the authenticity of the output signal is output by the discriminator network D to determine whether the mechanical equipment is normal or abnormal.

[0013] In the sixth step (S6) of the mechanical abnormality interpretable intelligent detection method based on algorithm unfolding adversarial learning, the test signal is input into the trained generator network G to reconstruct the overall feature And view the atomic features D (1，j) =D1D2…D j of each equivalent layer, and perform post-analysis.

[0014] In the mechanical abnormality interpretable intelligent detection method based on algorithm unfolding adversarial learning, the first step (S1) signal, is a noise interference, N is the length of the signal, and the vibration signal is collected by an acceleration sensor.

[0015] In the mechanical abnormality interpretable intelligent detection method based on algorithm unfolding adversarial learning, in the second step, the deep sparse coding model is:

[0016]

[0017] wherein, is the deep sparse coding to be solved, L is the number of coding layers, D i is the coding dictionary of the i-th layer, gamma i is the coding value of the i-th layer, i=1, 2,..., L, D (i，L) =D i D i+1 …D L is an equivalent dictionary multiplied by a multi-layer coding dictionary, ||·||1 represents the L1 norm, is the square of the L2 norm, in the deep sparse coding model, the first term is to ensure that the deep sparse coding a fidelity term reconstructing the feature data of the input vibration signal y, and a sparse regularization term ensuring the sparsity of each layer coding, λ i is a trade-off parameter between the data fidelity term and the sparse regularization term.

[0018] In the mechanical anomaly interpretable intelligent detection method based on algorithm unfolding adversarial learning, the coding dictionary and the coding of each layer satisfy:

[0019]

[0020] wherein, T i is the sparsity constraint constant of the i-th layer, i = 1, 2,..., L, γ i is the sparse coding value of each layer, thus D (i，L) γ L = γ i-1 , x = γ0is the ideal original feature signal.

[0021] In the mechanical anomaly interpretable intelligent detection method based on algorithm unfolding adversarial learning, the coding dictionary is a convolution dictionary, m i is the number of convolution kernels, wherein, is a circulant matrix composed of the j-th convolution kernel with a length of n i j = 1, 2,..., m i .

[0022] In the mechanical anomaly interpretable intelligent detection method based on algorithm unfolding adversarial learning, in the third step, the proximal gradient mapping algorithm of the deep sparse coding model is:

[0023] Initialize the coding value of each layer, so that wherein k = 0, 1,..., K, K is the number of algorithm iterations; i = 0, 1,..., L;

[0024] For each iteration k = 0, 1,..., K, the coding value of each layer i = 0, 1,..., L is solved in turn, and the solving formula is as follows:

[0025]

[0026] Output the deep sparse coding value

[0027] wherein, Soft T (·) is a soft threshold function, and its expression is Soft T (·) = max(|·|-T, 0) * sign(·), sign(·) is a sign function, and T is a threshold value, in the i-th layer, T = μi λ i , μ i is the step size of each iteration of the proximal gradient mapping algorithm.

[0028] In the mechanical anomaly interpretable intelligent detection method based on algorithm unfolding adversarial learning, in the deep sparse coding model, the coding dictionary D i is equivalent to the transpose convolution network layer, is equivalent to the convolution network layer, and the soft threshold function is the activation function between layers.

[0029] In the mechanical anomaly interpretable intelligent detection method based on algorithm unfolding adversarial learning, in the third step (S3), two coding matrices D μ , D σ respectively obtain the mean and standard deviation vector of the coding distribution p(z|y), and the coding matrix D γ reconstructs the deep coding wherein is the estimated value of the original feature signal x. The prior distribution p(z) is an independent Gaussian distribution, and the discriminator network D used in the adversarial training strategy is a full connection network, and the activation function thereof is selected as LeakyReLU.

[0030] In the mechanical anomaly interpretable intelligent detection method based on algorithm unfolding adversarial learning, in the fourth step, the loss function of the adversarial training is defined as:

[0031]

[0032]

[0033] wherein represents the expectation value of y and z, is the loss function of the generator network G, is the loss function of the discriminator network D, f(·) represents the discriminator network D, G e is the encoder part in the generator network G, in the composition of the generator loss function, is the adversarial loss, used to judge the gap between the coding distribution p(z|y) and the prior distribution p(z), is the reconstruction loss, w adv and w con are parameters for adjusting the proportion of the two, and the adversarial loss and the reconstruction loss are defined as follows:

[0034]

[0035] wherein represents the expectation value of y.

