A small sample fault diagnosis method based on improved TimeGAN model

Abstract

The present invention discloses a small sample fault diagnosis method based on an improved TimeGAN model, comprising: collecting elevator operation data, including normal samples and fault samples; establishing an improved TimeGAN model; inputting fault samples into the improved TimeGAN model to generate fault data; merging the elevator operation data and the generated fault data into a sample set, and extracting the time-frequency features of the sample set by using wavelet packet decomposition; dividing the extracted time-frequency features into a training set and a test set, and inputting the training set into a CNN model for training; inputting the test set into the trained CNN model, obtaining fault diagnosis results and verifying the fault diagnosis results. The method generates fault samples with time series features, reduces the imbalance of original data distribution, and fully extracts the time-frequency features of samples to obtain efficient and accurate fault diagnosis results, thereby realizing safe maintenance of intelligent equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular relates to a small sample fault diagnosis method based on an improved TimeGAN model. Background Art

[0002] In recent years, with the rapid economic development, the use of daily electrical equipment by residents has also increased significantly, such as elevators. Faults exist in every device that people use daily. Many devices have complex structures that are difficult to troubleshoot and have great potential threats. Once a fault occurs, it will inevitably cause huge losses. In addition, problems such as the scarcity of fault samples and the complexity of vibration signals also reduce the accuracy of fault diagnosis. Therefore, it is of great significance to diagnose and predict potential faults in the daily use of elevators and prevent the occurrence of faults in a timely manner to maintain the safety of residents' lives and property.

[0003] At present, with the development of big data and artificial intelligence technology, fault diagnosis has also been widely studied. However, artificial intelligence algorithm models are data-driven, so the combination of relevant algorithms to achieve fault diagnosis and prediction often requires massive amounts of data, especially fault data. The particularity of elevator equipment makes it difficult to collect fault data, and it is impossible to ensure that there are enough training samples to support model training, so the relevant research on sample generation and expansion is becoming more and more extensive. Generative models such as generative adversarial networks (GAN), least squares generative adversarial networks (LSGAN), and deep convolutional generative adversarial networks (DCGAN) can generate the required images, so they are widely used in image expansion and restoration. Since the samples in fault diagnosis are usually one-dimensional data, related studies first convert the samples into two-dimensional images, then use the generative model to expand the images, and finally input them into the relevant algorithms for training to obtain the diagnosis results. These methods solve the problem of insufficient training samples, but they do not fully consider the unique time correlation between vibration signals. Traditional generative models cannot ensure that the network can dynamically and effectively capture the step-by-step dependencies in the signal, nor do they fully utilize the time, frequency and spatial domain characteristics of vibration signals. Therefore, it is urgent to develop new elevator fault diagnosis methods to achieve efficient and accurate fault diagnosis and intelligent equipment safety maintenance. Summary of the invention

[0004] The purpose of the present invention is to propose a small sample fault diagnosis method based on an improved TimeGAN model to address the above problems. This method generates fault samples with time series characteristics, reduces the imbalance of the original data distribution, and fully extracts the time-frequency characteristics of the samples to obtain efficient and accurate fault diagnosis results, thereby realizing safe maintenance of intelligent equipment.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] The present invention proposes a small sample fault diagnosis method based on an improved TimeGAN model, comprising the following steps:

[0007] S1. Collect elevator operation data, which includes normal samples and fault samples;

[0008] S2. Establish an improved TimeGAN model. The improved TimeGAN model includes an embedding network, a recovery network, a sequence generator, and a sequence discriminator, where:

[0009] The embedding network transforms static features and temporal features into potential representations using an embedding function, which is implemented based on a recurrent network.

[0010] The restoration network uses the restoration function to restore the potential representation to the original static and temporal features. The restoration function is implemented based on the feedforward network.

[0011] The sequence generator samples and extracts static random vectors and temporal random vectors from the elevator operation data, and uses a generating function to convert the static random vectors and temporal random vectors into synthetic latent codes. The generating function is implemented based on a recurrent network.

[0012] The sequence discriminator receives the synthetic latent code using a discriminant function and returns the classification result. The discriminant function is implemented based on a bidirectional recurrent network with a feedforward output layer.

