Fault diagnosis methods, systems, media, and equipment that integrate multi-scale feature enhancement and intra-class variance constraints

Abstract

This invention relates to the field of industrial equipment condition monitoring and intelligent fault diagnosis, and discloses a fault diagnosis method, system, medium, and device that integrates multi-scale feature enhancement and intra-class variance constraint. The method includes: acquiring vibration signals under different operating conditions; calculating time-domain statistical features to construct a multi-channel index input matrix and assigning labels to each matrix sample; inputting the multi-channel index input matrix into a deep diagnostic model containing an MS-CA module; constructing an attention constraint mechanism based on physical index stability; calculating the cross-condition stability of each physical index channel in real time based on the multi-channel index input matrix and operating condition labels; guiding the weight allocation of the MS-CA module by constructing a feature-attention consistency loss function; training the deep diagnostic model based on the loss function to optimize the parameters; constructing a multi-channel index input matrix from the real-time vibration signals of the equipment to be diagnosed, transmitting it to the trained deep diagnostic model, and outputting the final fault category diagnosis result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment state monitoring and intelligent fault diagnosis, and in particular to a fault diagnosis method, system, medium and equipment fusing multi-scale feature enhancement and intra-class variance constraint. BACKGROUND

[0002] With the development of industrial intelligence, using deep learning to diagnose faults of industrial equipment has become the mainstream. However, in the actual variable working condition environment, the existing technology still has significant defects:

[0003] First, there is a lack of adaptive filtering and fusion capability for multi-dimensional physical indicators. Traditional diagnosis methods often rely on manual selection of specific time domain indicators based on human experience, which is inefficient and subjective. While existing deep learning methods can accept multi-channel indicator input, they mostly use simple concatenation or fully connected layers for fusion, treating all physical indicators equally. In fact, different time domain indicators have completely different sensitivity to faults (for example, kurtosis is sensitive to shock, while RMS is sensitive to energy), and there is a lack of a multi-scale mechanism to finely weight and feature reorganization according to the characteristics of the input indicators, making it difficult to effectively highlight key fault physical features.

[0004] Second, the distribution of physical indicators drifts seriously under cross-condition scenarios, and the generalization ability is poor. Mechanical equipment often operates in non-stationary working conditions with varying speeds and loads. Many time domain indicators (such as effective value, peak value) are not only related to faults, but also strongly related to working condition parameters (such as speed). This leads to a huge difference in indicator values for the same health state under different working conditions (i.e., the intra-class variance is too large). Existing training mainly relies on cross-entropy loss, which only focuses on class distinction and ignores the constraint on the drift of such "working condition sensitive" indicators, making the model prone to overfitting. Once the working condition changes, the diagnosis accuracy will decrease significantly.

[0005] Third, the existing neural network attention mechanism lacks perception of "indicator stability". In order to filter features, some studies have introduced channel attention mechanisms. However, existing attention mechanisms only allocate weights based on "numerical intensity" or "activation level". Under variable working conditions, some indicators (such as energy indicators that are strongly affected by speed) have large values, but their volatility is very strong, which are unstable features. Existing methods cannot evaluate the stability of indicators under cross-condition, and are prone to incorrectly assigning high weights to these unstable indicators, while ignoring essential features that are less numerically small but robust to working condition changes. SUMMARY

[0006] To address the aforementioned problems, the purpose of this invention is to provide a fault diagnosis method, system, medium, and device that integrates multi-scale feature enhancement and intra-class variance constraints. This method can adaptively weight multi-channel physical indicators and automatically suppress the influence of unstable indicators for cross-condition fault diagnosis.

[0007] To achieve the above objectives, in a first aspect, the technical solution adopted by the present invention is as follows: a fault diagnosis method integrating multi-scale feature enhancement and intra-class variance constraint, comprising: acquiring vibration signals of rotating machinery under various working conditions to obtain multi-working-condition vibration data; calculating time-domain statistical features to construct a multi-channel index input matrix; configuring fault category labels and working-condition labels for each matrix sample; inputting the multi-channel index input matrix into a deep diagnostic model containing an MS-CA module to output the probability distribution of fault categories; MS-CA being multi-scale channel attention; constructing an attention constraint mechanism based on physical index stability; based on the multi-channel index input matrix and working-condition labels, calculating the cross-working-condition stability of each physical index channel in real time during the forward propagation process of deep diagnostic model training; and guiding the weight allocation of the MS-CA module to maintain consistency with index stability through a loss function; training the deep diagnostic model based on the loss function to optimize parameters; constructing a multi-channel index input matrix from the real-time vibration signals of the equipment to be diagnosed, transmitting it to the trained deep diagnostic model, and outputting the final fault category diagnosis result.

