Motor composite fault diagnosis method based on empirical knowledge and data fusion driving

Abstract

The invention discloses an empirical knowledge and data fusion driven motor load fault diagnosis method, which comprises the steps of constructing an empirical knowledge feature space of a current / vibration signal, and obtaining a plurality of simulation signal samples; inputting the simulation signal sample into an improved generation network DDGAN for processing to obtain an expanded simulation signal sample, and combining an actual fault sample with the expanded simulation signal sample to obtain a fault signal sample; constructing an ICNLCNDTL diagnosis model, inputting the fault sample signal into the ICNLCNDTL diagnosis model for diagnosis processing, and obtaining a current signal diagnosis result and a vibration signal diagnosis result; and fusing the current signal diagnosis result and the vibration signal diagnosis result through a multi-sensing signal collaborative fusion motor composite fault diagnosis model to complete final fusion identification and obtain a fault diagnosis result. According to the invention, the composite fault cross-domain diagnosis performance is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a motor load fault diagnosis method driven by empirical knowledge and data fusion. Background Art

[0002] The basic premise of the new power system is to ensure energy and power security. To ensure the safe operation of the new power system, state perception technology is the cornerstone, and comprehensive, accurate and reliable fault diagnosis technology is the means. As an important energy conversion equipment in the new power system, the safe and stable operation of the asynchronous motor is an important part of ensuring the safe, efficient and continuous operation of the production process. However, affected by various destructive factors from itself and the outside world, the various parts of the motor have defects, damage or even serious failures, resulting in motor shutdown and reduced production efficiency. Timely detection of the fault state of the asynchronous motor and monitoring and maintenance in the early stage of the fault are conducive to extending the service life, improving the reliability of the entire production system, and preventing major safety accidents. Therefore, it is of great significance to study the fault diagnosis and identification method of asynchronous motors.

[0003] In order to ensure the stable operation of asynchronous motors, many mature and reliable relay protection devices were deployed on the motors, such as stator overcurrent protection, motor negative sequence current protection and motor differential protection. These protection measures are based on different protection principles, reasonably organize relays and set their parameter values. Once certain parameter values ​​exceed the threshold, the protection starts and an alarm is sounded. In severe cases, the tripping circuit is activated and the motor stops to prevent the accident from spreading further. However, this method cannot achieve the knowledge and control of the entire operation process of the motor. The protection scope of the relay protection device is limited to accidents and serious faults, and it does not have an early warning function. It is the last safety line of defense for the stable operation of the motor.

[0004] Later, in order to avoid the sudden stop of asynchronous motors, regular maintenance was adopted to conduct comprehensive maintenance and diagnosis of the motors. Regular maintenance can timely detect motor defects and avoid further expansion of faults to a certain extent, but this method is invasive and difficult to avoid hidden dangers caused by improper human operation. It is also blind and time-consuming and labor-intensive, affecting normal production and life. Thanks to the rapid development of the power Internet of Things, experts and scholars have realized the knowable and controllable operation of the entire asynchronous motor process through real-time analysis of sensor signals, and the diagnosis method based on signal processing has become an important research direction. The traditional motor fault diagnosis method is to compare the size relationship between the typical fault characteristics and thresholds in the monitoring signal, but due to environmental noise and interference from electronic devices, early fault characteristics are often masked, resulting in delayed fault detection, which in turn affects the timely maintenance of the motor. The rapid development of artificial intelligence algorithms provides huge opportunities for intelligent diagnosis of motors, getting rid of the dilemma of traditional fault diagnosis relying on expert experience and knowledge, but it also has limitations: from relying on expert experience to relying on training data, requiring independent and identical distribution of source domain and target domain data, etc.

[0005] Therefore, the present invention proposes a motor load fault diagnosis method driven by empirical knowledge and data fusion. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a motor load fault diagnosis method driven by empirical knowledge and data fusion. Starting from the current signal, the motor fault mechanism is analyzed, and the fault characteristics contained in the current signal are effectively extracted to solve the problem of diagnostic performance degradation caused by domain distribution differences in variable operating conditions. By establishing a collaborative diagnosis model of vibration signals and current signals based on multi-source information fusion theory, the cross-domain diagnosis performance of complex faults is further improved.

[0007] To achieve the above object, the present invention provides a motor load fault diagnosis method driven by empirical knowledge and data fusion, comprising:

[0008] Construct the empirical knowledge feature space of current / vibration signal and obtain several simulation signal samples;

[0009] Inputting the simulation signal sample into the improved generation network DDGAN for processing to obtain an expanded simulation signal sample, and combining the actual fault sample with the expanded simulation signal sample to obtain a fault signal sample;

[0010] Construct an ICNLCNDTL diagnostic model, input the fault sample signal into the ICNLCNDTL diagnostic model for diagnostic processing, and obtain current signal diagnostic results and vibration signal diagnostic results; wherein the ICNLCNDTL diagnostic model performs feature extraction through a dual-branch feature extraction network of IConvNeXt-LCNN;

[0011] The motor composite fault diagnosis model based on the coordinated fusion of multiple sensor signals is used to fuse the current signal diagnosis result and the vibration signal diagnosis result to obtain the fault diagnosis result.

[0012] Preferably, obtaining the plurality of simulation signal samples comprises:

[0013] A multi-fault finite element simulation model of an asynchronous motor is established, the stator current characteristics during faults are analyzed, and vibration and current signals under various operating conditions and fault types are generated. The fault mechanism, empirical knowledge, and time-frequency signal analysis method are combined to construct the empirical knowledge feature space of the current / vibration signal, and the time-frequency detail features of the SDP image, the time-frequency detail features of the ISDP image, and the TFSF mechanism knowledge features are obtained respectively.

[0014] Preferably, the improved generation network DDGAN uses the DDPG algorithm to optimize the generation strategy of the generator and adjusts the parameters of the generator according to the current state and reward signal.