[0013] The loss function of the improved TimeGAN model includes the reconstruction loss l R , fight against loss U and supervision loss l S , and construct the following objective function:

[0014]

[0015]

[0016]

[0017]

[0018] Among them, θ e ,θr ,θ g ,θ d represents the parameters of the embedding network, the recovery network, the sequence generator and the sequence discriminator respectively, λ≥0, η≥0, s is the static feature vector, is the restored static eigenvector, x t is the temporal feature vector at time t, is the temporal feature vector at time t after recovery, y S is the real static feature classification result, y t is the real temporal feature classification result at time t, is the generated static feature classification result, is the temporal feature classification result generated at time t, h S is the latent static representation, h t is the potential tense representation at time t, h t-1 is the potential temporal representation at time t-1, z t is the temporal random vector at time t, g χ Generate a network for temporal features, is subject to random variables s,x 1:T The expectation of a probability distribution P, is subject to random variables s,x 1:T Generate probability distribution expectations;

[0019] S3, input the fault samples into the improved TimeGAN model to generate fault data;

[0020] S4. The elevator operation data and the generated fault data are combined into a sample set, and the time-frequency characteristics of the sample set are extracted by wavelet packet decomposition. The wavelet packet decomposition includes decomposition operation and reconstruction operation, which are as follows:

[0021] 1) Assume the input signal is a discrete signal The decomposition operation is calculated as follows:

[0022]

[0023] Among them, n is the discrete time series, k is the time shift factor, N is the scale factor, and i is the number of layers. is the component to be decomposed (0≤i≤2 N -1), is the wavelet packet coefficient at low frequency, is the wavelet packet coefficient at high frequency, h 0 (·) is a low-pass filter, g 0 (·) is a high-pass filter;

[0024] 2) The reconstruction operation is calculated as follows:

[0025]

[0026] in, is the low-frequency signal component, is the high frequency signal component, h 1 (·) is a high-pass filter, g 1 (·) is a low-pass filter;

[0027] S5, dividing the extracted time-frequency features into a training set and a test set, and inputting the training set into the CNN model for training to obtain a trained CNN model;

[0028] S6. Input the test set into the trained CNN model to obtain the fault diagnosis results and verify the fault diagnosis results.

[0029] Preferably, the CNN model includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a first fully connected layer, a second fully connected layer and an output layer, which are connected in sequence.

[0030] Preferably, each pooling layer is a maximum pooling layer, and the output end of each convolutional layer and the fully connected layer is also provided with an activation function, the activation function is a ReLU function, and the output layer uses a SoftMax classifier.

[0031] Preferably, the wavelet packet decomposition adopts the db3 wavelet basis.

[0032] Preferably, the objective function satisfies λ=1,η=10.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] This method introduces the least squares loss function into the time series generative adversarial network (TimeGAN) to improve the objective function of the model, generate fault samples with time series characteristics, reduce the imbalance of the original data distribution, and solve the problem of insufficient samples; and use wavelet packets to decompose the expanded samples, calculate the wavelet packet coefficients of each node and reconstruct the signal to fully extract the time-frequency characteristics of the samples; finally, the powerful spatial feature extraction ability of the CNN model is combined to extract the spatial characteristics of the signal, establish a fault diagnosis model, and realize efficient and accurate fault diagnosis, thereby realizing the safe maintenance of intelligent equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flow chart of the small sample fault diagnosis method based on the improved TimeGAN model of the present invention;

[0036] Figure 2 A schematic diagram of time-frequency characteristics extracted by wavelet packet decomposition of the present invention;

[0037] Figure 3 It is a schematic diagram of the CNN model structure of the present invention;

[0038] Figure 4 This is a comparison chart of the accuracy of the five methods proposed in this invention, BP, LSTM, GAN+CNN and WPD+BP. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0040] It should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0041] like Figure 1-4 As shown in FIG. 1 , a small sample fault diagnosis method based on an improved TimeGAN model includes the following steps:

[0042] S1. Collect elevator operation data, which includes normal samples and fault samples. The elevator operation data can be obtained by collecting acceleration sensors.