[0008] Furthermore, vibration signals of rotating machinery under various operating conditions are collected to obtain multi-condition vibration data. Time-domain statistical features are calculated to construct a multi-channel index input matrix, and fault category labels and operating condition labels are configured for each matrix sample, including: Vibration signals of rotating machinery under various operating conditions are collected using accelerometers. Overlap sampling is performed on the acquired original one-dimensional vibration signal to divide the continuous long signal into a fixed-length sample sequence, and the samples are divided into training set and test set; For each segmented sample time segment, calculate N different physical time domain indices; The calculated N time-domain index values ​​are standardized to eliminate dimensional differences. Then, the N indexes are used to construct the network's input feature matrix, and each matrix sample is assigned a fault category label and an operating condition label as a multi-channel index input matrix.

[0009] Furthermore, the multi-channel indicator input matrix is ​​fed into the constructed deep diagnostic model containing the MS-CA module to output the probability distribution of fault categories, including: The input layer dimension of the deep diagnostic model is designed as (B, N, T); where B is the batch size; N is the number of channels, corresponding to N physical time-domain indicators; and T is the time step. The MS-CA module is embedded in the shallow feature extraction stage of the network as a physical index filter to achieve adaptive filtering of different physical indices. The index feature matrix after MS-CA weighting is input into the subsequent deep feature extraction network, and then outputs the probability distribution of fault categories through fully connected layers and the Softmax activation function.

[0010] Furthermore, the MS-CA module is embedded in the shallow feature extraction stage of the network, specifically as follows: The MS-CA module contains three parallel 1D convolutional branches of different scales, used to extract the temporal evolution features of each physical index in parallel along the time dimension T. After fusing branch features at different scales, the features are compressed into a dimensionless form using global average pooling and a multilayer perceptron. The channel weight vector W, where each element of the channel weight vector W directly corresponds to the importance of a physical index; The generated weight vector W is multiplied element-wise with the input multi-channel index matrix to suppress index channels that are judged to have low weight and enhance high-weight index channels, thus completing the index screening.

[0011] Furthermore, an attention constraint mechanism based on the stability of physical indicators is constructed. Based on the multi-channel indicator input matrix and operating condition labels, the cross-operating condition stability of each physical indicator channel is calculated in real time during the forward propagation process of deep diagnostic model training. The loss function guides the weight allocation of the MS-CA module to maintain consistency with indicator stability, including: A group sampling strategy is used when constructing training batches to calculate the intra-class cross-condition variance; For the multi-channel index input matrix before weighting by the MS-CA module, calculate the operating condition sensitivity of each channel; group the samples in the batch according to the fault category k, and for the first... For each physical index channel, calculate its characteristic mean under different operating conditions, and calculate the variance under different operating conditions. ; Traverse all N channels to obtain the physical index instability vector V; Establish the constraint relationship between the weight vector W generated by MS-CA and the index instability vector V, and construct the feature-attention consistency loss term. This is to ensure that the weight allocation of the MS-CA module remains consistent with the stability of the metrics.

[0012] Furthermore, a feature-attention consistency loss term is constructed. for:

[0013] In the formula, Assigning the first to MS-CA The importance weight of each indicator.

[0014] Furthermore, a deep diagnostic model is trained to optimize the parameters. The specific implementation process is as follows: Construct a total loss function to guide network parameter updates. ;in, Loss to the main task For feature-attention consistency loss, This is the balance coefficient; The training set containing samples from different working conditions is input into the deep diagnostic model in batches. Forward propagation is used to calculate the classification results and the current index stability vector V and weight vector W; the total loss is then calculated. The gradient is calculated using the backpropagation algorithm; An optimizer is used to update the parameters in the network. As training progresses, due to... With the presence of [something], the network will automatically adjust the parameters of the MS-CA module, causing the weights W generated to gradually tilt towards indicators with higher stability.

