A method and system for fault diagnosis of electromechanical equipment

Abstract

This invention discloses a fault diagnosis method and system for electromechanical equipment, relating to the field of computer technology. The fault diagnosis model of this application employs a residual network to extract fault-related features of the electromechanical equipment, then uses a weighted bidirectional feature pyramid network to fuse multi-scale features to enhance their discriminative power. A feature representation module with a cyclic stacked structure based on a scaling dot product attention mechanism is used for feature mining. Even with insufficient fault samples, the model learns temporal information from multi-scale features, thereby improving the accuracy and efficiency of fault diagnosis for electromechanical equipment.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and system for diagnosing faults in electromechanical equipment. Background Technology

[0002] In the context of today's industrial big data, the rapid development of artificial intelligence and deep learning has led to the increasing intelligence of fault diagnosis. Neural network-based fault diagnosis algorithms are gaining more and more attention and have become a new research hotspot in the field. Deep learning technologies such as convolutional neural networks, recurrent neural networks, and adversarial neural networks can automatically mine deep features from input information. By directly inputting raw information at the input end, fault diagnosis results can be obtained at the output end. These methods are highly valued in the field of fault diagnosis today.

[0003] Currently, most deep learning-based fault diagnosis algorithms extract features using multi-layer convolutional neural networks, fully connected attention systems, or recurrent neural networks employing long short-term memory. However, this approach suffers from several problems: First, feature fusion is incomplete. Traditional neural network models either lack feature fusion or only perform top-down feature fusion, relying on a one-way information flow and ultimately just a simple concatenation operation. This easily leads to sparsity in the fused features, resulting in information loss and reduced fault diagnosis accuracy. Second, data scarcity exists. Traditional deep learning algorithms require massive amounts of data for training due to their large parameter count. However, faults in electromechanical equipment are infrequent, making fault data extremely scarce and hindering model training, thus affecting the effectiveness of fault diagnosis. Summary of the Invention

[0004] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention proposes a fault diagnosis method and system for electromechanical equipment, which can improve the accuracy of fault diagnosis for electromechanical equipment.

[0005] On one hand, embodiments of the present invention provide a fault diagnosis method for electromechanical equipment, comprising the following steps:

[0006] Acquire electromechanical equipment data;

[0007] The electromechanical equipment data is input into the fault diagnosis model to obtain the diagnostic results of the electromechanical equipment;

[0008] The fault diagnosis model includes a residual network, a weighted bidirectional feature pyramid network, and a feature representation module. The feature representation module adopts a cyclic stacked structure based on a scaled dot product attention mechanism. The construction process of the fault diagnosis model includes:

[0009] Obtain the training dataset;

[0010] The training dataset is input into the residual network for feature extraction to obtain multi-layer first feature data;

[0011] Based on a weighted bidirectional feature pyramid network, feature fusion is performed on the first feature data of multiple layers to obtain the second feature data of multiple layers.

[0012] Based on the feature representation module, feature mining is performed on the multi-layer second feature data to obtain the prediction results of the training dataset;

[0013] Based on the prediction results, backpropagation is performed to update the model parameters, resulting in a trained fault diagnosis model.

[0014] According to some embodiments of the present invention, the process of constructing the fault diagnosis model further includes:

[0015] Acquire historical data of electromechanical equipment, wherein the historical data includes normal data and abnormal data of different fault types;

[0016] Perform a short-time windowed Fourier transform on the historical data to obtain data samples;

[0017] The data samples are divided into a training dataset, a validation dataset, and a test dataset. The training dataset is used to train the fault diagnosis model, the test dataset is used to detect the generalization ability of the fault diagnosis model after training, and the validation dataset is used to detect whether the fault diagnosis model after training is overfitting.

[0018] According to some embodiments of the present invention, performing a short-time windowed Fourier transform on the historical data to obtain data samples includes the following steps:

[0019] The historical data over the entire time segment is divided into multiple short-time segments of historical data;

[0020] Fourier transform is performed on historical data from multiple short time segments to obtain data samples.