[0043] S2. Establish an improved TimeGAN model. The improved TimeGAN model includes an embedding network, a recovery network, a sequence generator, and a sequence discriminator, where:

[0044] The embedding network transforms static features and temporal features into potential representations using an embedding function, which is implemented based on a recurrent network.

[0045] The restoration network uses the restoration function to restore the potential representation to the original static and temporal features. The restoration function is implemented based on the feedforward network.

[0046] The sequence generator samples and extracts static random vectors and temporal random vectors from the elevator operation data, and uses a generating function to convert the static random vectors and temporal random vectors into synthetic latent codes. The generating function is implemented based on a recurrent network.

[0047] The sequence discriminator receives the synthetic latent code using a discriminant function and returns the classification result. The discriminant function is implemented based on a bidirectional recurrent network with a feedforward output layer.

[0048] The loss function of the improved TimeGAN model includes the reconstruction loss l R , fight against loss U and supervision loss l S , and construct the following objective function:

[0049]

[0050]

[0051]

[0052]

[0053] Among them, θ e ,θ r ,θ g ,θ d represents the parameters of the embedding network, the recovery network, the sequence generator and the sequence discriminator respectively, λ≥0, η≥0, s is the static feature vector, is the restored static eigenvector, x t is the temporal feature vector at time t, is the temporal feature vector at time t after recovery, y S is the real static feature classification result, y t is the real temporal feature classification result at time t, is the generated static feature classification result, is the temporal feature classification result generated at time t, h S is the latent static representation, h t is the potential tense representation at time t, h t-1 is the potential temporal representation at time t-1, z t is the temporal random vector at time t, g χ Generate a network for temporal features, is subject to random variables s,x 1:T The expectation of a probability distribution P, is subject to random variables s,x 1:T Generate probability distribution expectations.

[0054] In one embodiment, the objective function satisfies λ=1, η=10.

[0055] Among them, the method obtains the LS-Time GAN model (i.e., the improved TimeGAN model) by combining the least square loss function in the TimeGAN model and the least square generative adversarial networks (LSGAN). The improved TimeGAN model includes: an embedding network, a recovery network, a sequence generator, and a sequence discriminator. The embedding network and the recovery network are jointly trained as autoencoder components with the sequence generator and the sequence discriminator as adversarial components, so that the model can simultaneously learn the feature distribution and the potential relationship between variables in the process of iteration over time. The embedding network provides a latent space (Latent Space) for the adversarial network (sequence generator and sequence discriminator), and the potential dynamic correlation (Temporal Dynamics) of the real data and the generated data is synchronized through the supervised loss, as follows:

[0056] 2.1) Embedding and restoring networks

[0057] The embedding function and the recovery function provide a mapping between features and latent space, allowing the adversarial network to learn the underlying temporal dynamics between data through low-dimensional representations.

[0058] Let H S , H χ are the latent vector spaces corresponding to the static feature vector space S and the temporal feature vector space χ, respectively, and the embedding function e:S×Π t χ→H S ×Π t H χ Convert static features and temporal features into latent representations (LatentRepresentations) S ,h 1:T =e(s,x 1:T ), the embedding function e is implemented through a recurrent network:

[0059] h S =e S (s), h t =e χ (h S ,h t-1 ,x t ) (1)

[0060] Among them, e S :S→H S is the embedding network of static features, e χ :H S ×H χ ×χ→Hχ is the embedding network of temporal features, h S is the latent static representation, h t is the potential tense representation at time t, h 1:T is a set of potential temporal representations in the interval [1, T], where T is time, s is a static feature vector, and x 1:T is the set of temporal feature vectors in the interval [1,T], h t-1 is the potential temporal representation at time t-1, x t is the temporal feature vector at time t.

[0061] In contrast, the recovery function r:H S ×Π t H χ →S×Π t χ will potentially represent h S ,h 1:T =e(s,x 1:T ) are restored to their original static and temporal features, and the results are expressed as The recovery function r is implemented through a feedforward network:

[0062]

[0063] Among them, r S :H S →S is the recovery network of static feature embedding, r χ :H χ →χ is the restoration network for temporal feature embedding, is the restored static feature vector, is the temporal feature vector at time t after recovery, is the set of temporal feature vectors of the [1,T] interval after recovery.