[0015] Secondly, the technical solution adopted by this invention is as follows: a fault diagnosis system integrating multi-scale feature enhancement and intra-class variance constraint, comprising: a data acquisition and processing module, which acquires vibration signals of rotating machinery under various working conditions to obtain multi-working-condition vibration data, calculates time-domain statistical features to construct a multi-channel index input matrix, and configures fault category labels and working condition labels for each matrix sample; a deep diagnosis module, which inputs the multi-channel index input matrix into a deep diagnosis model containing an MS-CA module to output the probability distribution of fault categories; MS-CA is multi-scale channel attention; an IVM module, which constructs an attention constraint mechanism based on physical index stability, and calculates the cross-working-condition stability of each physical index channel in real time during the forward propagation process of deep diagnosis model training based on the multi-channel index input matrix and working condition labels, and guides the weight allocation of the MS-CA module to maintain consistency with index stability through a loss function; and a training and optimization module, which trains the deep diagnosis model based on the loss function to optimize the parameters, constructs a multi-channel index input matrix from the real-time vibration signal of the equipment to be diagnosed, transmits it to the trained deep diagnosis model, and outputs the final fault category diagnosis result.

[0016] Thirdly, the technical solution adopted by the present invention is: a computer-readable storage medium for storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform any of the methods described above.

[0017] Fourthly, the technical solution adopted by the present invention is as follows: a computing device, comprising: one or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described above.

[0018] The present invention has the following advantages due to the adoption of the above technical solutions: 1. This invention introduces an improved MS-CA module, breaking through the limitations of traditional methods that merely concatenate or manually select multidimensional indicators. Through a multi-branch parallel structure, this module can not only automatically assess the importance of different physical indicators but also capture the multi-scale temporal evolution patterns and nonlinear correlations in the indicator sequence. This enables the model to automatically extract the combination of indicators with the strongest fault characterization capability from a redundant set of time-domain indicators.

[0019] 2. This invention, based on the feature-attention consistency constraint mechanism of IVM, addresses the problem that some time-domain indicators (such as RMS and peak values) drift drastically with changes in speed and load. This invention explicitly penalizes the weights of channels with high numerical intra-class variance indicators, thus guiding the model to "ignore" unstable indicators sensitive to operating conditions and instead rely on essential features robust to changes in operating conditions. This allows the model to maintain high diagnostic accuracy even under unseen operating conditions.

[0020] 3. This invention possesses strong physical interpretability. Unlike the "black box" feature extraction of traditional deep learning models, this invention directly uses explicit physical time-domain indicators as multi-channel inputs. By analyzing the weight distribution ultimately learned by the MS-CA module, technicians can directly determine which physical indicators (such as kurtosis vs. effective values) are most robust and valuable for cross-condition diagnosis in the current diagnostic task, thereby guiding subsequent sensor deployment or edge computing feature selection. Attached Figure Description

[0021] Figure 1 This is a flowchart of a fault diagnosis method that integrates multi-scale feature enhancement and intra-class variance constraint in an embodiment of the present invention. Detailed Implementation

[0022] To address the shortcomings of existing technologies, this invention provides a fault diagnosis method, system, medium, and device that integrates multi-scale feature enhancement and intra-class variance constraints. First, the original vibration signal is processed to construct a feature matrix containing multi-channel time-domain indices, where each channel corresponds to a physical statistical index (such as mean, kurtosis, root mean square, etc.). Based on this, a deep neural network containing a "multi-scale channel attention module (MS-CA)" is constructed. The MS-CA module is used to extract the temporal dependence and correlation features of each physical index channel at different scales in parallel and adaptively generate importance weights for each index channel. Furthermore, this invention introduces an intra-class cross-condition feature variance minimization (IVM) strategy and an attention mechanism to establish a linked constraint during the training phase. Specifically, this invention constructs a feature-attention consistency loss function, using the "physical index variance (i.e., index stability)" of the same fault category under different operating conditions to regularize the channel weights generated by MS-CA. This mechanism forces the network to automatically identify and suppress physical index channels that fluctuate drastically with operating conditions (high variance) during training, thereby enabling the model to automatically focus on robust index combinations with operating condition invariance, achieving high-precision diagnosis across operating conditions.

