Fault Diagnosis Method and System Based on Interpretable Multiplicative Convolutional Networks

Abstract

The application belongs to the technical field of fault diagnosis, and specifically discloses a fault diagnosis method and system based on a releaseable multiplicative convolution network, which comprises the following steps: collecting vibration signals under different health states of equipment, and constructing a frequency spectrum sample data set; designing a multiplicative feature separator, a convolution feature extractor and a discriminator to obtain a releaseable multiplicative convolution fault diagnosis evaluation model; taking the frequency spectrum sample data set as the input of the releaseable multiplicative convolution fault diagnosis evaluation model, and updating and optimizing the releaseable multiplicative convolution fault diagnosis evaluation model. According to the technical scheme, the multiplicative feature separator, the convolution feature extractor and the discriminator are used to enable the model to make scientific and reasonable decisions and recognitions based on fault information, so that the purpose of model releaseability is achieved.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, and relates to a fault diagnosis method and system based on interpretable multiplicative convolutional networks. Background Technology

[0002] In recent years, with the continuous innovation and development of science and technology and the rapid development of social industrialization, more and more industrial systems have been installed in large-scale industrial production settings. In this application environment, they create irreplaceable and enormous productivity, making the maintenance of these devices particularly important and challenging. However, many hidden influencing factors can cause these systems to malfunction or even fail. Therefore, governments and enterprises are paying increasing attention to the condition monitoring and fault diagnosis of these important systems.

[0003] Traditional intelligent fault diagnosis models have two main drawbacks:

[0004] (1) Traditional intelligent fault diagnosis models are based entirely on data and lack in-depth analysis of the potential generation mechanism of equipment vibration signals, resulting in the model structure not being well adapted to the data characteristics of vibration signals.

[0005] (2) Traditional intelligent fault diagnosis models increase the fault identification capability of the model by deep stacking, but make the higher layers more abstract and without clear physical meaning, and the model lacks explanation.

[0006] The above issues raise many questions about the model's decision-making mechanism and results, making it difficult for researchers and technicians to understand what data characteristics the intelligent fault diagnosis model uses to make scientific and reasonable decisions, and also limiting the further development and industrial application of intelligent fault diagnosis technology. Summary of the Invention

[0007] The purpose of this invention is to provide a fault diagnosis method and system based on interpretable multiplicative convolutional networks, so as to solve the problem that the model also lacks interpretation.

[0008] To achieve the above objectives, the basic solution of this invention is: a fault diagnosis method based on interpretable multiplicative convolutional networks, comprising the following steps:

[0009] Vibration signals from the equipment under different health conditions were collected, and a spectrum sample dataset was constructed.

[0010] A multiplicative feature separator, a convolutional feature extractor, and a discriminator are designed to obtain an interpretable multiplicative convolution fault diagnosis and evaluation model.

[0011] The spectral sample dataset is used as input to the interpretable multiplicative convolution fault diagnosis and evaluation model to update and optimize the model.

[0012] The working principle and beneficial effects of this basic scheme are as follows: The interpretable multiplicative convolutional fault diagnosis and assessment model consists of three layers and has the significant advantage of being lightweight. To enable the interpretable multiplicative convolutional fault diagnosis and assessment model to make scientific and reasonable decisions and identifications based on fault information, a multiplicative feature separator is designed in the first layer of the model to separate fault-sensitive information in the vibration signal. This ensures that subsequent feature extraction and decision-making are based on the extracted fault information, thereby achieving the goal of model interpretability.

[0013] Furthermore, the specific method for constructing the spectrum sample dataset is as follows:

[0014] Vibration signals of the device under different health conditions are collected, the sampling frequency is set to Fs, and the number of health condition categories is K;

[0015] Each vibration signal is divided into multiple vibration samples, and the length of each vibration sample is L = Fs. The sampling duration corresponding to each vibration sample is set.

[0016] Perform an FFT transform on each vibration sample to obtain its corresponding spectral sample x∈R 1×L And record the sample labels.