[0064] 2.2) Sequence Generator and Sequence Discriminator

[0065] The sequence generator does not directly output the generated data in the feature space, but first outputs it to the output space of the embedding network. Let Z S ,Z χ Represent the static feature vector space and temporal feature vector space of known distribution respectively, and extract random vectors from them as static random vectors and temporal random vectors as input to generate H S ,H χ .

[0066] Then generate the function g:Z S × t Z χ →H S × t H χConvert static and temporal random vectors to synthetic latent codes The generating function g is implemented through a recurrent network:

[0067]

[0068] Among them, g S :Z S →H S For static feature generation network, g χ :H S ×H χ ×Z χ →H χ Generate a network for temporal features, To synthesize latent static representation, is the set of synthetic temporal feature vectors in the interval [1, T], z S is a static random vector, z 1:T is a set of temporal random vectors in the interval [1,T], is the synthetic latent temporal representation at time t, is the synthetic latent temporal representation at time t-1, z t is a temporal random vector at time t, following a random process.

[0069] Finally, the sequence discriminator also operates from the output space of the embedding network. The discriminant function d:H S × t H χ →[0,1]×П t [0,1] receives the synthesized latent code and returns the classification result d is implemented via a bidirectional recurrent network with a feed-forward output layer:

[0070]

[0071] in, and denote the forward and backward hidden state sequences, respectively, is the forward loop function, is the backward loop function, d S is the static feature output layer classification function, d χ is the classification function of the temporal feature output layer, is the static feature classification result, is the temporal feature classification result at time t, is the set of temporal feature classification results in the interval [1, T], Represents h S or Represents h 1:T or Right now The real data (h * ) or synthetic data embedding; similarly, Represents the real data (y * ) or synthetic data The classification result is shown in Figure 2. The * represents an arbitrary subscript, and the real data (original data) is the collected elevator operation data.

[0072] 2.3) Objective Function

[0073] As a reversible mapping between feature space and latent space, the embedding function should be able to transform the original data s, x 1:T Extract its potential representation h S ,h 1:T , the recovery function can be obtained from its potential representation h S ,h 1:T Accurately reconstruct the characteristic distribution of the original data The overall definition of the objective function is:

[0074]

[0075] Among them, θ e ,θ r ,θ g ,θ d They represent the parameters of the embedding network, the recovery network, the sequence generator, and the sequence discriminator, respectively. λ≥0, η≥0 are hyperparameters for balancing the two losses, respectively. In this embodiment, λ=1, η=10.

[0076] Among them, the reconstruction loss (Reconstruction Loss) is calculated as follows:

[0077]

[0078] The adversarial loss uses the least squares loss function, which is calculated as follows:

[0079]

[0080] The existing adversarial loss adopts cross entropy loss, which only focuses on whether the classification result is correct, but not the distance, so it is easy to reach saturation. The least squares loss only reaches saturation at one point, which can make the training process more stable and generate higher quality data.

[0081] Apply the maximum likelihood method to define the supervised loss:

[0082]

[0083] Among them, when the adversarial loss lU Push the generator to create realistic sequences, and the supervision loss l S Further ensure that it produces similar step-by-step dependencies.

[0084] S3. Input the fault samples into the improved TimeGAN model to generate fault data.

[0085] Specifically, Table 1 shows the discriminative scores and predictive scores of the three methods: improved TimeGAN, TimeGAN, and 1DGAN. The discriminative score and predictive score are indicators that qualitatively measure the difference between the generated data and the original data. The smaller the value, the closer the two are.

[0086] Table 1

[0087]

[0088] As can be seen from Table 1, compared with other methods, the improved TimeGAN of this application scores lower in the discrimination score and prediction score of original data and generated data, indicating that the data it generates has the smallest gap with the original data and has better quality and practicality.

