Abnormal detection method for mechanical rotating equipment under varying working conditions based on feature alignment residual GAN

Abstract

The present invention discloses a method for detecting abnormalities in mechanical rotating equipment under variable working conditions based on feature alignment residual GAN. The method comprises: constructing a time-frequency graph from rotating machinery data, reconstructing the time-frequency graph using an encoder GE(x) and a decoder GD(x) respectively combined with an autoencoder network of a residual module ResNet, and using a discriminator D to distinguish the authenticity of the reconstructed data and the source data to determine abnormal data. The loss function of the network contains an adversarial loss L. adv , context loss L of the generator reconstruction data error con , encoder loss L enc , and the loss L of Wasserstein distance for constructing ranking parity data classification for eigenvalue alignment w‑gan The present invention has a good effect on detecting abnormalities of rotating machinery under variable working conditions, can detect abnormalities of rotating machinery equipment in advance, ensure the long-term stable operation of industrial systems, and avoid the occurrence of serious accidents, which has important social significance.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault anomaly detection, and in particular to a method for detecting abnormalities in mechanical rotating equipment under varying operating conditions based on feature alignment residual GAN. Background Art

[0002] Mechanical equipment plays a crucial role in industrial production, impacting the operational efficiency of the entire production line. Rotating machinery is widely used in steam turbines, generators, gas turbines, hydraulic turbines, aircraft generators, centrifugal compressors, pumps, various gears, speed increasers, and other mechanical equipment.

[0003] Gears and rolling element bearings, as key mechanical components in most rotating devices, often experience failures due to harsh operating environments and long operating times. Such failures can lead to machine failure or significant economic losses. Detecting rotating machinery failures and anomalies in advance is of great social significance in ensuring the long-term stable operation of industrial systems and avoiding serious accidents. Fault condition monitoring has always been an important component of the full life cycle management of rotating machinery and a research hotspot. In recent years, anomaly detection, as a hot technology in condition monitoring, has become a new development trend, helping to transform traditional corrective maintenance into a condition-based maintenance strategy.

[0004] However, the current methods used in the field of mechanical fault anomaly detection are mostly for stable working conditions or cross-speed and cross-load conditions, and there are few methods for anomaly detection of rotating machinery equipment under variable working conditions. However, in actual production, variable working conditions are very common. It is of great practical significance to study a rotating machinery anomaly detection method suitable for variable working conditions.

[0005] Wang Daolei et al. disclosed in Chinese invention patent CN112184654A a "high-voltage line insulator defect detection method based on generative adversarial networks." By using GAN to train the network with only normal data, it can effectively detect abnormal data. This patent invention can achieve good detection in the field of high-voltage line insulators. However, this method detects defects of the same type of high-voltage line insulators and is not effective in detecting defects in variable-operating mechanical equipment with weak correlation and large distribution differences. Summary of the Invention

[0006] In order to at least solve the technical problems existing in the prior art to a certain extent, the present invention provides a GAN-based method for detecting abnormalities in a rotating machine under variable operating conditions.

[0007] To achieve the purpose of the present invention, the present invention provides a method for detecting abnormalities in mechanical rotating equipment under varying operating conditions based on feature alignment residual GAN, comprising the following steps:

[0008] S1: Collect normal and abnormal vibration signal data of mechanical rotating equipment under different working conditions, perform continuous wavelet transformation on the collected signals, and convert them into time-frequency diagrams;

[0009] S2: Construct anomaly detection model RWGAN network based on the training idea of ​​feature alignment residual GAN;

[0010] S3: Input normal data under variable working conditions into the RWGAN network for iterative update training; S4: After multiple iterative trainings, the error curve tends to be stable, and the RWGAN network training is completed;

[0011] S5: Input the normal data time-frequency diagram and the abnormal data time-frequency diagram under the variable working conditions into the trained RWGAN network to detect the abnormal data of the equipment under the variable working conditions.