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] In one embodiment of the present invention, a fault diagnosis method integrating multi-scale feature enhancement and intra-class variance constraint is provided. The method utilizes an intra-class cross-condition feature variance minimization strategy to regularize the attention weights of multi-scale channels, thereby achieving high-precision and robust fault feature extraction and identification diagnosis under drastically changing operating conditions. This enables the training and application of a neural network model for this purpose. In this embodiment, as shown... Figure 1 As shown, the method includes the following steps: 1) Multi-condition vibration data acquisition and dataset construction: acquire vibration signals of rotating machinery under various working conditions to obtain multi-condition vibration data, calculate time-domain statistical features to construct a multi-channel index input matrix, and configure fault category labels and working condition labels for each matrix sample; 2) Construct a deep diagnostic model containing the MS-CA (Multi-Scale Channel Attention) module: Input the multi-channel index input matrix into the constructed deep diagnostic model containing the MS-CA module to output the probability distribution of fault categories; 3) Construct an attention constraint mechanism based on the stability of physical indicators: Based on the multi-channel indicator input matrix and working condition labels, the cross-working condition stability of each physical indicator channel is calculated in real time during the forward propagation process of deep diagnostic model training, and the weight allocation of the MS-CA module is guided by the loss function to be consistent with the indicator stability. 4) Model training and parameter optimization: The deep diagnostic model is trained based on the loss function to optimize the parameters. The real-time vibration signal of the device to be diagnosed is used to construct a multi-channel index input matrix, which is then transmitted to the trained deep diagnostic model to output the final fault category diagnosis result.

[0026] In step 1) above, vibration signals of rotating machinery under various operating conditions are collected to obtain multi-condition vibration data. Time-domain statistical features are calculated to construct a multi-channel index input matrix, and fault category labels and operating condition labels are configured for each matrix sample. This includes the following steps: 1.1) Vibration signals of rotating machinery under various working conditions are collected by an accelerometer.

[0027] In this embodiment, the operating conditions include different speeds (e.g., 1700 rpm, 1750 rpm, 1800 rpm) and different loads (e.g., 0 kW, 1 kW, 2 kW) to simulate the non-stable operating environment of actual industrial sites. The collected state categories include: normal state, inner ring fault, outer ring fault, rolling element fault, and k other states.

[0028] 1.2) Overlap sampling is performed on the acquired original one-dimensional vibration signal to divide the continuous long signal into a fixed-length sample sequence, and the samples are divided into training set and test set.

[0029] Specifically, the sample length is set to L sampling points with an overlap rate of 50%, and the continuous long signal is divided into a fixed-length sample sequence. The samples are then divided into a training set and a test set. It is ensured that the training set contains sample data for all operating conditions so that the model can learn the feature distribution across operating conditions.

[0030] 1.3) For each segmented sample time segment, calculate N different physical time domain indices.

[0031] In this embodiment, in order to comprehensively characterize the amplitude changes and waveform distortion of the signal, the following two types of indicators are selected: (1) Dimensional indicators: including mean, root mean square (RMS), root mean square (RMS), standard deviation (Std), peak value, and peak-to-peak value. These indicators are sensitive to changes in the energy and intensity of the signal.

[0032] (2) Dimensionless indices: These include kurtosis, skewness, shape factor, impulse factor, margin factor, and crest factor. These indices are sensitive to the distribution pattern of the signal and have a certain robustness to changes in the absolute amplitude of the signal.

[0033] 1.4) The calculated N time-domain index values ​​are standardized (Z-score standardization) to eliminate dimensional differences. Then, the N indices are used to construct the input feature matrix of the network, and each matrix sample is assigned a fault category label and an operating condition label as a multi-channel index input matrix.

[0034] In this embodiment, the fault category label is used to supervise the network to train fault classification; the operating condition label is used by the subsequent IVM module to identify the source of the samples and calculate the variance of the index of the same category under different operating conditions.

[0035] In step 2) above, the multi-channel indicator input matrix is ​​input into the constructed deep diagnostic model containing the MS-CA module to output the probability distribution of fault categories, including the following steps: 2.1) The input layer dimension of the deep diagnostic model is designed as (B, N, T); where B is the batch size; N is the number of channels, corresponding to N physical time-domain indicators (e.g., N=12, which represents 12 physical channels such as mean, kurtosis, and root mean square); and T is the time step, representing the evolution sequence of the indicators over time.

[0036] 2.2) The MS-CA module is embedded in the shallow feature extraction stage of the network as a physical index filter to achieve adaptive filtering of different physical indices.

[0037] In this embodiment, the MS-CA module is embedded in the shallow feature extraction stage of the network, specifically including the following steps: 2.2.1) Multi-scale perception: The MS-CA module contains three parallel 1D convolutional branches of different scales, which are used to extract the temporal evolution features of each physical index in parallel along the time dimension T; among them, the three parallel 1D convolutional branches of different scales have convolutional kernel sizes of 3×1, 5×1 and 7×1, respectively.