[0021] According to some embodiments of the present invention, the formula for the short-time windowed Fourier transform is:

[0022]

[0023] Where g(ut) is the window function, f(u) is the input historical data, and f(w,t) is a two-dimensional signal containing both frequency and time dimensions.

[0024] According to some embodiments of the present invention, the residual network adopts ResNet, which is represented as:

[0025] F(x) = H(x) + x;

[0026] Where H(x) is the feature extracted by the current layer, x is the feature extracted by the previous layer, i.e., the input feature of the current layer, and F(x) is the output feature of the current layer.

[0027] According to some embodiments of the present invention, the weighted bidirectional feature pyramid network adopts BiFPN, and BiFPN is repeatedly used to perform feature fusion of the first feature data of multiple layers from top to bottom and from bottom to top.

[0028] According to some embodiments of the present invention, BiFPN is represented as:

[0029]

[0030]

[0031] in, It is the input feature of the x-th layer, that is, the first feature data F output by the x-th layer of the residual network. x (x), It is an intermediate feature of the xth layer on the top-down path, and It is the output feature of the xth layer on the bottom-up path.

[0032] According to some embodiments of the present invention, the step of performing feature mining on the multi-layer second feature data based on the feature representation module to obtain the prediction result of the training dataset includes the following steps:

[0033] The second feature data of each layer is convolved and labeled to obtain the label embedding matrix;

[0034] The tag embedding matrix is ​​input into the Transformer encoder to obtain the encoder output matrix;

[0035] The importance weight value of each tag embedding matrix is ​​determined based on the encoder output matrix;

[0036] The importance weights of each label embedding matrix are passed through a fully connected layer and then output by the softmax function to obtain the prediction results.

[0037] According to some embodiments of the present invention, inputting the tag embedding matrix into the Transformer encoder to obtain the encoder output matrix includes the following steps:

[0038] A linear transformation is performed on the input tag embedding matrix to obtain the query matrix, key matrix, and value matrix;

[0039] Based on the multi-head attention mechanism, the encoder output matrix is ​​determined according to the query matrix, the key matrix, and the value matrix.

[0040] On the other hand, embodiments of the present invention also provide a fault diagnosis system for electromechanical equipment, comprising:

[0041] The first module is used to acquire data from electromechanical equipment;

[0042] The second module is used to input the electromechanical equipment data into the fault diagnosis model to obtain the diagnostic results of the electromechanical equipment.

[0043] The fault diagnosis model includes a residual network, a weighted bidirectional feature pyramid network, and a feature representation module. The feature representation module adopts a cyclic stacked structure based on a scaled dot product attention mechanism. The construction process of the fault diagnosis model includes:

[0044] Obtain the training dataset;

[0045] The training dataset is input into the residual network for feature extraction to obtain multi-layer first feature data;

[0046] Based on a weighted bidirectional feature pyramid network, feature fusion is performed on the first feature data of multiple layers to obtain the second feature data of multiple layers.

[0047] Based on the feature representation module, feature mining is performed on the multi-layer second feature data to obtain the prediction results of the training dataset;

[0048] Based on the prediction results, backpropagation is performed to update the model parameters, resulting in a trained fault diagnosis model.

[0049] The above-mentioned technical solution of the present invention has at least one of the following advantages or beneficial effects: The fault diagnosis model of this application uses a residual network to extract fault-related features of electromechanical equipment, and then uses a weighted bidirectional feature pyramid network to fuse multi-scale features to make the features have higher discriminative ability. It uses a feature representation module based on a cyclic stacked structure of scaling dot product attention mechanism to perform feature mining. In the case of insufficient fault samples, it learns the temporal information in multi-scale features, thereby improving the accuracy and efficiency of fault diagnosis of electromechanical equipment. Attached Figure Description

[0050] Figure 1 This is a flowchart of the fault diagnosis model construction process in the fault diagnosis method for electromechanical equipment provided in this embodiment of the invention;

[0051] Figure 2 This is a schematic diagram of the fault diagnosis model processing procedure provided in the embodiment of the present invention. Detailed Implementation

[0052] 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 originals 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.