[0015] Preferably, the ICNLCNDTL diagnostic model comprises:

[0016] Signal processing module: used to combine the analysis of fault response vibration / current signal experience knowledge, construct the mechanism feature matrix TFSF, and encode the current signal into two-dimensional space using the ISDP method;

[0017] Dual-branch feature extraction module: used to build a dual-branch feature extraction network based on IConvNeXt-LCNN to extract and fuse the local detail features and mechanism features of the current signal and vibration signal;

[0018] Domain adaptation strategy module: used to adapt the discriminative joint distribution to DJP-MMD loss and conditional domain discriminator to reduce the probability distribution difference between the source domain and target domain deep features;

[0019] Motor state classifier: used to learn the characteristic patterns corresponding to motors in different health states, identify the association between different features and the motor health state, and judge the motor health state based on its features through the trained motor state classifier.

[0020] Preferably, the dual-branch feature extraction network based on IConvNeXt-LCNN includes:

[0021] Multi-scale feature extraction sub-network: used to obtain local feature information of faults at different scales in ISDP images and fuse local fault information at each scale;

[0022] Lightweight convolutional neural sub-network: used to obtain deep features of TFSF.

[0023] Preferably, the backbone network of the multi-scale feature extraction subnetwork is a ConvNeXt network, each Stage module of the ConvNeXt network is formed by stacking a downsampling layer and a ConvNeXt Block, and the ordinary convolution layer in the downsampling layer in the Stage is replaced by an ODConv, wherein the convolution kernel size used by the ODConv contained in the Stage module is different, and the combination method of the downsampling layer in the Stage module is to first perform an ODConv operation, then perform layer normalization processing, and finally perform multi-scale fusion of the features output by the Stage module through an asymmetric feature fusion module;

[0024] The lightweight convolutional neural sub-network includes two convolutional layers and one fully connected layer.

[0025] Preferably, the ICNLCNDTL diagnostic model combines the discriminant joint distribution adaptation DJP-MMD loss and the conditional domain discriminator to reduce the probability distribution difference;

[0026] Among them, the discriminant joint distribution adaptation DJP-MMD loss is obtained by introducing the joint probability difference JPD based on the MMD definition; the conditional domain discriminator is used to apply the multilinear conditional adjustment method to improve the classification accuracy of the state classifier by optimizing the cross entropy between the deep features and the classifier predictions.

[0027] Preferably, the loss function of the ICNLCNDTL diagnostic model includes a source domain classification loss function, a DJP-MMD loss function between the source domain and the target domain, and a conditional domain adversarial loss function;

[0028] The source domain classification loss function adopts the cross entropy loss function to measure the difference between the real label and the predicted label of the source domain, and the DJP-MMD loss function between the source domain and the target domain uses DJP-MMD to measure the distribution difference between the source domain and the target domain.

[0029] Preferably, the motor composite fault diagnosis model with collaborative fusion of multiple sensor signals utilizes Lance distance to calculate credibility and an improved DS evidence fusion method based on Deng entropy to calculate uncertainty to fuse the current signal diagnosis results and the vibration signal diagnosis results.

[0030] Compared with the prior art, the present invention has the following advantages and technical effects:

[0031] The present invention effectively extracts the fault features contained in the current signal, constructs a migration strategy based on feature distribution alignment and discriminant distribution alignment, and solves the problem of diagnostic performance degradation caused by domain distribution differences in variable working conditions. It also establishes a vibration signal and current signal collaborative diagnosis model based on multi-source information fusion theory, which further improves the cross-domain diagnosis performance of complex faults.

[0032] This paper introduces the DDPG algorithm into GAN and designs an improved generative network DDGAN. Compared with the random parameter update method in the traditional GAN, in the DDGAN network model, DDPG can adjust the parameters of the generator in a targeted manner according to the current state and reward signal, so that it can be optimized faster in the direction of generating more realistic samples. The reward mechanism in DDPG can provide a clearer optimization goal for the generative adversarial network. By designing a suitable reward function, the generator can be guided to generate samples that better meet the requirements of specific tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0034] Figure 1 This is a schematic diagram of the ICNLCNDTL diagnostic model according to an embodiment of the present invention;

[0035] Figure 2 Ansys / Maxwell two-dimensional transient analysis flow chart of an embodiment of the present invention;

[0036] Figure 3 This is a DDGAN network structure diagram of an embodiment of the present invention;

[0037] Figure 4 Schematic diagram of the structure of a dual-branch feature extraction network based on IConvNeXt-LCNN in an embodiment of the present invention;

[0038] Figure 5 A schematic diagram of the structure of an AFF according to an embodiment of the present invention;

[0039] Figure 6 This is a diagram showing the difference between DJP-MMD and MMD in an embodiment of the present invention;

[0040] Figure 7 A framework diagram of a CDD according to an embodiment of the present invention;

[0041] Figure 8 Schematic diagram of motor compound fault diagnosis based on multi-source information fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0043] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0044] The present invention proposes a motor load fault diagnosis method driven by empirical knowledge and data fusion, comprising:

[0045] Construct the empirical knowledge feature space of current / vibration signal and obtain several simulation signal samples;

[0046] Inputting the simulation signal sample into the improved generation network DDGAN for processing to obtain an expanded simulation signal sample, and combining the actual fault sample with the expanded simulation signal sample to obtain a fault signal sample;

[0047] An ICNLCNDTL diagnostic model is constructed, and the fault sample signal is input into the ICNLCNDTL diagnostic model for diagnostic processing to obtain current signal diagnostic results and vibration signal diagnostic results; wherein, the ICNLCNDTL diagnostic model performs feature extraction through a dual-branch feature extraction network of IConvNeXt-LCNN. Figure 1 Schematic diagram of the ICNLCNDTL diagnostic model.

[0048] The motor composite fault diagnosis model based on the coordinated fusion of multiple sensor signals is used to fuse the current signal diagnosis result and the vibration signal diagnosis result to complete the final fusion judgment and obtain the fault diagnosis result.