[0012] Furthermore, step S1 specifically includes: collecting vibration signals of various mechanical equipment fault conditions through an acceleration sensor; then, dividing the collected original vibration signal into multiple samples of length L using a sliding sampling technique; finally, performing a continuous wavelet transform (CWT) on the corresponding samples to convert them into corresponding time-frequency graphs. The formula used in the continuous wavelet transform process is as follows:

[0013]

[0014]

[0015] L slide =χ ω / 2

[0016] Among them, χ ω It represents the sampling point, L total is the total length of the sample, L slide is the length of the sampling slide. The above three formulas mean that the length of each sample sampling point must be at least greater than one cycle, and there must be half of the overlapping sampling points between every two samples. Figure 1 .

[0017] The corresponding samples are transformed into the corresponding time-frequency graph by continuous wavelet transform (CWT). The formula defined by continuous wavelet transform is:

[0018]

[0019] Where ψ(ω) is the Fourier transform of the original input function ψ(t), X(t)∈L 2 (R), L 2 (R) represents a distribution of energy in a limited space, τ is a factor related to transformation, and a is a factor related to scale.

[0020] Furthermore, step S2 specifically includes: the constructed RWGAN network includes a generator G, a discriminator D, and a residual network ResNet, wherein the generator G consists of a self-decoder (encoder G E (χ) Combine the residual network and decoder G D (χ) combined with a residual network) to learn the data distribution of the input χ and reconstruct the image The initial learning rate L of the network parameters r Set to L r =0.001; the role of the discriminator D is to distinguish the real input data χ and the data generated by the generator G Compare and judge the normality and abnormality of the data. Specifically, after the network training converges, the encoder E calculates the L in all samples. enc When inferring, given an image, we can use the learned network to calculate its L enc If it is less than the discrimination threshold, it is judged as a normal sample, and if it is greater than the discrimination threshold, it is judged as an abnormal sample. ; Encoder G E (χ) and decoder G D (χ) are added with a residual network, which is a key design of the present invention compared with the common self-decoder. The structure of the generator G of the proposed invention method is detailed in the attached Figure 2 .

[0021] Furthermore, step S3 specifically includes: constructing training tasks of different cases using the time-frequency graphs under the variable working conditions in step S1, and the training set only contains time-frequency graphs of normal data, and defining the learning rate of the network as L r ,RWGAN is optimized by RMSprop algorithm.

[0022] Furthermore, step S4 specifically includes: the loss function of the RWGAN network consists of four parts, namely, the adversarial loss L adv , the generator reconstructs the data Error context loss L con , encoder loss L enc , and the loss L of Wasserstein distance for constructing ranking parity data classification and eigenvalue alignment w-gan , the formulas of the four loss functions are expressed as follows:

[0023]

[0024]

[0025]

[0026]

[0027] Where χ represents the input data, Represents the data reconstructed after the generator, and z represents the data after the encoder G E The vector after (χ), For reconstruction The data generated after encoding, Z even is the collection of all the odd-numbered data time-frequency graphs in the data set, Z odd It is the collection of time-frequency graphs of all even-numbered data in the dataset.

[0028] Further analysis reveals that this method utilizes the innovative principle of Wasserstein distance (optimal transmission distance). Common GAN-based methods often use KL divergence (Kullback-Leibler Divergence) and JS divergence (Jensen-Shannon Divergence) as indicators to measure the similarity between two distributions.

[0029] The KL divergence describes the distance between the probability distribution p(χ) obtained from the evaluation training and the target distribution q(χ), which can be expressed as:

[0030]

[0031] The ultimate goal of the machine learning algorithm is to reduce the value of the distribution D. Due to the relative position relationship between p(x) and q(χ) in the logarithmic term in the KL divergence, the KL divergence is actually asymmetric. That is, from the perspective of the physical reference system, it can be intuitively felt that if you want to evaluate the similarity between two objects (distributions), the value of the similarity (such as KL divergence) should not change due to the different reference targets (target distributions) selected. KL divergence is not symmetric, so still from the perspective of the reference system, it is sufficient to simply average the distances calculated in all reference systems (only discussing the target distribution and the predicted distribution reference systems). This is the idea of ​​JS divergence, which is specifically defined as:

[0032]

[0033] Then discuss:

[0034] right

[0035]

[0036]

[0037]

[0038] When θ = 0, the distributions of p and q completely overlap. It can be seen that the KL divergence will produce infinite results when the two distributions have no intersection at all, while the JS divergence will have a sudden step and is not differentiable at 0. Only the Wasserstein distance can provide a smoother result for parameter updates in the gradient descent method.