[0036] Compared with the prior art, the present application has the following advantages: the present application can quickly and effectively realize the identification of abnormal states by constructing a deep encoding model, expanding it into a network, and combining an adversarial training strategy, can improve the accuracy and reliability of abnormal detection, has the characteristics of interpretability in both structure and result, and is beneficial to safe, reliable, and transparent fault prediction and health management of mechanical equipment. BRIEF DESCRIPTION OF DRAWINGS

[0037] Various other advantages and benefits of the present application will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are intended to only illustrate preferred embodiments of the present application and are consequently not to be considered as limiting its scope. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained from these drawings without creative labor for those of ordinary skill in the art. Moreover, the same reference numerals are used to denote the same components throughout the drawings.

[0038] In the drawings:

[0039] Figure 1 is a step schematic diagram of a mechanical abnormality interpretable intelligent detection method based on algorithm expansion adversarial learning according to an embodiment of the present application;

[0040] Figure 2 is a schematic diagram of a Qianpeng gear test bench according to an embodiment of the present application;

[0041] Figures 3(a) to 3(h) is a signal time domain vibration diagram of normal state and abnormal state according to an embodiment of the present application, Fig. 3(a) is a normal state signal, Fig. 3(b) is a 1mm single-sided crack abnormal state signal, Fig. 3(c) is a 2mm single-sided crack abnormal state signal, Fig. 3(d) is a 1mm double-sided crack abnormal state signal, Fig. 3(e) is a 3mm single-sided crack abnormal state signal, Fig. 3(f) is a 4mm single-sided crack abnormal state signal, Fig. 3(g) is a 0.5mm single-sided wear abnormal state signal, and Fig. 3(h) is a 0.5mm double-sided wear abnormal state signal;

[0042] Figure 4 is a network structure diagram of a mechanical abnormality interpretable intelligent detection method based on algorithm expansion adversarial learning according to an embodiment of the present application;

[0043] Figure 5 is a performance comparison (including detection rate, false alarm rate, AUC area, F1 score, and accuracy) between a mechanical abnormality interpretable intelligent detection method based on algorithm expansion adversarial learning according to an embodiment of the present application and other methods;

[0044] Figures 6(a) to 6(d)This is a schematic diagram comparing the ability of an algorithm-based adversarial learning-based intelligent detection method for mechanical anomalies, according to one embodiment of the present invention, to distinguish between normal and abnormal states with other methods.

[0045] Figures 7(a) to 7(d) This is an embodiment of the present invention of an algorithm-based adversarial learning-based intelligent detection method for mechanical anomalies, and the reconstruction results of normal signals by other methods;

[0046] Figure 8 This is a schematic diagram of the learned atomic features of an algorithm-based adversarial learning-based intelligent detection method for mechanical anomalies according to an embodiment of the present invention.

[0047] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0048] The following will refer to the appendix. Figures 1 to 8 Specific embodiments of the invention will be described in more detail below. While specific embodiments of the invention are shown in the accompanying drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0049] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0050] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0051] To better understand, such as Figures 1 to 8 As shown, the intelligent detection method for mechanical anomalies based on algorithmic adversarial learning includes:

[0052] In the first step S1, the vibration signal y of the mechanical equipment during normal operation is collected and clipped into a training dataset according to a fixed signal length.

[0053] In the second step S2, a deep sparse coding model is established to perform deep coding on the signal y and then map the signal y to a d-dimensional hidden code z;

[0054] In the third step S3, a proximal gradient mapping algorithm is used to solve the established deep sparse coding model, and the optimization solving algorithm is unfolded as a sparse coding network, and the reconstruction of the original signal from the hidden code z is a decoding network, and the two together constitute a generator network G. A discriminator network D is constructed to constrain the distribution of the hidden code z and match the prior distribution p(z);

[0055] In the fourth step S4, the constructed adversarial generative network model is trained using the collected normal samples, and the network parameters are updated using the back propagation technique to reduce the reconstruction error of the network on the signal and the gap between the encoding distribution p(z|y) and the prior distribution p(z);

[0056] In the fifth step S5, the test signal is input into the trained network, and the authenticity of the output signal is output by the discriminator network D, i.e., whether it belongs to normal or abnormal;

[0057] In the sixth step S6, the overall features of the reconstructed signal and the atomic features learned by the network are viewed for post-hoc interpretability analysis.