[0089] S4. The elevator operation data and the generated fault data are combined into a sample set, and the time-frequency characteristics of the sample set are extracted by wavelet packet decomposition. The wavelet packet decomposition includes decomposition operation and reconstruction operation, which are as follows:

[0090] 1) Assume the input signal is a discrete signal The decomposition operation is calculated as follows:

[0091]

[0092] Among them, n is the discrete time series, k is the time shift factor, N is the scale factor, ie is the number of layers, Represents the component to be decomposed (0≤i≤2 N -1), and is the wavelet packet coefficient, h 0 (·) is a low-pass filter, g 0 (·) is a high-pass filter;

[0093] 2) The reconstruction operation is calculated as follows:

[0094]

[0095] in, and is the signal component, h 1 (·) is a high-pass filter, g1 (·) is a low-pass filter.

[0096] Among them, wavelet packet decomposition can decompose the low-frequency and high-frequency parts of the signal at the same time, and is suitable for processing non-stationary vibration signals with high-frequency disturbances and strong background noise. Wavelet packet is proposed to overcome the problem that wavelet decomposition has poor resolution in the high-frequency band and good resolution in the low-frequency band. It is a more sophisticated signal analysis method that improves the time-frequency resolution of the signal.

[0097] In one embodiment, the wavelet packet decomposition uses the db3 wavelet basis.

[0098] The original data and the generated fault data are combined into a sample set, and wavelet packet coefficients are obtained by decomposition operation using wavelet packet decomposition (WPD), and then reconstruction operation is performed to obtain components to extract the time-frequency characteristics of the sample set.

[0099] Specifically: Using the db3 wavelet basis, the elevator operation data (such as elevator vibration signal) is decomposed and reconstructed using a three-layer wavelet packet to obtain a feature vector consisting of eight fault feature values. like Figure 2 As shown, the time-frequency features include the frequency band signals of each node and the corresponding spectrum. The left side is the reconstructed frequency band signals of each node, and the right side is the corresponding Fourier spectrum.

[0100] The sensor sampling frequency used in this embodiment is 512Hz. According to Shannon sampling theorem, the Nyquist sampling frequency is 256Hz. After the signal is decomposed by three layers of wavelet packets, it contains 2^3=8 frequency segments. The frequency interval of each frequency segment is 256 / 8=32Hz. Therefore, the reconstructed signal after wavelet packet decomposition is It should be distributed between [0-32Hz, 32-64Hz, 64-96Hz, 96-128Hz, 128-160Hz, 160-192Hz, 192-224Hz, 224-256Hz], which is consistent with the spectrum diagram. It can be seen from the spectrum diagram that the spectrum of each node is significantly different, and the frequency domain interval overlaps less. Therefore, the three-layer decomposition of wavelet packets can effectively extract the time-frequency characteristics of the elevator vibration signal.

[0101] S5. Divide the extracted time-frequency features into a training set and a test set, and input the training set into the CNN model for training to obtain a trained CNN model.

[0102] In one embodiment, the CNN model includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a first fully connected layer, a second fully connected layer and an output layer, which are connected in sequence.

[0103] In one embodiment, each pooling layer is a maximum pooling layer, and the output end of each convolutional layer and the fully connected layer is also provided with an activation function, the activation function is a ReLU function, and the output layer uses a SoftMax classifier.

[0104] Specifically, the CNN model is Figure 3 As shown in the figure, the size of the input feature map (i.e., time-frequency feature) is 100×100×3; the number of convolution kernels of the first to fourth convolution layers are 32, 64, 128, and 128, respectively, the convolution kernel size of the first and second convolution layers is 5×5, the convolution kernel size of the third and fourth convolution layers is 3×3, and the activation function selects the ReLU function; the pooling method of the pooling layer selects the maximum pooling, that is, the maximum value of the p×p region of the input feature map is calculated, and the region size is 2×2; the number of nodes in the first fully connected layer is 1024, and the number of nodes in the second fully connected layer is 512; the number of nodes in the output layer is the number of categories of the number of input samples, and the classifier selects SoftMax classification.

[0105] Table 2 shows the accuracy of the extracted time-frequency features input into the CNN model after adding different numbers of generated samples to different numbers of original samples. Among them, the learning rate of the CNN model is uniformly set to 0.001, the ratio of the training set to the test set is 4:1, and the step size is 8.