[0039] The objective loss function is defined as:

[0040] L total =ω adv L adv +ω con L con +ω enc L enc +ω w-gan L w-gan

[0041] Among them, ω adv =1,ω con =50,ω enc =1,ω wgan_loss =1

[0042] Therefore, the addition of Wasserstein distance to the method proposed in the present invention can well align the different eigenvalues ​​of the variable working condition data, and can avoid the non-convergence or non-differentiation in the process of gradient reduction, greatly enhancing the generalization performance detection ability of the network, and better applicable to the abnormal detection of mechanical equipment under variable working conditions in the application scenario of the invention. After applying this module, the effect is shown in the attached Figure 4 : The left side is a schematic diagram of the loss function when the eigenvalues ​​are not aligned when detecting task h1, and the right side is a schematic diagram of the loss function after the eigenvalues ​​of detection task h1 are aligned.

[0043] Furthermore, step S5 specifically includes: constructing different training cases based on the time-frequency graphs under the variable working conditions in step S1. In addition to the normal data in the corresponding training set cases, the test set in the task also contains abnormal time-frequency graphs in the task that did not appear in the training set. The accuracy of the RWGAN detection results is measured by AUC (Area Under Curve).

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] Under variable operating conditions, only normal data is used to train the network, but abnormal data can be effectively detected with good generalization capabilities. Because the present invention incorporates the Wasserstein distance module, it utilizes loss-specific feature learning networks and adversarial training techniques to handle large distribution differences under different working conditions. The optimization algorithm in the network selects RMSProp to optimize the distribution probability of time-frequency images under different working conditions. Secondly, in order to extract richer feature representations and improve the performance of the model, a ResNet module is added. Based on these innovations, the method of the present invention can effectively detect anomalies in rotating machinery equipment "under variable working conditions." BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a schematic diagram of the data preprocessing method described in an embodiment of the present invention.

[0047] Figure 2 This is a structural diagram of the autoencoder described in an embodiment of the present invention.

[0048] Figure 3 This is a structural diagram of the RWGAN network described in an embodiment of the present invention.

[0049] Figure 4 Figure 1 compares the loss functions of the RWGAN network with and without eigenvalue alignment for detection task h1, as described in an embodiment of the present invention. (a) shows the loss function for detection task h1 without eigenvalue alignment, and (b) shows the loss function for detection task h1 after eigenvalue alignment. DETAILED DESCRIPTION

[0050] The present invention will be further described in detail below with reference to the accompanying drawings. The embodiments described with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention. The step numbers in the following embodiments are provided for ease of explanation only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0051] In the description of the present invention, unless otherwise clearly defined, terms such as parameter settings, data sets, and case conclusions should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0052] The present invention provides a method for detecting abnormalities in mechanical rotating equipment under varying operating conditions based on feature alignment residual GAN, comprising the following steps:

[0053] Step 1: Collect normal and abnormal vibration signal data of mechanical rotating equipment under different working conditions, perform continuous wavelet transformation on the collected signals, and convert them into time-frequency diagrams.

[0054] Step 1 includes the following sub-steps:

[0055] Step 1.1: Use acceleration sensors to collect vibration signals of mechanical equipment under various working conditions;

[0056] Step 1.2: Divide the original vibration signal collected in step 1.1 into multiple samples of length L using the sliding sampling technique;

[0057] Step 1.3: Perform continuous wavelet transform (CWT) on the corresponding samples and convert them into the corresponding time-frequency graph.

[0058] Among them, see Figure 1 In some embodiments of the present invention, the following formula means that the length of each sample sampling point is at least greater than one period, and there is half of the overlapping sampling points between every two samples:

[0059]

[0060] L slide =χ ω / 2

[0061] Where, χ ω It represents the sampling point, L total is the total length of the sample, L slide is the length of the sampling slide.

[0062] The formula defined by continuous wavelet transform is:

[0063]

[0064] WT(τ,a) represents the final transformation result; τ is a transformation-related factor, and a is a scale-related factor.

[0065] Step 2: Construct the anomaly detection method RWGAN network based on the feature alignment residual GAN ​​training idea.