[0058] In the preferred embodiment of the method, in the first step S1, the vibration signal can be decomposed as y=x+e, where is a characteristic signal, is a noise interference, and N is the length of the signal. The vibration signal is collected by an acceleration sensor.

[0059] In the preferred embodiment of the method, in the second step S2, the established deep sparse coding model is:

[0060]

[0061] wherein is the deep sparse code obtained by solving, L is the number of coding layers, D i is the coding dictionary of the i-th layer, γ i is the coding value of the i-th layer, i=1, 2,..., L, D (i，L) =D i D i+1 …D L is the equivalent dictionary of the multiplication of the multi-layer coding dictionaries. ||·||1 represents the L1 norm, is the square of the L2 norm.

[0062] In the deep sparse coding model, the first term is a data fidelity term that ensures that the deep sparse code reconstruct the input signal y; the rest are sparse regularization terms, which ensure that the encoding of each layer is sparse. λ i is a trade-off parameter between the data fidelity term and the sparse regularization term.

[0063] In a preferred embodiment of the method, the encoding dictionary of each layer and the encoding satisfy:

[0064]

[0065] where T i is the sparsity constraint constant of the i-th layer, i = 1, 2,..., L, γ i is the sparse encoding value of each layer. Therefore, D (i，L) γ L = γ i-1 , which specifies that x = γ0is the ideal original feature signal.

[0066] In a preferred embodiment of the method, the encoding dictionary is a convolution dictionary, m i is the number of convolution kernels. wherein, is a circulant matrix composed of the j-th convolution kernel with a length of n i j = 1, 2,..., m i .

[0067] In a preferred embodiment of the method, in the third step S3, the proximal gradient mapping algorithm of the deep sparse coding model is:

[0068] Initialize the encoding value of each layer so that where k = 0, 1,..., K, K is the number of iterations of the algorithm; i = 0, 1,..., L;

[0069] For each iteration k = 0, 1,..., K, the encoding value of each layer i = 0, 1,..., L is solved in turn, and the solving formula is as follows:

[0070]

[0071] Output the deep sparse coding value The above algorithm constitutes the main structure of the sparse coding network.

[0072] wherein, Soft T (·) is a soft threshold function, and its expression is Soft T (·) = max(|·|-T, 0) * sign(·), sign(·) is a sign function, and T is a threshold value, in the i-th layer, T = μ i λ i , μi Step size of each iteration of proximal gradient mapping algorithm.

[0073] In the preferred embodiment of the method, in the sparse coding network, D i Multiplying with vector is equivalent to transposed convolution network layer, Multiplying with vector is equivalent to convolution network layer, and the soft threshold function is the activation function between layers.

[0074] In the preferred embodiment of the method, in the third step S3, two encoding matrices D μ , D σ respectively obtain the mean and standard deviation vector of the hidden encoding distribution p(z|y), and the encoding matrix D γ is used to reconstruct the deep encoding Due to the reconstructability of the sparse representation framework, the reconstruction process from the deep encoding to the feature signal does not need to build another network, wherein is the estimated value of the original feature signal x. The prior distribution p(z) is an independent Gaussian distribution, and the discriminator network D used in the adversarial training strategy is a fully connected network, and the activation function thereof is selected as LeakyReLU.

[0075] In the preferred embodiment of the method, in the fourth step S4, the loss function of the adversarial training is defined as:

[0076]

[0077]

[0078] wherein represents the expectation value of y and z, is the loss function of the generator, is the loss function of the discriminator, f(·) represents the discriminator network, G e is the encoder part in the generator network. In the composition of the generator loss function, is the adversarial loss, used to judge the gap between the posterior distribution p(z|y) and the prior distribution p(z), is the reconstruction loss, used to make the generated signal and the original signal close enough, w adv and w con are parameters for adjusting the proportion of the two. The definitions of the adversarial loss and the reconstruction loss are as follows:

[0079]

[0080]

[0081] wherein E[y] represents the expectation value of y.