[0049] Further, obtaining the plurality of simulation signal samples includes:

[0050] A multi-fault finite element simulation model of an asynchronous motor is established, the stator current characteristics during faults are analyzed, and vibration and current signals under various operating conditions and fault types are generated. The fault mechanism, empirical knowledge, and time-frequency signal analysis method are combined to construct the empirical knowledge feature space of the current / vibration signal, and the time-frequency detail features of the SDP image, the time-frequency detail features of the ISDP image, and the TFSF mechanism knowledge features are obtained respectively.

[0051] Specifically, by establishing a multi-fault finite element simulation model of an asynchronous motor, analyzing the stator current characteristics during faults, and verifying the effectiveness of the simulation model, vibration and current signals under various operating conditions and fault types can be generated based on the simulation model. By combining the fault mechanism, empirical knowledge, and time-frequency signal analysis method, the empirical knowledge feature space of the current / vibration signal is constructed, and the time-frequency detail features of the SDP and ISDP images and the TFSF mechanism knowledge features are obtained.

[0052] In this embodiment, a finite element model of the motor in normal state is built on the Ansys / Maxwell platform. In order to improve the high fidelity of the model, the correlation coefficient between the time domain signals corresponding to the simulated motor and the physical model is calculated. When the correlation coefficient is greater than the threshold (set to 0.6 in this embodiment), the model is considered to have good accuracy. Otherwise, the model parameters are corrected and the model is updated until it meets the requirements. The analysis process is as follows: Figure 2 shown.

[0053] On the basis of simulating the motor in normal state, the motor with broken rotor bar fault and stator turn-to-turn short circuit fault is simulated by modifying the model parameters; the relative position of the moving parts is used to set the eccentric fault motor; the bearing fault simulation motor and the compound fault simulation motor are simulated by using the joint simulation method of Maxwell and Simplorer.

[0054] Based on the simulation model, vibration and current signals under various working conditions and fault types can be generated. By combining the fault mechanism, empirical knowledge, and time-frequency signal analysis methods, the empirical knowledge feature space of current / vibration signals can be constructed.

[0055] This embodiment comprehensively compares a variety of signal time-frequency analysis methods, adopts an improved SDP image coding method to analyze and process the current / vibration signal, and constructs an ISDP image for the three-phase current signal to fully mine the fault information contained in the amplitude and phase of the vibration / current signal; based on the statistical feature selection strategy, the screened fault-sensitive features are rearranged in dimension and reconstructed into a mechanism feature matrix TFSF, so as to mine the fault mechanism characteristics from the overall level.

[0056] Symmetrical point mode ISDP for three-phase current fusion:

[0057] Considering that the convolutional layer performs well when inputting images, it is necessary to use the signal-image conversion method to map the three-phase current signal into a two-dimensional image. Markov Transition Field (MTF), GAF and other time series signal-to-image conversion methods can convert one-dimensional time series signals into feature maps, but there is a common limitation: only the amplitude or phase of a single-phase current is used, and the characteristics of the three-phase amplitude and phase difference of the current signal are not fully explored. In order to make full use of the amplitude and phase information of the three-phase current signal, an improved symmetric dot pattern (Symmetric Dot Pattern with three phase current, ISDP) for three-phase current fusion is proposed. ISDP is obtained by reasonably sorting the phase currents based on SDP. The entire polar coordinate system is divided into six sectors, each sector in the upper half plane maps the same phase current, and each sector in the lower half plane maps two different phase currents. The sectors corresponding to the 0° axis of the polar coordinate rotated counterclockwise by 60° are mapped to the A phase current, B phase current, C phase current, AC phase current, AB phase current, and BC phase current. From a mathematical perspective, the equation (1) By taking reasonable values ​​for the fixed value θ contained in φ(i), the three-phase current signal can be mapped to the polar coordinate plane. The mathematical expression is as follows:

[0058]

[0059] In the formula, Respectively and the value of the fixed value θ in φ(i).

[0060] The subtle differences between different signals mainly depend on the parameters ζ and τ. Appropriate ζ and τ can effectively enhance the image resolution and amplify the differences between signals. Therefore, this embodiment uses the typical image evaluation index two-dimensional correlation coefficient method to select the optimal parameters ζ and τ. For an image of size m×n, r corr2 (A,B) can be expressed as:

[0061]

[0062] In the formula, are the average grayscale values ​​of images A and B respectively.

[0063] r corr2 The value range of (A,B) is between 0 and 1, r corr2 The larger the value of (A, B), the more similar image A and image B are. Otherwise, it means that the difference between image A and image B is large, so it is easy to identify image A and image B. For multiple images, their total r sum can be written as:

[0064]

[0065] In the formula, A i and B j is the corresponding image sample, and k is the total number of image samples. Therefore, the total r of multiple images can be used sum as a criterion for selecting parameters ζ and τ.

[0066] Time-frequency statistical feature matrix TFSF:

[0067] The time-frequency statistical features corresponding to the three-phase current samples are extracted, including peak-to-peak value, double frequency, skewness, etc. The specific types are shown in Table 1.

[0068] Table 1

[0069]

[0070] For the time-frequency statistical features extracted from each current sample, their dimensions are rearranged to construct a 4*5 time-frequency statistical feature matrix TFSF (mechanism feature matrix) as the input of the lightweight convolutional neural sub-network.

[0071] Furthermore, the improved generation network DDGAN uses the DDPG algorithm to optimize the generation strategy of the generator and adjusts the parameters of the generator according to the current state and reward signal, wherein the DDPG algorithm includes an Actor network and a Critic network, the Actor network is used to take the current state information of the generator as input to adjust the parameters of the generator; the Critic network is used to evaluate the quality of the action output by the Actor network in the current state and feed back to the Actor network.