[0017] It is easy to operate and obtains spectrum samples, which is convenient for subsequent use.

[0018] Furthermore, the method for designing a multiplication feature separator is as follows:

[0019] The multiplicative feature separator contains several analytical multiplicative filter kernels to separate fault-sensitive features from the spectral samples. Each multiplicative filter kernel has two learnable parameters: the center frequency ω. c and the bandwidth penalty coefficient λ, and the feature map h corresponding to the k-th multiplicative filter kernel of the multiplicative feature separator. k for:

[0020] h k =M(ω) ck ,λ k )·X(ω)

[0021] Where X(ω) is the training set composed of spectral samples, ω is the frequency; M(·) is the multiplication filter kernel; ω ck and λ k These are the center frequency and bandwidth penalty coefficient of the k-th multiplication filter core, respectively;

[0022] Let the depth of the multiplicative feature separator be N, and the number of multiplicative filter kernels be N. Stack the N feature maps to form the output mapping H of the multiplicative feature separator:

[0023] H = [h1; h2; ...; h N ].

[0024] By using an analytical multiplication filter kernel, fault-sensitive information in the vibration signal is separated, thus ensuring that subsequent feature extraction and decision-making are based on the extracted fault information.

[0025] Furthermore, to enable each multiplicative filter kernel to separate more discriminative fault information and improve the diversity of feature maps, an anti-aliasing strategy is introduced into the multiplicative feature separator. The k-th feature map h... k Rewritten as:

[0026]

[0027] Among them, h i This represents the i-th feature map.

[0028] An anti-aliasing strategy is introduced into the multiplicative feature separator, which forces each filter kernel to separate more discriminative fault information and improves the diversity of feature maps.

[0029] Furthermore, to achieve feature separation, the analytical expression for designing the multiplicative filter kernel can be:

[0030]

[0031] and / or

[0032] Where W(·) and G(·) represent the Wiener filter kernel and the Gaussian filter kernel, respectively, and α and σ are their bandwidth penalty coefficients; ω c It is the center frequency parameter.

[0033] It has a simple structure and is easy to use.

[0034] Furthermore, the method for designing a convolutional feature extractor is as follows:

[0035] Construct a 1D convolutional layer to extract high-dimensional features from the output map H of the multiplicative feature separator. The convolutional kernel has 64 output channels, and the convolution operation is represented as follows:

[0036] y = ReLU(w conv *H+b conv )

[0037] Among them, w conv and b conv These are the weights and biases of the convolutional feature extractor, respectively; ReLU(·) is the activation function, and its activation criterion is:

[0038]

[0039] Where y represents the node response output by the convolutional layer;

[0040] The feature y is reduced in dimensionality using the adaptive max pooling function AdaptiveMaxPooling(·), retaining the 10 largest output nodes. The pooling process is as follows:

[0041]

[0042] Among them, y l It is the convolution output vector of the l-th channel. This represents the i-th output of the l-th channel of the convolutional layer after adaptive max pooling;

[0043] The output Z of each channel of the convolutional feature extractor l Flattening the feature vector:

[0044] y←Fltten({y l}), l=1,2,...,64

[0045] Where Z∈R 640×1 It is the flattened feature vector, and Fltten indicates the flattening process.

[0046] Extracting the required feature vectors is beneficial for subsequent fault diagnosis.

[0047] Furthermore, the method for designing the discriminator is as follows:

[0048] Using a single fully connected layer, the feature vector Z in the convolutional feature extractor is reduced to the number of categories K:

[0049] p = w fc Z+b fc

[0050] Among them, w fc and b fc These are the weight matrix and bias vector of the fully connected layer, respectively; p∈R K×1 It is the reasoning result of the interpretable multiplicative convolution fault diagnosis and evaluation model;

[0051] The probability of each health status is:

[0052]

[0053] in, Let P be the probability value for assessing the i-th type of health status. j Let represent the output value of the j-th neuron at the output end of the fully connected layer, and the predicted probability of the j-th type of fault; e represents the natural exponent; P i This represents the output value of the i-th neuron at the output end of the fully connected layer, and the predicted probability of the i-th type of fault.