[0106] Table 2

[0107]

[0108]

[0109] As can be seen from Table 2, when the total number of samples is similar, the lower the imbalance ratio, the better the training effect, and when the imbalance ratio is the same, the more samples, the better the training effect. From the training results of sample set 1, it can be seen that although the accuracy of the training set can reach 91% when the amount of data is too small, the generalization of the model is poor, and the accuracy of the test set is only 75%, which verifies the necessity of sample expansion. The accuracy difference between sample sets 3 and 4 is small, indicating that the difference between the generated samples and the original samples is small, which once again proves the effectiveness of the improved TimeGAN. The accuracy of sample set 10 did not achieve the effect of balanced samples. Compared with sample sets 1, 3, 4, and 7 with the same imbalance ratio but fewer samples, the accuracy decreased. This is because the generator model itself also has errors. When the number of generated samples is more than the original data, the error may increase, resulting in a counterproductive effect. Therefore, the sample expansion is often stopped when the number of generated samples is equal to the number of original samples. In summary, adding 512 generated samples to the training set has the best training effect, so the training set accuracy comparison experiment uses sample set 7 as the training set.

[0110] Reference Figure 4, BP neural network has the longest training cycle and the lowest accuracy, LSTM has the shortest training cycle but the accuracy is also not high, the accuracy of the original GAN ​​and CNN combination method can reach 94.41%, which is slightly improved compared with the CNN model, and the WPD+BP neural network has a better effect, with an accuracy of 95.75%, which is nearly 6% higher than the BP neural network, verifying the effectiveness of WPD. Compared with the above four methods, the method in this paper combines the advantages of the generation model (TimeGAN model), WPD and CNN. Under the same data set, the training accuracy reached 99.01%, which significantly improved the model training effect.

[0111] S6. Input the test set into the trained CNN model to obtain the fault diagnosis results and verify the fault diagnosis results.

[0112] Table 3 shows the average test accuracy of different methods and different samples. The test set accuracy can reflect the generalization ability of the model and is also an important indicator for judging the final fault diagnosis effect of the model.

[0113] Table 3

[0114]

[0115]

[0116] As can be seen from Table 3, the accuracy of the Extreme Learning Machine (ELM) algorithm is the lowest, only 79.76%, followed by the BP neural network; after WPD decomposition, the accuracy of the BP neural network is increased to 94.05%, which is better than other algorithms, indicating that shallow machine learning algorithms such as BP, LSTM, and SVM are difficult to fully extract features from vibration signals, and the advantages of WPD in feature extraction; the accuracy of the CNN model is higher than that of other shallow machine learning algorithms, verifying its powerful feature extraction ability; and the accuracy of the classification after the original GAN ​​expansion sample is input into the CNN does not improve, and even decreases, because although the original GAN ​​and other generation models can generate realistic pictures, they cannot consider the temporal relationship between vibration signals and are difficult to process one-dimensional data. The accuracy of this method is higher than that of other methods in both balanced and unbalanced sample sets, and the accuracy of sample set 7 reaches 95.19%, which is significantly improved compared with the sample expansion and WPD before processing, which fully reflects the effectiveness of the small sample fault diagnosis method based on the improved TimeGAN model of this application.

[0117] This application introduces the least squares loss function into the time series generative adversarial network (TimeGAN) to improve the objective function of the model based on the small sample fault diagnosis method of the improved TimeGAN model, generates fault samples with time series characteristics, reduces the imbalance of the original data distribution, and solves the problem of insufficient samples; and uses wavelet packets to decompose the expanded samples, calculate the wavelet packet coefficients of each node and reconstruct the signal to fully extract the time-frequency characteristics of the samples; finally, the powerful spatial feature extraction capability of the CNN convolutional neural network is combined to extract the spatial features of the signal, establish a fault diagnosis model, and achieve efficient and accurate fault diagnosis, thereby realizing the safe maintenance of intelligent equipment.