[0072] Specifically, this embodiment introduces the DDPG algorithm into GAN, so that the generator can be adjusted more efficiently in the continuous parameter space, improve the quality and diversity of the generated samples, and design an improved generation network DDGAN. Through the above simulation modeling method, a large number of simulated signal samples can be generated. Combined with DDGAN, a limited number of actual fault samples and simulated signal samples are used for joint generation, which can generate a large number of fault samples that are close to the real signal characteristics, effectively expand the fault signal samples, and ultimately improve the performance of the diagnostic model. The DDGAN network structure diagram is shown in the figure. Figure 3 shown.

[0073] Generative Adversarial Network (GAN) consists of a generator and a discriminator. The goal of the generator is to generate samples that are as realistic as possible to deceive the discriminator, while the task of the discriminator is to distinguish between real samples and fake samples generated by the generator. DDPG can learn a smarter strategy to adjust the parameters of the generator. The DDGAN network model regards the training process of the generator in GAN as a continuous decision-making process and uses the DDPG algorithm to optimize the generator's generation strategy. Compared with the random parameter update method in traditional GAN, in the DDGAN network model, DDPG can adjust the parameters of the generator in a targeted manner according to the current state and reward signal, so that it can be optimized faster in the direction of generating more realistic samples. The reward mechanism in DDPG can provide a clearer optimization goal for the generative adversarial network. By designing a suitable reward function, the generator can be guided to generate samples that better meet the requirements of specific tasks. The DDPG algorithm mainly includes the Actor network and the Critic network.

[0074] The Actor network takes the current state information of the generator as input, and its task is to learn to generate appropriate actions to adjust the parameters of the generator so that it can generate more realistic samples. At the same time, the Actor adjusts its output actions according to the value estimation provided by the Critic network to optimize the generation strategy of the generator. By continuously adjusting the network parameters of the GAN generator, the samples it generates can gradually approach the distribution of real data.

[0075] The Critic network receives three pieces of information as input: fake data generated by the generator, real data, and the actions output by the Actor. By evaluating these inputs, it outputs a value estimate, which is used to evaluate the quality of the actions output by the actor network in the current state. The Critic's value estimate is fed back to the actor network to guide the actor network to output better actions, thereby indirectly optimizing the generator's generation strategy.

[0076] Furthermore, the ICNLCNDTL diagnostic model includes:

[0077] Signal processing module: used to combine the analysis of fault response vibration / current signal experience knowledge, construct the mechanism feature matrix TFSF, and encode the current signal into two-dimensional space using the ISDP method;

[0078] Dual-branch feature extraction module: used to construct a dual-branch feature extraction network based on IConvNeXt-LCNN, input the fault signal sample into the dual-branch feature extraction network based on IConvNeXt-LCNN for processing, and obtain the current signal diagnosis result and the vibration signal diagnosis result respectively;

[0079] Domain adaptation strategy module: used to adapt the discriminative joint distribution to DJP-MMD loss and conditional domain discriminator to reduce the probability distribution difference between the source domain and target domain deep features;

[0080] Motor state classifier: used to learn the characteristic patterns corresponding to motors in different health states, identify the association between different features and the motor health state, and judge the motor health state based on its features through the trained motor state classifier.

[0081] Specifically, the ICNLCNDTL diagnostic model includes a signal processing module, a dual-branch feature extraction module, a domain adaptation strategy, and a motor state classifier. In the signal processing module, the mechanism feature matrix TFSF is constructed by combining the analysis of the empirical knowledge of the fault response vibration / current signal, and the ISDP method is used to encode the current signal into a two-dimensional space; secondly, a dual-branch feature extraction module is designed, and the local detail features and mechanism features of the current / vibration signal are extracted and fused using a multi-scale fusion feature extraction module and a lightweight convolutional neural network; finally, the discriminant joint distribution adaptation and conditional domain discriminator are used to realize cross-domain fault diagnosis.

[0082] Domain Adaptation Strategy Module:

[0083] Since the data samples used in the training process of the fault diagnosis model are data under some working conditions (speed, power, operating conditions, etc.), and many new working conditions will appear in actual field diagnosis, the domain adaptation module is introduced to improve the migration and generalization ability of the model to adapt to more working conditions.

[0084] In this embodiment, in order to solve the problem of reduced diagnostic performance due to distribution differences in variable working condition data, the discriminant joint distribution adaptation DJP-MMD loss and the conditional domain discriminator are combined to narrow the probability distribution differences between the deep features of the source domain and the target domain, so as to improve the scenario adaptability of the model.

[0085] Furthermore, the dual-branch feature extraction network based on IConvNeXt-LCNN includes:

[0086] Multi-scale feature extraction sub-network: used to obtain local feature information of faults at different scales in ISDP images and fuse local fault information at each scale;

[0087] Lightweight convolutional neural sub-network: used to obtain deep features of TFSF.

[0088] The design details of the multi-scale feature extraction sub-network IConvNeXt and the lightweight convolutional neural network LCNN are as follows: Figure 4 .

[0089] Specifically, the multi-scale feature extraction sub-network IConvNeXt mines the deep local features contained in the ISDP image to obtain the local feature information of the fault at multiple scales. The asymmetric fusion module AFF is introduced to adaptively fuse the local fault information at each scale, further enhancing the model's ability to express the local characteristics of the fault. On the other hand, while capturing the local features, the lightweight convolutional neural network LCNN is used to process the mechanism feature matrix TFSF to obtain the deep features of TFSF.

[0090] Furthermore, the backbone network of the multi-scale feature extraction sub-network is a ConvNeXt network, each Stage module of the ConvNeXt network is formed by stacking a downsampling layer and a ConvNeXt Block, and the ordinary convolution layer in the downsampling layer in the Stage is replaced by an ODConv, wherein the convolution kernel size used by the ODConv contained in the Stage module is different, and the combination method of the downsampling layer in the Stage module is to first perform an ODConv operation, then perform layer normalization processing, and finally perform multi-scale fusion of the features output by the Stage module through an asymmetric feature fusion module;

[0091] This embodiment extracts the features after four Stage modules. The H×W×C of these features are different, and it is impossible to simply merge the features from a certain dimension. Therefore, an asymmetric feature fusion module is proposed to perform multi-scale feature fusion. The structure of AFF is as follows: Figure 5 shown.