[0053] In the description of this invention, the use of terms such as "first," "second," etc., is merely for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of technical features indicated, or implicitly indicating the order of the technical features indicated.

[0054] In related technologies, Liu et al. proposed a fault diagnosis method based on WOA-GRNN to address the problem of low accuracy in fault diagnosis of hydraulic pumps in complex systems in practical applications. This method employs a generalized regression neural network (GRNN), fully utilizing the trend characteristics of corresponding hydraulic pump signal monitoring to effectively characterize each signal sample, fully explore the correlation between signals, and optimize the GRNN parameters using the Whale Optimization Algorithm (WOA), thereby efficiently and accurately predicting the fault category of the hydraulic pump. The main workflow of this method is as follows: Analyze, classify, and preprocess the monitoring parameter data of a hydraulic pump to construct the hydraulic pump monitoring parameter data; design a generalized regression neural network (GRNN) fault diagnosis model using hydraulic pump health data and fault sample datasets; on this basis, optimize the GRNN parameters using the Whale Optimization Algorithm (WOA) to establish the optimal GRNN fault diagnosis model WOA-GRNN; finally, test the created WOA-GRNN fault diagnosis model using a test sample dataset. This method has advantages such as good local approximation, global optimality and fast computation speed, but it still has shortcomings. Because low-level features have higher resolution and contain detailed information, but due to fewer convolutions, their semantics are lower and there is more noise. High-level features have stronger semantic information, but their resolution is very low and their ability to perceive details is poor, resulting in information loss.

[0055] To address the issue of insufficient ability of convolutional neural networks to learn key fault features, thus affecting the accuracy of bearing fault diagnosis, Kang Tao et al. proposed an adaptive noise-resistant convolutional neural network model (MACNN) that integrates a multi-attention mechanism. This model utilizes a composite attention mechanism of channels and time to optimize the learning mechanism. The resulting multi-attention mechanism module suppresses the influence of noise and irrelevant signal components from different angles and adaptively enhances the response of fault features. The main workflow of this method is as follows: Analyze, classify, and preprocess rolling bearing vibration signals to construct a rolling bearing monitoring parameter dataset; train a multi-scale feature extractor layer by layer, calculate attention weight vectors c' and t, adjust the parameters of the feature extraction sub-network, use a channel attention mechanism to associate features between different scales, and submit the fused features to a classifier for rolling bearing fault diagnosis. This method utilizes a composite attention mechanism of channels and time to suppress the influence of noise and irrelevant signal components from different angles and adaptively enhances the response of fault features, exhibiting good noise resistance. However, it has the following drawback: the increased computational load due to the introduction of the multi-attention mechanism requires more time for model training. Furthermore, the effectiveness of this method depends on modeling large deep learning networks with a large amount of sensor data; when the amount of data is small, the model performs poorly.

[0056] Based on this, embodiments of the present invention provide a fault diagnosis method for electromechanical equipment, including but not limited to steps S110 and S120.

[0057] Step S110: Obtain electromechanical equipment data;

[0058] Step S120: Input the electromechanical equipment data into the fault diagnosis model to obtain the diagnostic results of the electromechanical equipment.

[0059] In some embodiments, the electromechanical equipment data may be the operating parameters of the electromechanical equipment, such as the voltage, current, temperature, and other data of the electromechanical equipment.

[0060] In some embodiments, the fault diagnosis model includes a residual network, a weighted bidirectional feature pyramid network, and a feature representation module. The feature representation module adopts a recurrent stacked structure based on a scaled dot product attention mechanism. The residual network can be ResNet, and the weighted bidirectional feature pyramid network can be BiFPN. The feature representation module is a lightweight Transformers module, which includes a convolutional labeling part and a sequence pooling strategy part. The convolutional labeling part includes convolution, pooling, and resizing. The sequence pooling strategy part includes a Transformers encoder, sequence pooling, and linear layers. Figure 2The fault diagnosis process of the fault diagnosis model is as follows: sensors monitor electromechanical equipment and acquire electromechanical equipment data. After the electromechanical equipment data undergoes short-time windowed Fourier transform, features are extracted by ResNet and then fused by BiFPN. The data then undergoes convolution, pooling, and resizing operations in the lightweight Transformers module, and after adding position markers, it is input into the Transformers encoder. The output of the Transformers encoder is processed by sequence pooling and linear layers to obtain the output of the lightweight Transformers module. The output of the lightweight Transformers module is then processed by a fully connected layer and the Softmax function to obtain the fault classification prediction result.