[0082] In a preferred embodiment of the method, in the sixth step S6, the network model that has completed training is input with a test set signal to reconstruct the overall feature and view the atomic features of each equivalent layer, that is, D (1，j) = D1D2…D j , j = 1, 2,..., L, and perform post-hoc interpretability analysis.

[0083] In order to further understand the present application, in one embodiment, Figure 1 FIG. 1 is a schematic diagram of steps of a mechanical anomaly interpretable intelligent detection method based on algorithm unfolding adversarial learning, including the following steps:

[0084] S1: Collecting vibration signals y of a mechanical device in normal operation, and cutting them into training data sets according to a fixed signal length;

[0085] S2: Establishing a deep sparse coding model, performing deep coding on the signal y, and then mapping to d-dimensional hidden coding z;

[0086] S3: Using a proximal gradient mapping algorithm to solve the established deep sparse coding model, and unfolding the optimization solving algorithm into a sparse coding network, and the reconstruction from hidden coding z to the original signal as a decoding network, both of which together constitute a generator network G, and a discriminator network D is constructed to constrain the distribution of hidden coding z and match the prior distribution p(z);

[0087] S4: Training the generator network G and the discriminator network D using the training data set by setting a fixed number of cycles, updating the network parameters using the back propagation technology, and reducing the reconstruction error of the network on the signal and the gap between the coding distribution p(z|y) and the prior distribution p(z);

[0088] S5: Inputting a test signal into the trained network, and outputting the authenticity of the signal by the discriminator network D, that is, whether it belongs to normal or abnormal;

[0089] S6: Viewing the overall feature of the reconstructed signal and the atomic feature learned by the network, and performing post-hoc interpretability analysis.

[0090] The above embodiment constitutes a complete technical solution of the present application. Unlike the prior art, the structure of the abnormal intelligent detection network built by the above embodiment is unfolded from the solving algorithm of the deep sparse coding model, and the hidden coding distribution is constrained by means of the adversarial training strategy. This model can inherit the interpretability of traditional models, and at the same time, use network training to better mine the features of normal state signals, and improve the precision and reliability of mechanical anomaly detection.

[0091] Figure 2is a schematic diagram of the Qianpeng gear test bench. The test system is mainly composed of a variable speed drive motor, a parallel shaft gearbox, a shaft, a speed regulator and the like, and various faults are quickly simulated through the organic combination of components. The test system includes an eddy current acceleration sensor, a laser speed sensor key signal, a Premax Yiheng data acquisition system, and a ThinkPad Lenovo notebook. It is used for helical gear wear and crack fault simulation experiments, in which the parameters of the helical gear are as follows: the number of gear teeth of the input shaft gear Z1 = 53, the number of gear teeth of the output shaft gear Z2 = 75, the modulus m = 2, the tooth width b = 20 mm, and the helix angle β is 10.0633°. A total of 1 normal state, 7 fault states, 1, 2, 3, 4 mm deep single-tooth crack, 2 mm deep multi-tooth crack, 0.5 mm deep single-tooth uniform wear, and 0.5 mm deep multi-tooth uniform wear data are collected.

[0092] In this embodiment, in step S1, the vibration signal is collected by an eddy current acceleration sensor, the input shaft speed is 1000 r / min, the sampling frequency is 10240 Hz, the sampling time is 1.28 min, the signal is intercepted using a window length of 1024 points with a sliding window ratio of 0.8, and the data samples of the normal and abnormal states are obtained. The ratio of normal samples to abnormal samples is set to 5:1, the ratio of training samples to test samples is set to 1:1, and finally 3090 normal samples are obtained for training, and 3090 normal samples and 1159 abnormal samples are obtained for testing. Figures 3(a) to 3(h) The time domain waveform diagram of the vibration signal under the normal operation state and the abnormal operation state is shown.

[0093] In this embodiment, in step S2, a 4-layer deep coding model is established:

[0094]

[0095] The input signal is deep coded, and then transformed into 10-dimensional hidden coding z using two weight matrices D μ , D σ The decoding part first restores the low-dimensional hidden coding to deep coding using the weight matrix D γ Then, according to the reconstruction characteristics of sparse representation, the input signal is reconstructed:

[0096]

[0097] Both constitute an autoencoder, that is, the generator part in the generative adversarial network.