[0092] The input of AFF1 is the features output by the first three stages, and the input of AFF2 is the features output by the last three stages. The shuffled attention mechanism is used to allow AFF to adaptively assign weights to features of different scales, and the convolutional neural network is combined to fuse multi-scale features. The mathematical expression of AFF is as follows:

[0093]

[0094] In the formula, Represents the output of the nth AFF; Represents the output features of the nth stage; upsampling (↑) and downsampling (↓) are applied to connect features of different scales.

[0095] Since the TFSF size is 4*5, which is very small, a too deep network structure may cause the model to learn useless features, so a lightweight convolutional neural network (LCNN) is designed to extract deep feature information. LCNN contains two regular convolutional layers and one fully connected layer, cancels the global pooling operation, the convolution kernel size is 3*3, the convolution step is set to 1, and the padding is set to 0 or 1.

[0096] Furthermore, the ICNLCNDTL diagnostic model combines the discriminant joint distribution adaptation DJP-MMD loss and the conditional domain discriminator to narrow the probability distribution difference;

[0097] Among them, the discriminant joint distribution adaptation DJP-MMD loss is obtained by introducing the joint probability difference JPD based on the traditional MMD definition; the conditional domain discriminator is used to apply the multilinear conditional adjustment method to improve the classification accuracy of the state classifier by optimizing the cross entropy between the deep features and the classifier predictions.

[0098] Discriminative joint distribution adaptation:

[0099] The optimization goal of existing domain adaptive distribution alignment methods focuses on improving the transferability of features, but ignores the discriminability of features, resulting in poor classification performance. Therefore, a discriminative joint distribution adaptation (Discriminative Joint Probability Maximum Mean Discrepancy, DJP-MMD) is proposed, which aims to simultaneously minimize the difference in joint probability distribution between features of the same class in different domains to achieve transferability, and maximize the difference in joint probability distribution between features of different classes in different domains to achieve discriminability.

[0100] Based on the traditional MMD definition, the Joint Probability Discrepancy (JPD) is introduced. Given c = {1, 2, ..., C} and They are X s and X t The label set of X is C, and C is the total number of labels. Define P(X|Y) and P(Y) as the class conditional probability and prior probability, respectively. Therefore, according to Bayes' theorem, D s and D t The JPD expression between them is as follows:

[0101]

[0102] Where M same-class JPD, M, represents the same class feature data in the source domain and the target domain. different-classJPD representing different classes of feature data in two domains.

[0103] Comparing the expression of JPD with the traditional MMD definition, although they are equivalent in theory, the difference between JPD and the traditional MMD definition is that the latter is based on the product of marginal probability and posterior probability. Formula (5) is calculated directly from the feature data without taking approximate calculations, and integrates class discriminability into distribution differences. Figure 6 This is the difference diagram between DJP-MMD and MMD.

[0104] Conditional Domain Discriminator:

[0105] When you need to align the joint distribution of categories and features (i.e., the marginal probability distribution P(x) and the conditional probability distribution Q(y|x)) at the same time, but do not satisfy It is insufficient to use the classic domain adversarial network to adapt the feature representation. As the number of model layers increases, the features learned by the model are transformed from general features to data set private features, and the transferability is significantly reduced. In order to solve this problem, the conditional adversarial generative network (Conditional Adversarial Domain Adaptation, CDAN) uses the category label predicted by the classifier as a constraint for model construction, and at the same time performs multilinear mapping on the category vector and the deep feature representation to obtain a joint vector constraint discriminator. CDAN contains two conditional adjustment methods, namely multilinear conditional adjustment and entropy adjustment. This embodiment mainly applies the multilinear conditional adjustment method to improve the classification accuracy of the state classifier by optimizing the cross entropy between the deep features and the classifier predictions.

[0106] Based on the idea of ​​CDAN, a conditional domain discriminator (CDD) is designed. In addition to focusing on the deep features of the source domain and the target domain, CDD also introduces the label information predicted by the state classifier to establish a domain adversarial adaptation network to reduce the joint distribution difference of features and labels. The framework of CDD is as follows Figure 7 As shown. The input of CDD is the feature and the classifier prediction label, and the output is a binary label (0 or 1) used to distinguish whether the feature is from the source domain or the target domain. The CDD in this embodiment consists of a fully connected layer, batch normalization, ReLU activation function, GRL and sigmoid activation function.

[0107] Furthermore, the loss function of the ICNLCNDTL diagnostic model includes a source domain classification loss function, a DJP-MMD loss function between the source domain and the target domain, and a conditional domain adversarial loss function;

[0108] The source domain classification loss function adopts the cross entropy loss function to measure the difference between the real label and the predicted label of the source domain, and the DJP-MMD loss function between the source domain and the target domain uses DJP-MMD to measure the distribution difference between the source domain and the target domain.

[0109] The ICNLCNDTL diagnosis model includes a dual-branch feature extractor G DFEN , Motor health status classifier G y and domain adaptation network G DA In order to train the ICNLCNDTL model, three loss functions are designed, namely, source domain classification loss DJP-MMD loss between source and target domains and conditional domain adversarial loss By reasonably adjusting the weights of the three losses, the joint loss L is minimized ICNLCNDTL The value of is used to update the model parameters to improve the cross-domain fault diagnosis performance.