[0038] 2.2.2) Physical Meaning Fusion: After fusing branch features from different scales, the features are compressed into a dimension of 125-dimensional features using Global Average Pooling (GAP) and Multilayer Perceptron (MLP). The channel weight vector W, where each element of the channel weight vector W... This directly corresponds to the importance of a physical indicator; among them, , These represent the importance weight coefficients assigned by the MS-CA module to the physical index channels such as root mean square (RMS), kurtosis, and skewness.

[0039] 2.2.3) Feature recalibration: The generated weight vector W is multiplied element-wise with the input multi-channel index input matrix to suppress index channels that are judged as low weight and enhance high weight index channels, thus completing the index screening.

[0040] 2.3) The index feature matrix after MS-CA weighting is input into the subsequent deep feature extraction network, and then the probability distribution of the fault category is output through the fully connected layer and the softmax activation function.

[0041] In step 3) above, an attention constraint mechanism based on the stability of physical indicators is constructed. Based on the multi-channel indicator input matrix and operating condition labels, the cross-operating condition stability of each physical indicator channel is calculated in real time during the forward propagation process of deep diagnostic model training. The weight allocation of the MS-CA module is guided by the loss function to maintain consistency with the indicator stability. This includes the following steps: 3.1) When constructing training batches, a group sampling strategy is adopted to calculate the intra-class cross-condition variance; ensure that each batch contains sample data of the same fault category (e.g., all are "inner ring faults") but from different operating conditions (e.g., speeds of 1700 rpm and 1800 rpm).

[0042] 3.2) For the multi-channel index input matrix (i.e., N physical index channels) before weighting by the MS-CA module, calculate the operating condition sensitivity of each channel; group the samples in the batch according to the fault category k, for the... For each physical index channel (e.g., i=1 corresponds to the RMS index), calculate its eigenvalue (Centroid) under different operating conditions, and calculate the variance under different operating conditions. ; The larger the value, the more drastic the drift of the i-th physical index (such as RMS) with changes in operating conditions (speed / load), indicating poor stability of that index. By traversing all N channels, the physical index instability vector V is obtained; where, , These represent the characteristic variances (i.e. instability measures) of the root mean square (RMS), kurtosis, and skewness index channels across different operating conditions under the same fault category.

[0043] 3.3) Establish the constraint relationship between the weight vector W generated by MS-CA and the index instability vector V, and construct the feature-attention consistency loss term. This is to ensure that the weight allocation of the MS-CA module remains consistent with the stability of the metrics.

[0044] In this embodiment, a feature-attention consistency loss term is constructed. for:

[0045] In the formula, Assigning the first to MS-CA The importance weight of each indicator; It is the first The variance of the cross-operating condition fluctuation of each indicator.

[0046] Formula Mechanism Explanation: If a certain indicator (such as "Root Mean Square Value RMS") is greatly affected by rotational speed, its variance... It will be very large. At this point, if MS-CA still assigns it a high weight ( ),but This will lead to A surge. To reduce the loss, the backpropagation algorithm guides MS-CA to reduce... This means that unstable RMS indicators are automatically "shielded". Conversely, if some indicators are not sensitive to operating conditions, the model can confidently assign them high weights.

[0047] In step 4) above, the deep diagnostic model is trained to optimize the parameters. The specific implementation process includes the following steps: 4.1) Construct the total loss function to guide network parameter updates. ;in, Loss to the main task For feature-attention consistency loss, This is the balance coefficient.

[0048] In this embodiment, Cross-entropy loss is used to measure the accuracy of the model in classifying fault categories (such as inner ring faults and outer ring faults). Used to penalize models for their dependence on physical properties with high variance (i.e., poor stability across operating conditions). This is a hyperparameter used to adjust the strength of the constraint.

[0049] 4.2) Input the training set containing samples from different working conditions into the deep diagnostic model in batches, and perform forward propagation to calculate the classification result and the current index stability vector V and weight vector W; calculate the total loss. The gradient is calculated using the backpropagation algorithm.

[0050] 4.3) The optimizer (Adam) is used to update the parameters in the network. As training progresses, due to... With the presence of [something], the network will automatically adjust the parameters of the MS-CA module, causing the weights W generated to gradually tilt towards indicators with higher stability.