[0066] In this step, the RWGAN network includes a generator G and a discriminator D. The generator G is used to learn the data distribution of the input image χ and reconstruct the image χ. The generator G includes an encoder G E (χ) Combine the residual network and decoder G D (χ) combined with residual network. The generator G first reads an input image χ and passes it forward to its encoder GE(χ) and the network combined with the residual network, and then uses convolutional layer, batch-normal and ReLU() activation respectively. The encoder G E (χ) compresses the input image χ into a vector z, and then the generator G continues to pass through a decoder G D(χ) and the network combined with the residual network reconstructs the vector z into the image χ.

[0067] The encoder GE(x) of the generator combined with the residual network structure is shown below, including a two-dimensional convolutional layer, a residual network layer and a normalization layer.

[0068] Table 3

[0069]

[0070]

[0071] Among them, Conv 1, Conv 2, Conv 3, and Conv 4 are two-dimensional convolutional layers, Residual 1, Residual 2, Residual 3, and Residual 4 are residual network layers, and BN 1 and BN 2 are normalization layers.

[0072] This method is based on the principle of adding ResNet module to the autoencoder. Figure 3 . First, the low-resolution time-frequency image is input into the autoencoder network, and the convolution layer is used to extract the shallow features of the input image; secondly, the convolution kernels of different sizes in the autoencoder module are used to obtain the local features of the image under different receptive fields; then, the outputs of the residual blocks are combined and global features are fused using a 1×1 convolution layer; finally, the local features obtained by the residual blocks are combined with the global features using a residual connection, making full use of the low-resolution image features of the input network. The idea of ​​using residual connections can reduce the problem of input image features being lost during network transmission. A deep residual module is constructed in the network to alleviate the gradient vanishing and training stability problems of the generative adversarial network, while increasing the diversity of the generated images. The method proposed in the present invention can effectively avoid the vanishing gradient problem and degradation problem of the time-frequency graph during the encoding-decoding process. Assuming that the output result to be obtained is A, the function of the generator to be learned is represented by the function G, and the output of the residual module is represented by B, then:

[0073] A=G(G E ,G D )+B

[0074] From the above formula, the learning goal of the learning generator is G(G E ,G D )=AB, which is a linear identity mapping, and the eigenvalue learning of the autoencoder becomes easy.

[0075] The decoder G of the generator D (χ) The combined residual network structure is shown below, including a two-dimensional deconvolution layer, a residual network layer, and a normalization layer.

[0076] Table 4

[0077]

[0078] Among them, ConvTransposed 1, ConvTransposed 1, ConvTransposed 1, ConvTransposed 1 are two-dimensional deconvolution layers, Residual 5, Residual 6, Residual 7, Residual 8 are residual network layers, BN 3, BN 4 are normalization layers.

[0079] In the RWGAN network of the present invention, the encoder E and the discriminator D are connected to the encoder G. E (χ) is the same structure. Encoder E is used to calculate the encoder loss L in all samples enc Value, encoder E and discriminator D both include convolutional layers and BN layers.

[0080] The discriminator D is used to discriminate the real input data - the input image χ and the reconstructed image χ produced by the generator G, and judge whether the data is normal or abnormal. Specifically, in some embodiments of the present invention, after the network training converges, the encoder E calculates the encoder loss L in all samples enc When inferring, given an image, use the learned network to calculate its encoder loss L enc If it is less than the discrimination threshold, it is judged as a normal sample; if it is greater than the discrimination threshold, it is judged as an abnormal sample.

[0081] In some embodiments of the present invention, the parameters in the RWGAN network are set as follows:

[0082] Table 5

[0083]

[0084] Step 3: Input the normal data under variable working conditions into the RWGAN network for iterative update training.

[0085] In some embodiments of the present invention, in this step, the time-frequency graph under the variable working condition in step S1 is used to construct training tasks of different cases, and the training set in the task only has the time-frequency graph of normal data. The learning rate of the network is defined as L r , the RWGAN network is optimized using the RMSprop algorithm.