[0110] Specifically, the source domain classification loss The motor health status classifier consists of a fully connected layer FC and a Softmax activation function. The loss function uses the cross entropy loss function to measure the difference between the true source domain label and the predicted label. Source domain classification loss The expression is:

[0111]

[0112] DJP-MMD loss between source and target domains Due to changes in working conditions, there is a large difference in the distribution of source and target domain datasets, which leads to a decrease in model diagnosis performance. DJP-MMD is used to measure the distribution difference between the source and target domains. The smaller the value, the closer the data distribution of the source domain and the target domain is. Based on formula (7), The expression is:

[0113]

[0114] Where μ is a trade-off parameter, greater than 0; M same-class and M differnent-class They are used to measure the transferability of the same class and the discriminability of different classes respectively;

[0115] Conditional Domain Adversarial Loss G d is the domain discriminator, domain discriminator G d Determine the domain label of the input sample features through adversarial training. Loss function The backpropagation path consists of backpropagating the update G DFEN The parameters and Gd The parameters of G are reversed by gradient DFEN and G d Create an adversarial relationship. The expression function is:

[0116]

[0117] In the formula, is a multilinear map.

[0118] The expression of ICNLCNDTL joint loss function is:

[0119]

[0120] Wherein, the hyperparameters α and β are weight adjustment coefficients.

[0121] By minimizing the joint loss function L ICNLCNDTL , use the back-propagation mechanism and stochastic gradient descent optimization algorithm to update the parameters of ICNLCNDTL. Assume that θ y ,θ d and θ DFEN They represent the parameters of the motor health status classifier, conditional domain discriminator and dual-branch feature extractor respectively, λ represents the learning rate of the diagnosis model, ▽ represents the gradient value calculated for the corresponding loss, and the parameters of each module are updated according to formula (10):

[0122]

[0123] The ICNLCNDTL diagnostic method designed in this embodiment is shown in Table 2 below.

[0124] Table 2

[0125]

[0126] Furthermore, the motor composite fault diagnosis model of multi-sensor signal collaborative fusion utilizes Lance distance to calculate credibility and an improved DS evidence fusion method based on Deng entropy to calculate uncertainty to fuse the current signal diagnosis result and the vibration signal diagnosis result.

[0127] The proposed IConvNeXt-LCNN dual-branch deep feature extraction transfer learning diagnosis method is used to diagnose the current signal and vibration signal respectively. On this basis, an improved DS evidence fusion method is designed to calculate the credibility using Lance distance and the uncertainty based on Deng entropy to fuse the diagnosis results of the current signal and vibration signal to complete the final fusion judgment.

[0128] The motor composite fault diagnosis framework based on the coordinated fusion of multi-sensor signals is as follows: Figure 8As shown, it includes four steps: signal generation and processing, evidence acquisition, decision fusion and state classification.

[0129] Signal generation and processing. For the limited number and types of fault current and vibration signal samples on site, a large number of fault signal samples that approximate the characteristics of real signal samples can be generated by combining the simulation model and the DDGAN generation network model. The original three-phase current signal is ISDP converted, statistical features are extracted and reconstructed into a TFSF matrix to form an ISDP-TFSF data set; the radial vibration signal and the axial vibration signal are SDP converted to form the X-SDP data set and the Y-SDP data set.

[0130] Evidence acquisition. The ISDP-TFSF of the current signal sample uses the ICNLCNDTL model to extract the target domain deep features and obtains the probability of each fault type through the Softmax activation function as evidence E1; the multi-scale feature extraction module IConvNeXt combined with the migration strategy (DJP-MMD and CDD) is used to obtain the target domain deep features for the radial vibration signal sample X-SDP and the axial vibration signal sample Y-SDP, and the predicted probability is obtained through the activation function Softmax as E2 and E3. {E1, E2, E3} constitute the original evidence set.

[0131] Decision fusion. The original evidence set is modified using the discount coefficient calculated by the credibility and uncertainty of the evidence, and the modified evidence is fused twice to obtain the final fusion result.

[0132] State classification. Based on the final evidence fusion result, the label corresponding to the largest probability value is taken as the classification result, and the model diagnosis accuracy is calculated.

[0133] Based on the idea of ​​resolving conflict problems by modifying the body of evidence, an improved DS evidence fusion method is designed, which uses Lance distance to calculate credibility and Deng entropy to calculate uncertainty. Firstly, the credibility and uncertainty of evidence are determined by calculating the Lance distance and Deng entropy between evidences; secondly, the discount coefficient is calculated based on the credibility and uncertainty; then, the original body of evidence is modified using the calculated discount coefficient; finally, the fusion of the modified evidence is realized by combining the DS combination rule.

[0134] Specifically, the Lance distance is used to measure the distance between evidence bodies. The smaller the value, the more similar the two evidence bodies are. First, define the recognition framework Θ = {A1, A2, A3, ... A K}, the system has n pieces of evidence E, and the corresponding mass functions are E1, E2, …, E i ,…,E n , where E i ={mi (A1),m i (A2),…m i (A K )}. For the body of evidence E i and E j The Lance distance is defined as:

[0135]

[0136] Calculate the Lance distances between N pieces of evidence respectively and construct an n×n Lance distance metric matrix D Lance Obviously, the diagonal elements indicate that the Lance distance between the same evidence is 0, D Lance Defined as:

[0137]

[0138] The similarity between the evidences is defined as s(E i ,E j )=1-d ij (E i ,E j ), then the similarities between the N pieces of evidence can form a similarity matrix S of size n×n Lance , similarly S Lance The diagonal elements of S are 1, Lance is defined as:

[0139]

[0140] s ij The larger the value, the more evidence E i ,E j The higher the degree of consistency between them, the more evidence there is. Define the remaining evidence to be i The total similarity Sup(E i ) is the sum of the squares of all elements in each row of the matrix except the diagonal elements and then takes the square root. The expression is:

[0141]

[0142] For the total similarity Sup(E i ) is normalized, we can get evidence E i Credibility Rel(E i ):

[0143]

[0144] Calculation of evidence uncertainty based on Deng entropy:

[0145] Information entropy is an important concept in information theory, which is mainly used to measure the uncertainty or amount of information. Shannon, the founder of information theory, proposed Shannon entropy to measure the uncertainty of evidence. i The Deng entropy is defined as:

[0146]

[0147] In the formula, |A p | is a subset A p The number of elements contained. The previous item is m i (A p ) is a measure of the complete nonspecificity of the evidence, and the latter is a measure of the incongruity between the focal elements of the evidence.