[0061] In other embodiments, the loop stacking structure of the scaling dot product attention mechanism is optimized by using max pooling, and a pre-training-fine-tuning transfer learning method is used to transfer the pre-trained model parameters to the target domain and fine-tune the model, which can avoid overfitting caused by insufficient data.

[0062] Reference Figure 1 The process of constructing a fault diagnosis model includes, but is not limited to, the following steps:

[0063] Step S210: Obtain the training dataset;

[0064] Step S220: Input the training dataset into the residual network for feature extraction to obtain multilayer first feature data;

[0065] Step S230: Based on the weighted bidirectional feature pyramid network, feature fusion is performed on the multi-layer first feature data to obtain multi-layer second feature data;

[0066] Step S240: Based on the feature representation module, perform feature mining on the multi-layer second feature data to obtain the prediction results of the training dataset;

[0067] Step S250: Update the model parameters by backpropagation based on the prediction results to obtain the trained fault diagnosis model.

[0068] Furthermore, prior to step S210, the process of constructing the fault diagnosis model also includes the following steps:

[0069] Step S310: Obtain historical data of electromechanical equipment, wherein the historical data includes normal data and abnormal data of different fault types;

[0070] Step S320: Perform a short-time windowed Fourier transform on the historical data to obtain data samples;

[0071] Step S330: Divide the data samples into a training dataset, a validation dataset, and a test dataset. The training dataset is used to train the fault diagnosis model, the test dataset is used to detect the generalization ability of the fault diagnosis model after training, and the validation dataset is used to detect whether the fault diagnosis model after training is overfitting.

[0072] Specifically, normal data and abnormal data of different fault types are collected and normalized to prevent partial data from affecting the overall effect. The above data undergoes a short-time windowed Fourier transform and is then divided into training, validation, and test datasets according to a certain ratio.

[0073] ResNet is used as the backbone network of the model to extract features from the input. BiFPN is used as the feature network, and the first feature data {P3, P4, P5, P6} from different layers of the backbone network are repeatedly fused using BiFPN in both top-down and bottom-up directions. A lightweight Transformers module processes the fused features extracted by the model to obtain the prediction results for the training data. Backpropagation updates the parameters in ResNet, BiFPN, and the lightweight Transformers module until all data in the training dataset participates in the training, resulting in the fault diagnosis model.

[0074] Import the test dataset into the trained fault diagnosis model to view the prediction effect, and use the validation dataset to detect whether the fault diagnosis model has overfitting.

[0075] According to some embodiments of the present invention, the standard Fourier transform transforms the signal from the time domain to the frequency domain, thereby uncovering fault features hidden in the time domain. However, due to the long acquisition period, content from different frequencies is easily mixed, making features indistinguishable. Therefore, the standard Fourier transform has significant limitations in practical applications. The embodiments of the present invention employ a short-time windowed Fourier transform, which extracts multiple small segments from the entire time slice using a window function, and then performs a Fourier transform on each small segment to obtain the frequency variation over time.

[0076] This invention employs a Gaussian window as the window function. As a smoothing function, the Gaussian window avoids frequency domain oscillations compared to the commonly used rectangular window. The formula for the entire short-time windowed Fourier transform is:

[0077]

[0078] Where g(ut) is the window function, f(u) is the input historical data, and f(w,t) is a two-dimensional signal containing both frequency and time dimensions.