[0086] The loss function of the RWGAN network consists of four parts: adversarial loss L adv , contextual loss L of the error of the generator reconstructing the image x con , encoder loss Lenc And the loss L of Wasserstein distance for constructing parity data classification and eigenvalue alignment w-gan ,The target loss function is defined as:

[0087] L total =ω adv L adv +ω con L con +ω enc L enc +ω w-gan L w-gan

[0088] ω adv 、ω con 、ω enc 、ω w-gan is the weight parameter corresponding to each loss. In some embodiments of the present invention, ω adv =1,ω con =50,ω enc =1,ω w-gan =1.

[0089] Among them, L w-gan This is a critical module of the present invention because the correlation between the corresponding data in the generalization performance of rotating machinery detected by the present invention is very weak. The Wasserstein distance metric can be used to compare completely different probability distributions. Training the generative model can minimize the divergence between the model and the true distribution of the data, and can measure the distance between any two distributions, that is, the degree of similarity between any two distributions.

[0090] The correlation between the corresponding data in the generalization performance of rotating machinery detected is very weak. By adding the Wasserstein distance, this application can effectively align the different eigenvalues ​​of the variable working condition data, and can avoid non-convergence or non-differentiability during the gradient reduction process. This greatly enhances the network's ability to detect generalization performance under variable working conditions, and is better suited for anomaly detection under variable working conditions of mechanical equipment in the application scenarios of the invention. The feature alignment loss function is defined as:

[0091]

[0092] Z even is the collection of all the odd-numbered data time-frequency graphs in the data set, Z odd It is the set of time-frequency graphs of all even-numbered data in the data set, which conform to a joint distribution γ,γ~Π(P1,P2) with probabilities P1 and P2 respectively.

[0093] In addition, the expressions of the other three loss functions are

[0094]

[0095]

[0096]

[0097] Where χ represents the input image data, Represents the reconstructed image data after the generator, and z represents the image data after the encoder G E The vector after (χ), For reconstruction The data generated after encoding, Z even is the collection of all the odd-numbered data time-frequency graphs in the data set, Z odd is the set of all the even-numbered data time-frequency graphs in the data set, f(χ) and They are the original image output and the reconstructed image output respectively.

[0098] Step 4: After multiple iterations of training, the error curve tends to be stable, and the RWGAN network training is completed.

[0099] Step 5: Input the normal data frequency graph and abnormal data frequency graph under variable working conditions (i.e., the test set) into the trained RWGAN network to detect abnormal equipment data under variable working conditions.

[0100] Different training cases are constructed using the time-frequency graphs under the varying operating conditions in step S1. In addition to the normal time-frequency graphs corresponding to the training set cases, the test set also contains abnormal time-frequency graphs from the task that were not present in the training set. The accuracy of the RWGAN detection results is measured using the Area Under Curve (AUC).

[0101] In some embodiments of the present invention, the effectiveness of the method of the present invention is verified by using specific cases. The case data set comes from a failure experiment of a five-speed three-axis automobile transmission, which is carried out on a mechanical transmission test platform. In the measurement test, the motor is connected to the drive shaft through a belt, and the vibration signal is captured by an accelerometer installed on the output shaft with a sampling frequency of 12kHz. These bearing rotations include four different operating speeds (from 1500rpm to 500rpm). At each different speed, the load on the output shaft is 50 nanotorques, as shown in Table 1, which will simulate eight different health conditions under variable working conditions.

[0102] Table 1

[0103]

[0104]

[0105] The data preprocessing process of the present invention is shown in Figure 2 :

[0106] First, the sampling length selection problem, such as 500r / min, each circle takes 0.12s, the sampling frequency is 12KHz. 12k samples per second (1k=10 3 ) sample points, so approximately 1440 sample points are sampled per revolution. To fully learn the characteristics of the bearing under variable operating conditions, four to five revolutions of sampling points are used as one sample length, which translates to over 6000 sample points. Secondly, to expand data capacity, 50% of the sampling points overlap between the front and back samples.