[0054] The fault diagnosis results are obtained through the discriminator, and the operation is simple.

[0055] Furthermore, the method for updating the interpretable multiplicative convolution fault diagnosis and evaluation model is as follows:

[0056] The cross-entropy loss function is used as the optimization function for the interpretable multiplicative convolutional fault diagnosis and evaluation model. By minimizing the cross-entropy loss, the interpretable multiplicative convolutional fault diagnosis and evaluation model is updated. The update function is:

[0057]

[0058] Where X is the training set composed of spectral samples, and θ = {ω c ,λ,w conv ,b conv ,w fc ,b fc} represents the model parameters of the interpretable multiplicative convolution fault diagnosis and evaluation model, ω. c λ and w are the center frequency and bandwidth penalty coefficient of the multiplicative feature separator, respectively. conv and b conv These are the weights and biases of the convolutional feature extractor, w fc and b fc These are the weight matrix and bias vector of the discriminator, respectively; and These are the actual fault label value and the predicted label value, respectively; K is the number of health status categories; argmin is the mathematical function representation symbol.

[0059] According to the chain rule, the gradient of the model parameters is calculated and updated layer by layer. The parameters of the multiplicative feature separator in the first layer are updated as follows:

[0060]

[0061] in, It is a partial derivative operator, η is the learning rate; L is the length of the vibration sample; For the probability value of health status assessment; p∈R K×1 This is the inference result of the interpretable multiplicative convolutional fault diagnosis and evaluation model; y represents the node response of the convolutional layer output; H is the output mapping of the multiplicative feature separator. and It is the partial derivative of the analytical expression of the filter kernel with respect to the frequency parameter and the bandwidth penalty coefficient parameter; ω ci and λ i Here, are the center frequency and bandwidth penalty coefficient of the i-th multiplication filter core, respectively, and M represents the multiplication filter core;

[0062] For the Wiener filter kernel and the Gaussian filter kernel, their partial derivatives are calculated respectively:

[0063]

[0064]

[0065]

[0066]

[0067] Where W(·) and G(·) represent the Wiener filter kernel and the Gaussian filter kernel, respectively, and α and σ are their bandwidth penalty coefficients; α i and σ ci ω represents the bandwidth penalty coefficients of the i-th Wiener filter kernel and the i-th Gaussian filter kernel, respectively, and ω represents the frequency.

[0068] Optimize the model to enhance the accuracy of model fault diagnosis.

[0069] The present invention also provides a fault diagnosis system based on an interpretable multiplicative convolutional network, including a data acquisition module and a fault diagnosis module. The data acquisition module is used to acquire vibration signals of the equipment under different health conditions. The output end of the data acquisition module is connected to the input end of the fault diagnosis module. The fault diagnosis module executes the method described in the present invention to perform fault diagnosis.

[0070] This system is simple to operate and easy to use for equipment fault diagnosis.

[0071] The present invention also provides a gearbox fault testing system, including a two-stage gearbox, a magnetic powder brake, a load controller, a remote terminal, a motor, and the fault diagnosis system described in the present invention;

[0072] The magnetic powder brake is connected to the secondary gearbox via a coupling, and the output end of the load controller is connected to the control end of the magnetic powder brake.

[0073] The first control signal output terminal of the remote terminal is connected to the load controller, and the second control signal output terminal of the remote terminal is connected to the control terminal of the motor after being amplified. The rotating shaft of the motor is connected to the secondary gearbox.

[0074] The data acquisition module of the fault diagnosis system is connected to the secondary gearbox and is used to collect the vibration signal of the faulty gear in the secondary gearbox.

[0075] The gearbox fault testing system uses a fault diagnosis system to detect gear faults, which is beneficial for fault diagnosis and equipment maintenance. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating the fault diagnosis method based on interpretable multiplicative convolutional networks of the present invention.