[0079] According to some embodiments of the present invention, the residual network adopts ResNet. In neural network models, the depth of the network is crucial to the performance of the model. When the number of network layers is increased, the network can extract implicit features related to faults in the sensor data of electromechanical equipment. Traditional deep networks often face the problem of degradation, that is, when the network depth increases, the network accuracy saturates or even decreases. Moreover, important information is lost with each convolution operation. When building a deep network by stacking new layers, it is easy for the higher layers to fail to extract features, or even for degradation to occur. However, the residual blocks in ResNet process the information processed in the previous step and the information needed in the current step together, which has a loss reduction effect, making the extracted features contain more valuable information.

[0080] Traditional neural network models lack cross-layer connections. When the input feature is x, after feature extraction, the output is H(x), which is then fed into the next layer for feature extraction. ResNet, however, introduces cross-layer connections, fusing the features extracted in the current layer with the features extracted in the previous layer (i.e., the input features of this layer). The resulting fused features are then fed into the next layer for feature extraction. ResNet is represented as follows:

[0081] F(x) = H(x) + x;

[0082] Where H(x) is the feature extracted by the current layer, x is the feature extracted by the previous layer, i.e., the input feature of the current layer, and F(x) is the output feature of the current layer.

[0083] According to some embodiments of the present invention, the weighted bidirectional feature pyramid network adopts BiFPN, and BiFPN is repeatedly used to perform top-down and bottom-up feature fusion on the multi-layer first feature data obtained by ResNet.

[0084] Specifically, traditional feature fusion models only perform top-down feature fusion, based on information flow in one direction, and then perform a concatenation operation. This easily leads to sparsity of the fused features, resulting in the loss of fault-related information and failing to achieve accurate fault diagnosis. The embodiments of this invention employ BiFPN, which has two characteristics: bidirectional cross-scale connectivity and weighted feature fusion. (Continue to refer to...) Figure 2In the BiFPN part, firstly, the nodes with only one input in the top layer P6(F6(x)) and the bottom layer P3(F3(x)) of ResNet are removed because these nodes only contain one input and do not perform feature fusion, thus contributing very little to the feature network. Secondly, for the same layer, a connection from the initial input to the output is added. The feature output of each layer not only fuses the output features of the previous layer but also the output features of the next layer. During fusion, the contribution of features at different resolutions is different, so an additional weight is added to each input during feature fusion, and the network learns the importance of each input feature. Weight normalization is used to limit the value range of each weight. The weighted input features are as follows:

[0085]

[0086] Among them, by each w i Then add a ReLU to ensure w i >=0, ε=0.0001 is a very small value used to avoid numerical instability.

[0087] Finally, the BiFPN structure is repeated multiple times to achieve more advanced feature fusion, the formula of which is:

[0088]

[0089]

[0090] in, It is the input feature of the x-th layer, i.e., F x (x), It is an intermediate feature of the xth layer on the top-down path. This represents the output feature of the x-th layer along the bottom-up path. Furthermore, to optimize efficiency, depthwise separable convolutions can be used for feature fusion, with batch normalization (BN) and an activation function added after each convolution.

[0091] According to some embodiments of the present invention, step S240, which involves performing feature mining on the multi-layer second feature data based on the feature representation module to obtain the prediction result of the training dataset, includes the following steps:

[0092] Step S410: Convolve and label the second feature data of each layer to obtain the label embedding matrix;

[0093] Step S420: Input the tag embedding matrix into the Transformer encoder to obtain the encoder output matrix;

[0094] Step S430: Determine the importance weight value of each tag embedding matrix based on the encoder output matrix;

[0095] In step S440, the importance weight values ​​of each label embedding matrix are passed through a fully connected layer and then output by the softmax function to obtain the prediction result.

[0096] In step S420, the step of inputting the tag embedding matrix into the Transformer encoder to obtain the encoder output matrix includes the following steps:

[0097] Step S421: Perform a linear transformation on the input tag embedding matrix to obtain the query matrix, key matrix, and value matrix;

[0098] Step S422: Based on the multi-head attention mechanism, determine the encoder output matrix according to the query matrix, key matrix, and value matrix.