[0148] When|A p |=1, Deng entropy degenerates into Shannon entropy, which is expressed as:

[0149]

[0150] Subset A p The more elements contained in Evidence E i The larger the Deng entropy value, the more evidence E i The higher the uncertainty, the more evidence E i The corresponding weight should be smaller; otherwise, evidence E i The smaller the Deng entropy value, the more evidence E i The lower the uncertainty, the greater the corresponding weight should be. i The weight is 0, and the evidence E is calculated using the exponential form of Deng entropy. i The weight value of:

[0151]

[0152] As with the trustworthiness, normalizing formula (18) yields evidence E i Uncertainty Unc i (E i )for:

[0153]

[0154] Weighted fusion of evidence based on credibility and uncertainty:

[0155] Evidence E i Credibility Rel(E i ) value, the greater the support of the evidence by other evidence, and the higher the weight should be given when weighted fusion of evidence; E iThe smaller the uncertainty, the smaller the amount of information provided by the evidence, the clearer the evidence, and the higher the weight should be. The weight value of the evidence is determined by combining the credibility and uncertainty, and each body of evidence is modified according to the weight value, and finally fused n-1 times according to the DS evidence combination rule. The specific steps are:

[0156] Step 1: Calculate evidence E according to formulas (11) to (19) i Credibility Rel(E i );

[0157] Step 2: Calculate evidence E according to formulas (16) to (19) i Uncertainty Unc i (E i );

[0158] Step 3: Calculate the fusion weight value, i.e., the discount coefficient W, based on the credibility and uncertainty calculated in steps 1 and 2. i , discount factor W i Defined as And the discount coefficient is normalized:

[0159]

[0160] Step 4: According to the discount factor W i Revised BPA, Revised Evidence E i The corresponding mass function value

[0161]

[0162] Step 5: Apply the DS evidence combination rule to fuse the revised evidence n-1 times and obtain the evidence fusion result m fused , that is, the fault diagnosis result.

[0163] When a motor fails, the relatively balanced electromagnetic relationship inside is destroyed. Traditional methods such as multi-loop analysis cannot deeply analyze the complex electromagnetic relationship of the motor. Therefore, it is necessary to establish an equivalent simulation model based on the understanding of the asynchronous motor failure mechanism and understand the current characteristics of the faulty motor.

[0164] Deep Deterministic Policy Gradient (DDPG) is specially designed to handle problems in continuous action space. It can effectively learn the optimal strategy in continuous space through the collaborative work of policy network and value network. This paper introduces DDPG algorithm into GAN, so that the generator can be adjusted more efficiently in continuous parameter space, improving the quality and diversity of generated samples.

[0165] In order to improve the diagnostic accuracy and generalization ability of the fault diagnosis model in actual industrial applications, a motor cross-domain composite fault diagnosis method ICNLCNDTL based on IConvNeXt-LCNN dual-branch deep feature extraction is proposed. Firstly, twenty time-frequency statistical features of the test bench current signal are extracted and reconstructed, and the global statistical feature matrix TFSF is constructed. The three-phase current signal is converted into a two-dimensional image using the ISDP image coding method; then a dual-branch feature extraction network DFEN is established to extract and fuse local features and global features to enhance the fault feature expression ability of the model; finally, the discriminant joint distribution adaptation DJP-MMD and the conditional domain discriminator CDD are introduced to use the advantages of distribution alignment and discriminant alignment to reduce the difference in feature distribution probability between domains;

[0166] Aiming at the problem that the diagnostic model built by relying only on a single current signal has poor diagnostic effect on compound faults under variable working conditions, the present invention establishes a cross-domain fault diagnosis method for motors with multiple faults and unified multi-source information based on the theory of multi-source information fusion:

[0167] Since vibration signals are more sensitive to mechanical faults, vibration signals and current signals are introduced for joint diagnosis. Considering that vibration signals are single-phase, the SDP method is used for image encoding to obtain X-SDP images and Y-SDP images. The current signal samples are processed using the time-frequency statistical feature construction method and the ISDP image encoding method to obtain ISDP&TFSF inputs.

[0168] ICNLCNDTL is used to perform cross-domain diagnosis on the samples ISDP&TFSF corresponding to the current signal, and the predicted probability values ​​of the target domain samples are obtained as evidence; IConvNeXt combined with the migration strategy (DJP-MMD&CDD) is used for the X-SDP image and the Y-SDP image to obtain the predicted probability values ​​of the target domain samples as evidence E1; IConvNeXt combined with the migration strategy (DJP-MMD&CDD) is used for the X-SDP image and the Y-SDP image to obtain the predicted probability values ​​of the target domain samples as evidence E2 and E3. {E1, E2, E3} constitute the original evidence;

[0169] The credibility of evidence is calculated using Lance distance, the uncertainty of evidence is calculated based on Deng entropy theory, and the evidence discount coefficient is calculated according to the credibility and uncertainty to correct the original evidence; the DS evidence combination rule is used to fuse the corrected evidence to obtain the final diagnosis result;

[0170] The present invention produces the following technical effects:

[0171] In the simulation modeling and analysis of asynchronous motors, the finite element analysis method can convert actual engineering problems into simple small problems, and it has high calculation accuracy and good adaptability, making it a practical engineering analysis method.