[0077] Figure 2This is a schematic diagram of the gearbox fault testing system of the present invention;

[0078] Figure 3 This is a schematic diagram of the health spectrum sample in the extensibility verification of the gearbox fault testing system of the present invention;

[0079] Figure 4 This is a schematic diagram of the crack fault spectrum sample in the releasability verification of the gearbox fault testing system of the present invention.

[0080] Figure 5 This is a schematic diagram illustrating the interpretability of the gearbox fault testing system of the present invention. Detailed Implementation

[0081] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0082] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0083] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0084] This invention discloses a fault diagnosis method based on interpretable multiplicative convolutional networks (MCNs). Combining the multi-scale characteristics of vibration signals, an interpretable multiplicative convolutional network (MCN) architecture is proposed. Figure 1 As shown, the method includes the following steps:

[0085] Vibration signals from the equipment under different health conditions were collected, and a spectrum sample dataset was constructed.

[0086] A multiplicative feature separator, a convolutional feature extractor, and a discriminator are designed to obtain an interpretable multiplicative convolution fault diagnosis and evaluation model.

[0087] The spectral sample dataset is used as input to the interpretable multiplicative convolutional fault diagnosis and evaluation model to update and optimize the model (updating stops when the maximum number of iterations is reached or the cross-entropy loss is less than 0.0001).

[0088] In a preferred embodiment of the present invention, the specific method for constructing the spectrum sample dataset is as follows:

[0089] Vibration signals of the device under different health conditions are collected, the sampling frequency is set to Fs, and the number of health condition categories is K;

[0090] Each vibration signal is divided into multiple vibration samples, each with a length of L = Fs. The sampling duration for each vibration sample is set to 1s to ensure the spectral resolution after FFT transformation.

[0091] Perform an FFT transform on each vibration sample and label each sample with a number, such as 0, 1, 2, 3, 4, to obtain its corresponding spectral sample x∈R. 1×L The sample labels are recorded. The sample labels correspond to the fault classification numbers. Taking gear fault identification as an example, if the health status corresponding to the spectrum sample includes five categories: normal, spalling, tooth root crack, tooth surface pitting, and broken tooth, then the corresponding labels are 0, 1, 2, 3, and 4.

[0092] In a preferred embodiment of the present invention, the method for designing the multiplication feature separator is as follows:

[0093] The multiplicative feature separator comprises several analytical multiplicative filter kernels used to separate fault-sensitive features from the spectral samples. Each multiplicative filter kernel has two learnable parameters: the center frequency ω. c and the bandwidth penalty coefficient λ, and the feature map h corresponding to the k-th multiplicative filter kernel of the multiplicative feature separator. k for:

[0094] h k =M(ω) ck ,λ k )·X(ω)

[0095] Where X(ω) is the training set composed of spectral samples, ω is the frequency, and the spectral samples can be represented as discrete functions of frequency; M(·) is the multiplication filter kernel; ω ck and λ k These are the center frequency and bandwidth penalty coefficient of the k-th multiplication filter core, respectively;

[0096] Let the depth of the multiplicative feature separator be N, and the number of multiplicative filter kernels be N. Stack the N feature maps to form the output mapping H of the multiplicative feature separator:

[0097] H = [h1; h2; ...; h N].

[0098] To achieve feature separation, the analytical expression for designing the multiplicative filter kernel can be:

[0099]

[0100] and / or

[0101] Where W(·) and G(·) represent the Wiener filter kernel and the Gaussian filter kernel, respectively, and α and σ are their bandwidth penalty coefficients; ω c It is the center frequency parameter.

[0102] By using an analytical multiplication filter kernel, fault-sensitive information in the vibration signal is separated, thus ensuring that subsequent feature extraction and decision-making are based on the extracted fault information.

[0103] In a preferred embodiment of the present invention, in order to force each filter kernel to separate more discriminative fault information and improve the diversity of feature maps, an anti-aliasing strategy is introduced in the multiplicative feature separator. The k-th feature map h... k Rewritten as:

[0104]

[0105] Among them, h i Let L represent the i-th feature map, with a latitude of 1×L.