[0051] In step 4) above, the cross-condition online diagnosis specifically involves: acquiring real-time vibration signals of the equipment to be diagnosed. Following the method described above, N time-domain indicators (including root mean square value, kurtosis, etc.) of the signal are calculated in real time to construct a multi-channel indicator input matrix. This matrix is ​​then input into the trained network model. The trained MS-CA module automatically performs weighted filtering on the N input physical indicators based on its learned "stability prior knowledge," suppressing indicators highly susceptible to operating condition interference and retaining key indicators. The model outputs the final fault category diagnosis result.

[0052] In one embodiment of the present invention, a fault diagnosis system integrating multi-scale feature enhancement and intra-class variance constraint is provided, comprising: The data acquisition and processing module collects vibration signals of rotating machinery under various working conditions to obtain multi-condition vibration data, calculates time-domain statistical features to construct a multi-channel index input matrix, and configures fault category labels and working condition labels for each matrix sample. The deep diagnostic module takes the multi-channel indicator input matrix as input to the deep diagnostic model containing the MS-CA (Multi-Scale Channel Attention) module, and outputs the probability distribution of the fault category. The IVM module constructs an attention constraint mechanism based on the stability of physical indicators. Based on the multi-channel indicator input matrix and working condition labels, it calculates the cross-working condition stability of each physical indicator channel in real time during the forward propagation process of deep diagnostic model training. The loss function guides the weight allocation of the MS-CA module to be consistent with the indicator stability. The training and optimization module trains the deep diagnostic model based on the loss function to optimize the parameters. It constructs a multi-channel index input matrix from the real-time vibration signal of the device to be diagnosed and transmits it to the trained deep diagnostic model to output the final fault category diagnosis result.

[0053] In the above embodiments, vibration signals of rotating machinery under various operating conditions are collected to obtain multi-condition vibration data. Time-domain statistical features are calculated to construct a multi-channel index input matrix, and fault category labels and operating condition labels are configured for each matrix sample, including: Vibration signals of rotating machinery under various operating conditions are collected using accelerometers. Overlap sampling is performed on the acquired original one-dimensional vibration signal to divide the continuous long signal into a fixed-length sample sequence, and the samples are divided into training set and test set; For each segmented sample time segment, calculate N different physical time domain indices; The calculated N time-domain index values ​​are standardized to eliminate dimensional differences, and then the N indices are used to construct the network's input feature channels as a multi-channel index input matrix. Each constructed multi-channel indicator input matrix is ​​labeled with two types of tags: fault category tag and operating condition tag.

[0054] In the above embodiments, the multi-channel indicator input matrix is ​​input into the constructed deep diagnostic model containing the MS-CA module to output the probability distribution of fault categories, including: The input layer dimension of the deep diagnostic model is designed as (B, N, T); where B is the batch size; N is the number of channels, corresponding to N physical time-domain indicators; and T is the time step. The MS-CA module is embedded in the shallow feature extraction stage of the network as a physical index filter to achieve adaptive filtering of different physical indices. The index feature matrix after MS-CA weighting is input into the subsequent deep feature extraction network, and then outputs the probability distribution of fault categories through fully connected layers and the Softmax activation function.

[0055] In the above embodiments, the MS-CA module is embedded in the shallow feature extraction stage of the network, specifically as follows: The MS-CA module contains three parallel 1D convolutional branches of different scales, used to extract the temporal evolution features of each physical index in parallel along the time dimension T. After fusing branch features at different scales, the features are compressed into a dimensionless form using global average pooling and a multilayer perceptron. The channel weight vector W, where each element of the channel weight vector W directly corresponds to the importance of a physical index; The generated weight vector W is element-wise multiplied with the input multi-channel index matrix to suppress index channels that are judged to have low weight and enhance high-weight index channels, thus completing the index screening.

[0056] In the above embodiments, the multi-channel indicator input matrix is ​​input to the constructed attention constraint mechanism based on physical indicator stability, so as to calculate the cross-condition stability of each physical indicator channel in real time during the forward propagation process of deep diagnostic model training, and guide the weight allocation of the MS-CA module to be consistent with the indicator stability through the loss function, including: A group sampling strategy is used when constructing training batches to calculate the intra-class cross-condition variance; For the multi-channel index input matrix before weighting by the MS-CA module, calculate the operating condition sensitivity of each channel; group the samples in the batch according to the fault category k, and for the first... For each physical index channel, calculate its characteristic mean under different operating conditions, and calculate the variance under different operating conditions. ; Traverse all N channels to obtain the physical index instability vector V; Establish the constraint relationship between the weight vector W generated by MS-CA and the index instability vector V, and construct the feature-attention consistency loss term. This is to ensure that the weight allocation of the MS-CA module remains consistent with the stability of the metrics.