[0107] The present invention establishes different cases for the data set to verify the anomaly detection effect of the proposed method. Taking the case of verifying the generalization performance as an example, in this case, the training set only contains time-frequency graphs of normal data, and the test set contains both time-frequency graphs of normal data and time-frequency graphs of abnormal data. In addition, for the selected tasks, the training set of task h1 only contains normal data of speed increase, and verifies the abnormal data of speed increase; the training set of task h2 only contains normal data of speed increase, and verifies the abnormal data of speed increase; the training set of task h3 contains normal data of speed increase and speed increase, and verifies the abnormal data of speed increase; the training set of task h4 contains normal data of speed increase and speed increase, and verifies the abnormal data of speed increase. Therefore, the task settings of h1-h4 are to verify the generalization performance of the RWGAN of the present invention. The case tasks are shown in the figure below:

[0108] Table 2

[0109]

[0110]

[0111] The results of the anomaly detection loss function for task h1 in the case corresponding to the RWGAN network in this invention are shown in the attached Figure 4 , (a) is the loss function result of feature alignment without adding Wasserstein distance (optimal transmission distance), (b) is the loss function result of feature alignment with adding Wasserstein distance w-gan The loss function results of the two figures are easy to see. The loss function of the Wasserstein distance of the eigenvalue alignment operation performed by the present invention is the loss L w-gan It can smooth the loss function and accelerate the convergence of the model with obvious effect. enc It takes 50 iterations to converge, and other losses converge within 10 iterations. The performance of the entire network is very good.

[0112] In the anomaly detection case corresponding to the RWGAN network in the embodiment of the present invention, 3,000 time-frequency graphs were selected as training samples and 600 time-frequency graphs were selected as test samples. The three methods used for comparison included CAE, DSVDD, and f-AnoGAN. Specifically, CAE's design philosophy is that the encoder only performs representation learning, while the decoder is responsible for solving the pretext task, fully tapping the encoder's potential. DSVDD uses a jointly trained deep neural network to map normal sample data to a minimum volume hypersphere. Samples mapped within the hypersphere are classified as normal, while samples mapped outside the hypersphere are considered abnormal. The f-AnoGAN concept is divided into two steps: the first step is to train a generative adversarial network, and the second step is to train an encoder using the weights of the first generative adversarial network. The test of each detection task in the case was repeated five times to calculate the average accuracy to reduce randomness. The AUC (Area Under Curve) accuracy results are shown in Table 6.

[0113] Table 6

[0114]

[0115]

[0116] Without loss of generality, in this case, in the verification of the generalization performance under variable working conditions, the proposed method RWGAN achieves anomaly detection accuracy that is at least 15% higher than that of commonly used anomaly detection methods.

[0117] In summary, the present invention targets variable operating conditions, takes rolling bearings as the research object, and the proposed method effectively improves the abnormality detection capability of industrial equipment.

[0118] Generally speaking, the correlation between the corresponding data in the generalization performance of rotating machinery detected is very weak. The method proposed in this application adds Wasserstein distance, which can well align the different eigenvalues ​​of variable working condition data, and can avoid non-convergence or non-differentiability in the process of gradient reduction, greatly enhancing the network's ability to detect generalization performance under variable working condition scenarios, and better applicable to the abnormal detection of mechanical equipment under variable working conditions in the application scenarios of the invention.

[0119] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

[0120] The description of the above specific implementation methods is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting abnormalities in mechanical rotating equipment under varying working conditions based on feature alignment residual GAN, characterized in that: The following steps are involved: S1: Collect normal and abnormal vibration signal data of mechanical rotating equipment under different working conditions, and convert the collected signals into time-frequency diagrams; S2: Based on the feature alignment residual GAN ​​training idea, an anomaly detection method RWGAN network is constructed. The RWGAN network includes a generator G and a discriminator D. The generator G is equipped with a residual module ResNet. The generator G is used to learn the data distribution of the input image χ and reconstruct the image χ. The discriminator D is used to compare the input image χ with the reconstructed image χ produced by the generator G to determine whether the data is normal or abnormal. S3: Input the normal data under the variable working conditions into the RWGAN network for iterative update training; S4: After multiple iterations of training, the error curve stabilizes and the RWGAN network training is completed. During iterative training, the loss function of the RWGAN network includes the Wasserstein distance loss L for constructing parity bit data classification and eigenvalue alignment. w-gan , loss L w-ga The expression is Where Z even is the collection of all the odd-numbered data time-frequency graphs in the data set, Z odd It is the set of time-frequency graphs of all even-numbered data in the data set, which conform to a joint distribution γ,γ~Π(P1,P2) with probabilities P1 and P2 respectively; S5: Input the normal data time-frequency diagram and the abnormal data time-frequency diagram under the variable working conditions into the trained RWGAN network to detect the abnormal data of the equipment under the variable working conditions.