[0106] In a preferred embodiment of the present invention, the method for designing a convolutional feature extractor is as follows:

[0107] Construct a 1D convolutional layer to extract high-dimensional features from the output map H of the multiplicative feature separator. The convolutional kernel has 64 output channels, and the convolution operation is represented as follows:

[0108] y = ReLU(w conv *H+b conv )

[0109] Among them, w conv and b conv These are the weights and biases of the convolutional feature extractor, respectively; ReLU(·) is the activation function, and its activation criterion is:

[0110]

[0111] Where y represents the node response output by the convolutional layer;

[0112] The feature y is reduced in dimensionality using the adaptive max pooling function AdaptiveMaxPooling(·), retaining the 10 largest output nodes. The pooling process is as follows:

[0113]

[0114] Among them, y l It is the convolution output vector of the l-th channel. This represents the i-th output of the l-th channel of the convolutional layer after adaptive max pooling;

[0115] The output Z of each channel of the convolutional feature extractor l Flattening the feature vector:

[0116] y←Fltten({y l}), l=1,2,...,64

[0117] Where Z∈R 640×1 This is the flattened feature vector, where Fltten indicates the flattening process. The flattening process involves concatenating the outputs of the 64 convolutional layers after pooling in sequence, resulting in a concatenated dimension of 1×640.

[0118] In a preferred embodiment of the present invention, the method for designing the discriminator is as follows:

[0119] Using a single fully connected layer, the feature vector Z in the convolutional feature extractor is reduced to the number of categories K. The operation process is as follows:

[0120] p = w fc Z+b fc

[0121] Among them, w fc and b fc These are the weight matrix and bias vector of the fully connected layer, respectively; p∈R K×1 It is the reasoning result of the interpretable multiplicative convolution fault diagnosis and evaluation model;

[0122] The probability of each health status is:

[0123]

[0124] in, Let P be the probability value for assessing the i-th type of health status. j Let represent the output value of the j-th neuron at the output end of the fully connected layer, and the predicted probability of the j-th type of fault; e represents the natural exponent; P i This represents the output value of the i-th neuron at the output end of the fully connected layer, and the predicted probability of the i-th type of fault.

[0125] In a preferred embodiment of the present invention, the method for updating the interpretable multiplicative convolution fault diagnosis and evaluation model is as follows:

[0126] The cross-entropy loss function is used as the optimization function for the interpretable multiplicative convolutional fault diagnosis and evaluation model. By minimizing the cross-entropy loss, the interpretable multiplicative convolutional fault diagnosis and evaluation model is updated. The update function is:

[0127]

[0128] Where X is the training set composed of spectral samples, and θ = {ω c ,λ,w conv ,b conv ,w fc ,b fc} represents the model parameters of the interpretable multiplicative convolution fault diagnosis and evaluation model, ω. c λ and w are the center frequency and bandwidth penalty coefficient of the multiplicative feature separator, respectively. conv and b conv These are the weights and biases of the convolutional feature extractor, w fc and b fc These are the weight matrix and bias vector of the discriminator, respectively; and These are the actual fault label value and the predicted label value, respectively. The actual fault label is known during data preprocessing. Taking gear fault identification as an example, if the health status corresponding to the spectrum sample includes 5 categories (K=5): normal, spalling, tooth root crack, tooth surface pitting, and broken tooth, then the actual labels are 0, 1, 2, 3, and 4, respectively. During data preprocessing, the actual labels of the spectrum samples obtained from the corresponding vibration signals are known. K is the number of health status categories; argmin is the mathematical function notation, below which are the optimization parameters, followed by the loss function expression.