[0118] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0119] The above-described embodiments only express the more specific and detailed embodiments described in this application, but they cannot be understood as limiting the scope of the patent application. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of this application, which all belong to the protection scope of this application. Therefore, the protection scope of the patent application shall be based on the attached claims.

Claims

1. A small sample fault diagnosis method based on an improved TimeGAN model, characterized in that: The small sample fault diagnosis method based on the improved TimeGAN model includes the following steps: S1. Collecting elevator operation data, where the elevator operation data includes normal samples and fault samples; S2. Establish an improved TimeGAN model, wherein the improved TimeGAN model includes an embedding network, a recovery network, a sequence generator, and a sequence discriminator, wherein: The embedding network transforms static features and temporal features into potential representations using an embedding function, wherein the embedding function is implemented based on a recurrent network; The recovery network uses a recovery function to recover the potential representation to the original static features and temporal features, and the recovery function is implemented based on a feedforward network; The sequence generator samples and extracts static random vectors and temporal random vectors from elevator operation data, and converts the static random vectors and temporal random vectors into synthetic latent codes using a generating function, wherein the generating function is implemented based on a recurrent network; The sequence discriminator receives the synthesized latent code using a discriminant function and returns a classification result, wherein the discriminant function is implemented based on a bidirectional recurrent network with a feedforward output layer; The loss function of the improved TimeGAN model includes the reconstruction loss l R , fight against loss U and supervision loss l S , and construct the following objective function: Among them, θ e ,θ r ,θ g ,θ d represents the parameters of the embedding network, the recovery network, the sequence generator and the sequence discriminator respectively, λ≥0, η≥0, s is the static feature vector, is the restored static eigenvector, x t is the temporal feature vector at time t, is the temporal feature vector at time t after recovery, y S is the real static feature classification result, y t is the real temporal feature classification result at time t, is the generated static feature classification result, is the temporal feature classification result generated at time t, h S is the latent static representation, h t is the potential tense representation at time t, h t-1 is the potential temporal representation at time t-1, z t is the temporal random vector at time t, g χ Generate a network for temporal features, is subject to random variables s,x 1:T The expectation of a probability distribution P, is subject to random variables s,x 1:T Generate probability distribution expectations; S3, inputting the fault sample into the improved TimeGAN model to generate fault data; S4. The elevator operation data and the generated fault data are combined into a sample set, and the time-frequency characteristics of the sample set are extracted by wavelet packet decomposition. The wavelet packet decomposition includes a decomposition operation and a reconstruction operation, which are as follows: 1) Assume the input signal is a discrete signal The decomposition operation is calculated as follows: Among them, n is the discrete time series, k is the time shift factor, N is the scale factor, i is the number of layers, n, k, is the component to be decomposed, 0≤i≤2 N -1, is the wavelet packet coefficient at low frequency, is the wavelet packet coefficient at high frequency, h0(·) is a low-pass filter, and g0(·) is a high-pass filter; 2) The reconstruction operation is calculated as follows: in, is the low-frequency signal component, is the high-frequency signal component, h1(·) is a high-pass filter, and g1(·) is a low-pass filter; S5, dividing the extracted time-frequency features into a training set and a test set, and inputting the training set into the CNN model for training to obtain a trained CNN model; S6. Input the test set into the trained CNN model to obtain the fault diagnosis results and verify the fault diagnosis results.

2. The small sample fault diagnosis method based on the improved TimeGAN model according to claim 1, characterized in that: The CNN model includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a first fully connected layer, a second fully connected layer and an output layer, which are connected in sequence.

3. The small sample fault diagnosis method based on the improved TimeGAN model according to claim 2, characterized in that: Each of the pooling layers is a maximum pooling layer, and the output end of each of the convolutional layers and the fully connected layer is also provided with an activation function, the activation function is a ReLU function, and the output layer uses a SoftMax classifier.

4. The small sample fault diagnosis method based on the improved TimeGAN model according to claim 1, characterized in that: The wavelet packet decomposition adopts the db3 wavelet basis.

5. The small sample fault diagnosis method based on the improved TimeGAN model according to claim 1, characterized in that: The objective function satisfies λ=1, η=10.

Images