2. The method for detecting abnormalities in mechanical rotating equipment under variable working conditions based on feature alignment residual GAN ​​according to claim 1 is characterized in that: Step S1 specifically includes: The vibration signals of mechanical equipment under various working conditions are collected through acceleration sensors; The collected original vibration signal is divided into multiple samples of length L using sliding sampling technology; the corresponding samples are subjected to continuous wavelet transform and converted into corresponding time-frequency diagrams.

3. The method for detecting abnormalities in mechanical rotating equipment under variable working conditions based on feature alignment residual GAN ​​according to claim 2 is characterized in that: The formula for continuous wavelet transform is: WT(τ,a) represents the final transformation result, χ(t)∈L 2 (R), L 2 (R) represents a distribution in a finite space of energy, τ is a transformation-related factor, and a is a scale-related factor.

4. The method for detecting abnormalities in mechanical rotating equipment under variable working conditions based on feature alignment residual GAN ​​according to claim 1 is characterized in that: In step S2, the generator G includes a self-decoder, which includes an encoder G E (χ) and decoder G D (χ), and the encoder G E (χ) and decoder G D (χ) are combined with residual networks respectively. The generator G first reads an input image χ and passes it forward to its encoder G E (χ) and the network combined with the residual network, and then using convolutional layers, batch-normal and ReLU() activation, the encoder G E (χ) compresses the input image χ into a vector z, and then the generator G continues to pass through a decoder G D (χ) and the network combined with the residual network reconstructs the vector z into the image χ.

5. The method for detecting abnormalities in mechanical rotating equipment under variable working conditions based on feature alignment residual GAN ​​according to claim 1 is characterized in that: After the network training converges, calculate the encoder loss L in all samples enc Value, select the largest one as the discrimination threshold. When judging the data in step S2, given a picture, use the learned network to calculate its encoder loss L enc If it is less than the discrimination threshold, it is judged as a normal sample; if it is greater than the discrimination threshold, it is judged as an abnormal sample.

6. The method for detecting abnormalities in mechanical rotating equipment under variable working conditions based on feature alignment residual GAN ​​according to claim 1 is characterized in that: Step S3 specifically includes: constructing training tasks of different cases using the time-frequency graphs under the variable working conditions in step S1, and the training set in the task only has the time-frequency graphs of normal data, and defining the learning rate of the network as L r , RWGAN is optimized using the RMSprop algorithm.

7. The method for detecting abnormalities in mechanical rotating equipment under varying working conditions based on feature alignment residual GAN ​​according to claim 1 is characterized in that: In step S4, the loss of the RWGAN network also includes the adversarial loss L adv , context loss L of the error of the generator reconstructing the data x con , encoder loss L enc , then the objective loss function is L total =ω adv L adv +oh con L con +oh enc L enc +oh w-gan L w-gan ω adv 、ω con 、ω enc 、ω w-gan are the weight parameters corresponding to each loss.

8. The method for detecting abnormalities in mechanical rotating equipment under varying working conditions based on feature alignment residual GAN ​​according to claim 7 is characterized in that: Context loss L con , encoder loss L enc and adversarial loss L adv The expression is Where χ represents the input image data, Represents the reconstructed image data after the generator, and z represents the image data after the encoder G E The vector after (χ), For reconstruction The data generated after encoding, Z even is the collection of all the odd-numbered data time-frequency graphs in the data set, Z odd It is the collection of time-frequency graphs of all even-numbered data in the dataset.

9. The method for detecting abnormalities in mechanical rotating equipment under varying working conditions based on feature alignment residual GAN ​​according to claim 1 is characterized in that: In step S5, the test set includes not only the normal data time-frequency graphs corresponding to the tasks in the training set cases, but also the abnormal data time-frequency graphs that have not appeared in the training set. The test set is input into the trained RWGAN network to detect the abnormal data time-frequency graphs.

10. The method for detecting abnormalities in mechanical rotating equipment under varying working conditions based on feature alignment residual GAN ​​according to claim 9 is characterized in that: The accuracy of the detection results is measured by AUC.

Images