[0129] According to the chain rule, the gradient of the model parameters is calculated and updated layer by layer. The parameters of the multiplicative feature separator in the first layer are updated as follows:

[0130]

[0131] in, It is a partial derivative operator, η is the learning rate; L is the length of the vibration sample; For the probability value of health status assessment; p∈R K×1 This is the inference result of the interpretable multiplicative convolutional fault diagnosis and evaluation model; y represents the node response of the convolutional layer output; H is the output mapping of the multiplicative feature separator. and It is the partial derivative of the analytical expression of the filter kernel with respect to the frequency parameter and the bandwidth penalty coefficient parameter; ω ci and λ i Here, are the center frequency and bandwidth penalty coefficient of the i-th multiplication filter core, respectively, and M represents the multiplication filter core;

[0132] For the Wiener filter kernel and the Gaussian filter kernel, their partial derivatives are calculated respectively:

[0133]

[0134]

[0135]

[0136]

[0137] Where W(·) and G(·) represent the Wiener filter kernel and the Gaussian filter kernel, respectively, and α and σ are their bandwidth penalty coefficients; α i and σ ci ω represents the bandwidth penalty coefficients of the i-th Wiener filter kernel and the i-th Gaussian filter kernel, respectively, and ω represents the frequency.

[0138] This invention also provides a fault diagnosis system based on an interpretable multiplicative convolutional network, including a data acquisition module and a fault diagnosis module. The data acquisition module is used to acquire vibration signals of the equipment under different health states. The output of the data acquisition module is electrically connected to the input of the fault diagnosis module. The fault diagnosis module executes the method described in this invention to perform fault diagnosis. Using this system for equipment fault diagnosis is simple to operate and easy to use.

[0139] This invention also provides a gearbox fault testing system, such as... Figure 2 As shown, the system includes a secondary gearbox, a magnetic particle brake, a load controller, a remote terminal, a motor, and the fault diagnosis system described in this invention. The magnetic particle brake is connected to the secondary gearbox via a coupling. A speed-torque sensor is mounted on the coupling, and its signal output is electrically connected to the signal input of the remote terminal. The output of the load controller is connected to the control terminal of the magnetic particle brake. The first control signal output of the remote terminal is electrically connected to the load controller, and the second control signal output of the remote terminal is amplified and then electrically connected to the control terminal of the motor. The motor's rotating shaft is fixedly connected to the rotating shaft of the secondary gearbox. The remote terminal can be a computer or other similar device. The data acquisition module of the fault diagnosis system is electrically connected to the secondary gearbox and is used to collect vibration signals from the faulty gear within the secondary gearbox. An acceleration sensor is mounted on the faulty gear to collect corresponding acceleration signals, which is beneficial for control.

[0140] For example, the health status of gears in a secondary gearbox includes normal, cracked, spalling, pitting, and broken teeth. The input speed varies from 200 rpm to 1100 rpm, with a speed gradient of 50 rpm, and the vibration signal acquisition frequency Fs = 2560 Hz. Interpretability verification: The health spectrum samples at 1100 rpm and the crack fault spectrum samples are input into the interpretability multiplicative convolutional fault diagnosis and evaluation model. The Wiener filter kernel and its output feature map are compared before and after training. Figure 3 and Figure 4 As shown in the figure, f r1 f m1 and f m2 These are the shaft frequency and gear meshing frequency, from which their amplitude information and band frames can be observed, revealing fault sensitivity. Through training, the Wiener filter kernel, driven by frequency domain fault knowledge, accurately converges to these frequency bands, enabling the extraction of fault-related frequency domain knowledge to guide the MCN model in making interpretable and evidence-based decisions. Figure 5 As shown, the interpretability of the binary classification-based MCN fault diagnosis model is illustrated. (a) shows the uniformly initialized Wiener filter kernels in the binary classification MCN. The MCN sets up 8 filter kernels, with their center frequencies evenly spaced within the bandwidth of interest, and each having a unit amplitude response at its respective center frequency. (b) shows the learned filter kernels and their feature mappings to healthy spectrum samples. (c) shows the learned filter kernels and their feature mappings to crack fault spectrum samples. The dashed lines in (b) and (c) describe the learned filter kernels, while the solid lines describe the modes mapped from the 8 filter kernels. The difference between Figure (b) and Figure (c) is that Figure (b) uses healthy spectrum samples from the test set as input into the trained binary classification MCN to obtain the feature mappings; Figure (c) uses crack fault spectrum samples from the test set as input into the trained binary classification MCN to obtain the feature mappings. Therefore, the dashed lines in Figures (b) and (c) are in the same position, both describing the learned Wiener filter kernels, while the modes (solid lines) are different because the input test spectrum samples are different.