[0057] In the above embodiments, a feature-attention consistency loss term is constructed. for:

[0058] In the formula, Assigning the first to MS-CA The importance weight of each indicator; It is the first The variance of the cross-operating condition fluctuation of each indicator.

[0059] In the above embodiments, the deep diagnostic model is trained to optimize the parameters. The specific implementation process is as follows: Construct a total loss function to guide network parameter updates. ;in, Loss to the main task For feature-attention consistency loss, This is the balance coefficient; The training set containing samples from different working conditions is input into the deep diagnostic model in batches. Forward propagation is used to calculate the classification results and the current index stability vector V and weight vector W; the total loss is then calculated. The gradient is calculated using the backpropagation algorithm; An optimizer is used to update the parameters in the network. As training progresses, due to... With the presence of [something], the network will automatically adjust the parameters of the MS-CA module, causing the weights W generated to gradually tilt towards indicators with higher stability.

[0060] The system provided in this embodiment is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0061] In one embodiment of the present invention, a computing device is provided. This computing device can be a terminal and may include a processor, a communication interface, memory, a display screen, and an input device. The processor, communication interface, and memory communicate with each other via a communication bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. When the computer programs are executed by the processor, they implement the methods described in the above embodiments. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, a management network, NFC (Near Field Communication), or other technologies. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad mounted on the casing of the computing device, or an external keyboard, touchpad, or mouse. The processor can call logical instructions stored in the memory.

[0062] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0063] In one embodiment of the present invention, a computer program product is provided, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, and when the program instructions are executed by a computer, the computer is able to perform the methods provided in the above-described method embodiments.

[0064] In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided, which stores server instructions that cause a computer to perform the methods provided in the above embodiments.

[0065] The computer-readable storage medium provided in the above embodiments has a similar implementation principle and technical effect to the above method embodiments, and will not be described again here.

[0066] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault diagnosis method integrating multi-scale feature enhancement and intra-class variance constraint, characterized in that, include: Vibration signals of rotating machinery under various working conditions are collected to obtain multi-condition vibration data. Time-domain statistical features are calculated to construct a multi-channel index input matrix, and fault category labels and working condition labels are configured for each matrix sample. The multi-channel indicator input matrix is ​​input into a deep diagnostic model containing the MS-CA module to output the probability distribution of fault categories. MS-CA stands for Multi-Scale Channel Attention; An attention constraint mechanism based on the stability of physical indicators is constructed. Based on the multi-channel indicator input matrix and working condition labels, the cross-working condition stability of each physical indicator channel is calculated in real time during the forward propagation process of deep diagnostic model training. The weight allocation of the MS-CA module is guided by the loss function to be consistent with the indicator stability. The deep diagnostic model is trained based on the loss function to optimize the parameters. The real-time vibration signal of the device to be diagnosed is used to construct a multi-channel index input matrix, which is then transmitted to the trained deep diagnostic model to output the final fault category diagnosis result.

2. The fault diagnosis method integrating multi-scale feature enhancement and intra-class variance constraint as described in claim 1, characterized in that, Vibration signals of rotating machinery under various operating conditions are collected to obtain multi-condition vibration data. Time-domain statistical features are calculated to construct a multi-channel index input matrix, and fault category labels and operating condition labels are assigned to each matrix sample, including: Vibration signals of rotating machinery under various operating conditions are collected using accelerometers. Overlap sampling is performed on the acquired original one-dimensional vibration signal to divide the continuous long signal into a fixed-length sample sequence, and the samples are divided into training set and test set; For each segmented sample time segment, calculate N different physical time domain indices; The calculated N time-domain index values ​​are standardized to eliminate dimensional differences. Then, the N indexes are used to construct the network's input feature matrix, and each matrix sample is assigned a fault category label and an operating condition label as a multi-channel index input matrix.