[0098] ​In this embodiment, in step S3, by unfolding the deep sparse coding model, the sparse unfolding network obtained can learn the numerous parameters in the model end to end, wherein the parameter settings within the model are shown in Table 2. With the strategy of adversarial training, the latent encoding is subjected to a set Gaussian prior distribution, which is realized by a discriminator. The overall structure of the network is shown in Figure 4

[0099] Table 2 Structure parameters of the algorithm unfolding adversarial learning network

[0100]

[0101]

[0102] In this embodiment, in step S4, the algorithm unfolding adversarial learning anomaly intelligent detection network AAU-Net is implemented by using PyTorch. When the generator is updated, the reconstruction error is reduced, and at the same time, the discriminator judges the posterior distribution and the prior distribution as true and false. When the discriminator is updated, the posterior distribution and the prior distribution are judged as false and true, so as to realize the adversarial training. In the training process, the parameters of the network training are set as follows: the weights of the adversarial loss and the reconstruction loss are w adv = 0.001, w con = 0.999, the iteration is 15 times, the batch size is 64 samples in each iteration process, the learning rate of the generator is 5e-4, the learning rate of the discriminator is 2e-4, the optimizer is Adam, and the two indicators of the momentum of the optimizer are 0.5 and 0.999 respectively. The discriminator is updated once every 5 times of updating the generator.

[0103] In this embodiment, in step S5, the test sample is input into the algorithm unfolding adversarial learning anomaly detection network. Through the deep sparse coding and the mapping to the low-dimensional latent space, the discriminator finally judges whether the latent encoding of the sample is subjected to the normal distribution, so as to realize the purpose of anomaly detection. The comparison experiments with other neural network-based anomaly detection methods further illustrate the technical scheme of the present application. Figure 5 As shown in Table 3, the detection rate, the false alarm rate, the AUC area, the F1 score and the accuracy of various methods on the Qianpeng gear data can be compared. It can be seen that the present method has the highest detection rate, false alarm rate, AUC area and F1 score, and the lowest false alarm rate, which is superior to other methods. At the same time, the discrimination ability of various methods for normal and abnormal data is reflected by the scoring results of the discriminator on all test data, as shown in Figures 6(a) to 6(d) Fig. 3, the present method cannot successfully distinguish the normal and abnormal states from the abnormal score values, and the distributions of the two types of data are more concentrated within the class and more separated between the classes.

[0104] ​In this embodiment, in step S6, the reconfigurability of the sparse representation framework is utilized to perform post-hoc interpretability analysis by reconstructing the input signal and examining the learned dictionary atoms. The sparse expansion network learns the characteristics of normal operation data using normal samples for training, and thus can reconstruct normal samples well, Figures 7(a) to 7(d) For the overall signal reconstruction results of the method AAU-Net and other methods, all the methods have a certain reconstruction ability due to the structure containing the autoencoder, but the reconstruction result of the method contains less noise and can better reflect the characteristics of the input signal. The equivalent dictionary atoms at each level are extracted from the network trained by the method, as shown in Figure 8 The first two levels of dictionary convolution kernel length are still very short, and only some local features are learned, while the convolution kernel length of the third and fourth levels of equivalent dictionary atoms can cover the basic characteristics of the signal, including the double frequency characteristics and the six times frequency characteristics.

[0105] Although the embodiments of the present application are described above with reference to the drawings, the present application is not limited to the specific embodiments and application fields described above, and the specific embodiments described above are only illustrative and guiding, but not limiting. Those skilled in the art can make many forms under the guidance of this specification and without departing from the scope protected by the claims of the present application, which all belong to the protection of the present application.