[0141] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0142] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A fault diagnosis method based on a releasability multiplication convolution network, characterized by, The method comprises the following steps: Collect vibration signals under different health states of the equipment, and construct a frequency spectrum sample data set; A multiplication feature separator, a convolution feature extractor, and a discriminator are designed to obtain a releaseable multiplication convolution fault diagnosis evaluation model; The frequency spectrum sample data set is taken as an input of the releaseable multiplication convolution fault diagnosis evaluation model, and the releaseable multiplication convolution fault diagnosis evaluation model is updated and optimized; The method for updating the releaseable multiplication convolution fault diagnosis evaluation model is as follows: A cross-entropy loss function is taken as an optimization function of the releaseable multiplication convolution fault diagnosis evaluation model, the releaseable multiplication convolution fault diagnosis evaluation model is updated by minimizing the cross-entropy loss, and the update function is as follows: ， where X is the training set composed of spectrum samples, θ = {ω c , λ, w conv , b conv , w fc , b fc} is the model parameter of the releaseability multiplicative convolution fault diagnosis evaluation model, ω c and λ are the center frequency and bandwidth penalty coefficient of the multiplicative feature separator respectively, w conv and b conv are the weight and bias of the convolution feature extractor respectively, w fc and b fc are the weight matrix and bias vector of the discriminator respectively; and are the real fault label value and the predicted label value respectively, K is the number of health state categories; argmin is a mathematical function representing symbol; According to a chain rule, the gradient of the model parameters is calculated layer by layer and updated, the multiplication feature separator parameter update of the first layer is as follows: ， Wherein, ∂ is the partial derivative operator, η is the learning rate; L is the length of the vibration sample; For health status evaluation probability value; Is the inference result of the releaseable multiplication convolution fault diagnosis evaluation model; y represents the node response of the convolution layer output; H is the output mapping of the multiplication feature separator; And The partial derivative of the filter kernel analytical expression with respect to the frequency parameter and the bandwidth penalty coefficient parameter; ω ci And λ i The center frequency and bandwidth penalty coefficient of the i-th multiplication filter kernel, respectively, M represents the multiplication filter kernel; For the Wiener filter kernel and the Gaussian filter kernel, the partial derivatives are calculated as follows: ， ， ， ， where W(·) and G(·) denote the Wiener and Gaussian filter kernels, respectively, and are their bandwidth penalty coefficients, respectively; a i and are the bandwidth penalty coefficient of the i-th Wiener filter kernel and the bandwidth penalty coefficient of the i-th Gaussian filter kernel, respectively, and ω denotes the frequency. 2.The releaseability-multiplication convolution network-based fault diagnosis method of claim 1, wherein, The specific method for constructing the frequency spectrum sample data set is as follows: Collect vibration signals under different health states of the equipment, set a sampling frequency as Fs, and set the number of health states as K; Each vibration signal is divided into a plurality of vibration samples, the length of each vibration sample is L=Fs, and the sampling duration corresponding to the vibration sample is set; The FFT transform is performed on each vibration sample to obtain the corresponding frequency spectrum sample as and record the sample label. 3.The releaseability-multiplication convolution network based fault diagnosis method of claim 1, wherein, The method for designing the multiplication feature separator is as follows: The multiplication feature separator includes a plurality of analytical multiplication filter kernels for separating the fault-sensitive features in the spectral sample, each multiplication filter kernel having two learnable parameters: a center frequency ω c and a bandwidth penalty coefficient λ, the feature map h k corresponding to the kth multiplication filter kernel of the multiplication feature separator is: ， where X(ω) is a training set consisting of spectral samples, ω is frequency; M(·) is a multiplication filter kernel; ω ck and λ k are the center frequency and bandwidth penalty coefficient of the kth multiplication filter kernel, respectively. N multiplication filter kernels are stacked into an output mapping H of the multiplication feature separator, the depth of the multiplication feature separator is N, and the number of the multiplication filter kernels is N: 。 4. The fault diagnosis method based on the releaseability multiplication convolution network according to claim 3, wherein, To make each multiplication filter kernel separate more discriminative fault information and improve the diversity of feature maps, an anti-aliasing strategy is introduced in the multiplication feature separator. The kth feature map h k is rewritten as ， wherein h i represents the i-th feature map. 5.The releaseability-multiplication convolution network based fault diagnosis method of claim 3, wherein, In order to realize feature separation, the multiplication filter kernel design analytical expression can be as follows: ， and / or , where W(·) and G(·) denote the Wiener and Gaussian filter kernels, respectively, and a and b are their bandwidth penalty coefficients; ω are the center frequency parameters. c are the center frequency parameters.