3. The fault diagnosis method integrating multi-scale feature enhancement and intra-class variance constraint as described in claim 1, characterized in that, The multi-channel indicator input matrix is ​​fed into a constructed deep diagnostic model containing an MS-CA module to output the probability distribution of fault categories, including: The input layer dimension of the deep diagnostic model is designed as (B, N, T); where B is the batch size; N is the number of channels, corresponding to N physical time-domain indicators; and T is the time step. The MS-CA module is embedded in the shallow feature extraction stage of the network as a physical index filter to achieve adaptive filtering of different physical indices. The index feature matrix after MS-CA weighting is input into the subsequent deep feature extraction network, and then outputs the probability distribution of fault categories through fully connected layers and the Softmax activation function.

4. The fault diagnosis method integrating multi-scale feature enhancement and intra-class variance constraint as described in claim 3, characterized in that, The MS-CA module is embedded in the shallow feature extraction stage of the network, specifically as follows: The MS-CA module contains three parallel 1D convolutional branches of different scales, used to extract the temporal evolution features of each physical index in parallel along the time dimension T. After fusing branch features at different scales, the features are compressed into a dimensionless form using global average pooling and a multilayer perceptron. The channel weight vector W, where each element of the channel weight vector W directly corresponds to the importance of a physical index; The generated weight vector W is multiplied element-wise with the input multi-channel index matrix to suppress index channels that are judged to have low weight and enhance high-weight index channels, thus completing the index screening.

5. The fault diagnosis method integrating multi-scale feature enhancement and intra-class variance constraint as described in claim 1, characterized in that, An attention constraint mechanism based on the stability of physical indicators is constructed. Based on a multi-channel indicator input matrix and operating condition labels, the cross-operating condition stability of each physical indicator channel is calculated in real time during the forward propagation of the deep diagnostic model training. The loss function guides the weight allocation of the MS-CA module to maintain consistency with indicator stability, including: A group sampling strategy is used when constructing training batches to calculate the intra-class cross-condition variance; For the multi-channel index input matrix before weighting by the MS-CA module, calculate the operating condition sensitivity of each channel; group the samples in the batch according to the fault category k, and for the first... For each physical index channel, calculate its characteristic mean under different operating conditions, and calculate the variance under different operating conditions. ; Traverse all N channels to obtain the physical index instability vector V; Establish the constraint relationship between the weight vector W generated by MS-CA and the index instability vector V, and construct the feature-attention consistency loss term. This is to ensure that the weight allocation of the MS-CA module remains consistent with the stability of the metrics.

6. The fault diagnosis method integrating multi-scale feature enhancement and intra-class variance constraint as described in claim 5, characterized in that, Constructing a feature-attention consistency loss term for: In the formula, Assigning the first to MS-CA The importance weight of each indicator.

7. The fault diagnosis method integrating multi-scale feature enhancement and intra-class variance constraint as described in claim 1, characterized in that, The deep diagnostic model is trained to optimize the parameters. The specific implementation process is as follows: Construct a total loss function to guide network parameter updates. ;in, Loss to the main task For feature-attention consistency loss, This is the balance coefficient; The training set containing samples from different working conditions is input into the deep diagnostic model in batches. Forward propagation is used to calculate the classification results and the current index stability vector V and weight vector W; the total loss is then calculated. The gradient is calculated using the backpropagation algorithm; An optimizer is used to update the parameters in the network. As training progresses, due to... With the presence of [something], the network will automatically adjust the parameters of the MS-CA module, causing the weights W generated to gradually tilt towards indicators with higher stability.

8. A fault diagnosis system integrating multi-scale feature enhancement and intra-class variance constraint, characterized in that, include: The data acquisition and processing module collects vibration signals of rotating machinery under various working conditions to obtain multi-condition vibration data, calculates time-domain statistical features to construct a multi-channel index input matrix, and configures fault category labels and working condition labels for each matrix sample. The deep diagnostic module takes the multi-channel indicator input matrix as input to the deep diagnostic model containing the MS-CA module, and outputs the probability distribution of fault categories. MS-CA stands for Multi-Scale Channel Attention; The IVM module constructs an attention constraint mechanism based on the stability of physical indicators. Based on the multi-channel indicator input matrix and working condition labels, it calculates the cross-working condition stability of each physical indicator channel in real time during the forward propagation process of deep diagnostic model training. The loss function guides the weight allocation of the MS-CA module to be consistent with the indicator stability. The training and optimization module trains the deep diagnostic model based on the loss function to optimize the parameters. It constructs a multi-channel index input matrix from the real-time vibration signal of the device to be diagnosed and transmits it to the trained deep diagnostic model to output the final fault category diagnosis result.

9. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described in claims 1 to 7.

10. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods described in claims 1 to 7.

Images