Claims

1. A mechanical anomaly explainable intelligent detection method based on algorithm unfolding confrontation learning, characterized in that, It comprises the following steps, In a first step (S1), a vibration signal of a normal operation of the mechanical equipment is collected The vibration signal is cut into a training data set according to a fixed signal length. In the second step (S2), a deep sparse coding model is established to perform deep coding on the vibration signal and then map to dimensional latent coding . In a third step (S3), the deep sparse coding model is solved using a proximal gradient mapping algorithm, and the optimization solving algorithm is unfolded as a sparse coding network, from the hidden encoding to the reconstruction of the original signal as a decoding network, both of which together constitute a generator network , a discriminator network is constructed to constrain the distribution of the hidden encoding , matching the prior distribution ; In the fourth step (S4), the generator network is trained using the training dataset by setting a fixed number of iterations. and discriminator network Backpropagation technology is used to update network parameters, reducing the network's reconstruction error and coding distribution of vibration signals. With prior distribution The difference, the loss function for adversarial training is defined as: ， ， wherein denotes the expectation over and denotes the expectation value, is the loss function of the generator network , is the loss function of the discriminator network , denotes the discriminator network , is the encoder part in the generator network , in the composition of the generator loss function, is the adversarial loss, to judge the gap between the encoding distribution and the prior distribution , is the reconstruction loss, is a parameter to adjust the proportion of the two, and the definitions of the adversarial loss and the reconstruction loss are as follows: ， ， in Indicates to To find the expected value, during the training process, the network training parameters are set as follows: the weights of the adversarial loss and the reconstruction loss are... The algorithm iterates 15 times, with a batch size of 64 samples per iteration. The learning rate of the generator is 5e-4, the learning rate of the discriminator is 2e-4, and the optimizer is Adam. The two momentum metrics of the optimizer are 0.5 and 0.999, respectively. The discriminator is updated once every 5 generator iterations. In a fifth step (S5), the test signal is input into the trained generator network The discriminator network outputs the authenticity of the signal to determine whether the mechanical equipment is normal or abnormal.

2. The method of claim 1, wherein, In a sixth step (S6), the test signal is input into the trained generator network to reconstruct the overall features and to view the atomic features of the respective equivalent layers with the convolution kernels for post-analysis.

3. The method of claim 1, wherein, In a first step (S1), the vibration signal is decomposed into wherein is a characteristic signal, is a noise disturbance, is a signal length, the vibration signal being acquired by an acceleration sensor.

4. The method of claim 1, wherein, In the second step (S2), the deep sparse coding model is: ， wherein, is the deep sparse coding to be solved, is the number of encoding layers, is the encoding dictionary of the th layer, is the encoding value of the th layer, , is the equivalent dictionary of the multiplication of the multi-layer encoding dictionaries, denotes the L1 norm, is the square of the L2 norm, in the deep sparse coding model, the first term is to ensure that the input can be reconstructed from the deep sparse coding vibration signal characteristic data fidelity term, the remaining terms are sparse regularization terms to ensure that the encoding of each layer is sparse, is the trade-off parameter between the data fidelity term and the sparse regularization term.

5. The method of claim 4, wherein, The layers coding dictionary and coding satisfy: ， wherein, is the first is the sparsity constraint constant for the layer, , is the sparsity encoding value for each layer, thus having , is the ideal original feature signal.

6. The method of claim 5, wherein, coding dictionary is a convolution dictionary, is a number of convolution kernels, wherein, is a cyclic matrix composed of a shift of the th convolution kernel along the column direction, is a cyclic matrix composed of a shift of the th convolution kernel along the column direction.

7. The method of claim 6, wherein, In the third step (S3), the proximal gradient mapping algorithm of the deep sparse coding model is: The encoding values of the layers are initialized such that wherein , is the number of iterations of the algorithm; , ; For each iteration , the encoding values of each layer are solved in turn, with the following formula: , Output deep sparse code values , wherein is a soft threshold function, whose expression is , is a sign function, is a threshold value, in the first layer, , is the step size of each step iteration of the proximal gradient mapping algorithm.

8. The method of claim 7, wherein, In a deep sparse coding model, the coding dictionary is equivalent to a transposed convolutional network layer, is equivalent to a convolutional network layer, and the soft threshold function is the activation function between layers.

9. The method of claim 8, wherein, In the third step (S3), the two encoding matrices are used to obtain the mean and standard deviation vectors of the encoding distribution respectively, and the deep encoding is reconstructed as , , where is the estimate of the original feature signal ; the prior distribution is an independent Gaussian distribution, and the discriminator network used in the adversarial training strategy is a fully connected network with LeakyReLU as the activation function.

Images