[0172] The DDPG algorithm is introduced into GAN to design an improved generative network DDGAN. Compared with the random parameter update method in traditional GAN, in the DDGAN network model, DDPG can adjust the parameters of the generator in a targeted manner according to the current state and reward signal, so that it can be optimized faster in the direction of generating more realistic samples. The reward mechanism in DDPG can provide a clearer optimization goal for the generative adversarial network. By designing a suitable reward function, the generator can be guided to generate samples that better meet the requirements of specific tasks.

[0173] On the SQI-MFS test platform, the same working condition experiment and cross-working condition experiment were carried out respectively. The results of the same working condition experiment show that the dual-branch feature extraction network DFEN has excellent feature extraction ability and effectively improves the model's ability to express multiple faults. The results of the variable working condition experiment show that the designed DJP-MMD loss combined with the migration strategy of the conditional domain discriminator CDD improves the cross-domain diagnosis accuracy and has good generalization ability.

[0174] Based on the multi-source information fusion, the diagnosis of motor compound fault migration is evaluated by fusion of the diagnosis results of vibration and current signals to complete the diagnosis of motor compound fault. The same working condition and cross-working condition experiments are carried out on the SQI-MFS test platform. The experimental results show that after evidence fusion, the average diagnostic accuracy in the same working condition experiment is 99.75%; the average diagnostic accuracy in the variable working condition task is as high as 99%.

[0175] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A motor load fault diagnosis method driven by empirical knowledge and data fusion, characterized in that: include: Construct the empirical knowledge feature space of current / vibration signal and obtain several simulation signal samples; Inputting the simulation signal sample into the improved generation network DDGAN for processing to obtain an expanded simulation signal sample, and combining the actual fault sample with the expanded simulation signal sample to obtain a fault signal sample; Construct an ICNLCNDTL diagnostic model, input the fault sample signal into the ICNLCNDTL diagnostic model for diagnostic processing, and obtain current signal diagnostic results and vibration signal diagnostic results; wherein the ICNLCNDTL diagnostic model performs feature extraction through a dual-branch feature extraction network of IConvNeXt-LCNN; The motor composite fault diagnosis model based on the coordinated fusion of multiple sensor signals is used to fuse the current signal diagnosis result and the vibration signal diagnosis result to obtain a fault diagnosis result.

2. The method according to claim 1, characterized in that Obtaining the plurality of simulation signal samples comprises: A multi-fault finite element simulation model of an asynchronous motor is established, the stator current characteristics during faults are analyzed, and vibration and current signals under various operating conditions and fault types are generated. The fault mechanism, empirical knowledge, and time-frequency signal analysis method are combined to construct the empirical knowledge feature space of the current / vibration signal, and the time-frequency detail features of the SDP image, the time-frequency detail features of the ISDP image, and the TFSF mechanism knowledge features are obtained respectively.

3. The method according to claim 1, characterized in that The improved generation network DDGAN uses the DDPG algorithm to optimize the generation strategy of the generator and adjusts the parameters of the generator according to the current state and reward signal.

4. The method according to claim 1, characterized in that: The ICNLCNDTL diagnostic model includes: Signal processing module: used to combine the analysis of fault response vibration / current signal experience knowledge, construct the mechanism feature matrix TFSF, and encode the current signal into two-dimensional space using the ISDP method; Dual-branch feature extraction module: used to build a dual-branch feature extraction network based on IConvNeXt-LCNN to extract and fuse the local detail features and mechanism features of the current signal and vibration signal; Domain adaptation strategy module: used to adapt the discriminative joint distribution to DJP-MMD loss and conditional domain discriminator to reduce the probability distribution difference between the source domain and target domain deep features; Motor state classifier: used to learn the characteristic patterns corresponding to motors in different health states, identify the association between different features and the motor health state, and judge the motor health state based on its features through the trained motor state classifier.

5. The method according to claim 4, characterized in that The dual-branch feature extraction network based on IConvNeXt-LCNN includes: Multi-scale feature extraction sub-network: used to obtain local feature information of faults at different scales in ISDP images and fuse local fault information at each scale; Lightweight convolutional neural sub-network: used to obtain deep features of TFSF.

6. The method according to claim 5, characterized in that The backbone network of the multi-scale feature extraction sub-network is a ConvNeXt network, each Stage module of the ConvNeXt network is formed by stacking a downsampling layer and a ConvNeXt Block, and the ordinary convolution layer in the downsampling layer is replaced by an ODConv, wherein the convolution kernel size used by the ODConv contained in the Stage module is different, and the combination method of the downsampling layer in the Stage module is to first perform an ODConv operation, then perform layer normalization processing, and finally perform multi-scale fusion of the features output by the Stage module through an asymmetric feature fusion module; The lightweight convolutional neural sub-network includes two convolutional layers and one fully connected layer.

7. The method according to claim 4, characterized in that The ICNLCNDTL diagnostic model combines the discriminative joint distribution adaptation DJP-MMD loss and the conditional domain discriminator to narrow the probability distribution difference; Among them, the discriminant joint distribution adaptation DJP-MMD loss is obtained by introducing the joint probability difference JPD based on the MMD definition; the conditional domain discriminator is used to apply the multilinear conditional adjustment method to improve the classification accuracy of the state classifier by optimizing the cross entropy between the deep features and the classifier predictions.

8. The method according to claim 7, characterized in that The loss function of the ICNLCNDTL diagnostic model includes a source domain classification loss function, a DJP-MMD loss function between the source domain and the target domain, and a conditional domain adversarial loss function; The source domain classification loss function adopts the cross entropy loss function to measure the difference between the real label and the predicted label of the source domain, and the DJP-MMD loss function between the source domain and the target domain uses DJP-MMD to measure the distribution difference between the source domain and the target domain.

9. The method according to claim 1, characterized in that: The motor composite fault diagnosis model of multi-sensor signal collaborative fusion utilizes Lance distance to calculate credibility and an improved DS evidence fusion method based on Deng entropy to calculate uncertainty to fuse the current signal diagnosis results and the vibration signal diagnosis results.

Images