6. The releaseability-multiplication convolution network-based fault diagnosis method of claim 1, wherein, The method for designing the convolution feature extractor is as follows: A one-dimensional convolution layer is constructed, which is used to extract high-dimensional features in the multiplication feature separator output mapping feature map H, the output channel of the convolution kernel is 64, and the convolution operation is represented as follows: ， where w conv and b conv are the weights and biases of the convolutional feature extractor, respectively; ReLU(·) is an activation function with the activation criterion: ， Wherein, y represents a node response of the convolution layer output; The feature y is reduced in dimension through an adaptive maximum pooling function AdptiveMaxPooling(·), and the maximum 10 outputs are retained, and the pooling process is as follows: ， where y l is the convolution output vector of the lth channel, denotes the ith output of the adaptive max-pooling of the lth channel of the convolution layer. y = f(x) (1) l Flattened feature vector y: ， wherein, is the flattened feature vector, Flatten denotes the flattening process.

7. The releaseability-multiplication convolution network-based fault diagnosis method of claim 1, wherein, The method for designing the discriminator is as follows: The feature vector in the convolutional feature extractor is reduced in dimensionality to the number of classes K using a single layer of fully connected layers dimensionality to the number of classes K using a single layer of fully connected layers ， where w fc and b fc are the weight matrix and bias vector of the fully connected layer, respectively; is the inference result of the releaseable multiplication convolution fault diagnosis evaluation model; The evaluation probability of each health state is as follows: ， wherein, is the i-th health state evaluation probability value, represents the output value of the j-th neuron at the output end of the full connection layer, and the prediction probability of the j-th fault; e represents a natural index; represents the output value of the i-th neuron at the output end of the full connection layer, and the prediction probability of the i-th fault.

8. A fault diagnosis system based on a releasability multiplication convolution network, characterized by, The system comprises a data acquisition module and a fault diagnosis module, the data acquisition module is used to collect vibration signals under different health states of the equipment, an output end of the data acquisition module is connected with an input end of the fault diagnosis module, the fault diagnosis module executes the method in any one of claims 1-7, and performs fault diagnosis.

9. A gearbox fault testing system characterised in that, The system comprises a secondary gear box, a magnetic powder brake, a load controller, a remote terminal, a motor, and the fault diagnosis system in claim 8; The magnetic powder brake is connected with the secondary gear box through a shaft coupling, and an output end of the load controller is connected with a control end of the magnetic powder brake; A first control signal output end of the remote terminal is connected with the load controller, a second control signal output end of the remote terminal is connected with a control end of the motor through an amplifier, and a rotating shaft of the motor is connected with the secondary gear box; The data acquisition module of the fault diagnosis system is connected with the secondary gear box, and is used to collect vibration signals of a fault gear in the secondary gear box.

Images