[0099] In this embodiment, the lightweight Transformers module performs dimensionality reduction on the features through convolution before inputting them into the Transformers encoder, thereby achieving compression of the second feature data and efficient extraction of long-term data. Specifically, the second feature data obtained from each layer of BiFPN is... As input to the lightweight Transformers module, it is represented as: x∈R H×W×C Then, using a convolutional tokenizer, rich token information, i.e., a token embedding matrix, is generated through operations such as convolution, activation, and pooling. Its formula is:

[0100] x0=MaxPool(ReLU(Conv2d(x)));

[0101] In the Transformers encoder, three parameter matrices W are defined. Q W K W V A linear transformation is performed on x0 to obtain the corresponding query vector, key vector, and value vector. These vectors are then concatenated to obtain the query matrix Q, key matrix K, and value matrix V, thus obtaining the attention weights for each tag embedding matrix. A scaling factor is then used. After scaling and passing through the softmax function, multiplying by the value matrix V yields the attention mechanism function:

[0102]

[0103] Here, Attention represents the attention mechanism function, and Softmax is the normalized exponential function.

[0104] A multi-head attention mechanism is employed to extract sequence information from multiple dimensions, as detailed below:

[0105] MultiHead(Q,K,V)=Concat(head1,...,head h );

[0106]

[0107]

[0108] Among them, MultiHead is a multi-head attention mechanism, Concat means concatenation, and head i Let d represent the i-th head, h be the number of heads in the attention mechanism, O be the input matrix, and d be the number of heads in the input matrix. V Let d be the dimension of the value matrix. m This is the length of the vector.

[0109] This embodiment introduces a novel sequence pooling strategy to replace the traditional [class] tag, preserving relevant information about different parts of the input features contained in the output sequence, reducing the number of forwarding tokens. Its output sequence is mapped as T: Encoder output matrix X L as follows:

[0110] X L =f(x)∈R b×n×d ;

[0111] Among them, X L is the output of the L-layer encoder f(), b is the batch size, n is the sequence length, and d is the total embedding dimension.

[0112] After obtaining the encoder output matrix, X L Input to linear layer g(x) L )∈R d×1 And output it through the softmax function:

[0113] X′ L =softmax(g(x) L ) T )∈R b×1×n ;

[0114] This leads to the importance weight z for each input token, as shown in the following formula:

[0115] z = X′ L X L =softmax(g(x) L ) T )×X L ∈R b×1×d ;

[0116] Finally, the result z is passed through a fully connected layer and output by the softmax function.

[0117] The embodiments of the present invention have the following beneficial effects:

[0118] 1. Achieve efficient feature fusion: By using the bidirectional cross-scale connection of the BiFPN network and the weighted feature fusion strategy, features of different scales are fused to obtain features with better discriminative power.

[0119] 2. To alleviate the problem of limited performance of deep learning in fault diagnosis under small sample size, a lightweight Transformer module is introduced, which has the characteristics of low parameter quantity and strong feature learning ability. It can learn the temporal information in multi-scale features when there are insufficient fault samples, and achieve good fault diagnosis effect.

[0120] This invention also provides a fault diagnosis system for electromechanical equipment, comprising:

[0121] The first module is used to acquire data from electromechanical equipment;

[0122] The second module is used to input electromechanical equipment data into the fault diagnosis model to obtain the diagnostic results of the electromechanical equipment.

[0123] The fault diagnosis model includes a residual network, a weighted bidirectional feature pyramid network, and a feature representation module. The feature representation module adopts a cyclic stacked structure based on a scaled dot product attention mechanism. The construction process of the fault diagnosis model includes:

[0124] Obtain the training dataset;

[0125] The training dataset is input into the residual network for feature extraction to obtain multi-layer first feature data.

[0126] Based on a weighted bidirectional feature pyramid network, feature fusion is performed on multi-layer first feature data to obtain multi-layer second feature data;

[0127] Feature mining is performed on the multi-layer second feature data based on the feature representation module to obtain the prediction results of the training dataset;

[0128] The model parameters are updated by backpropagation based on the prediction results to obtain the trained fault diagnosis model.

[0129] It is understood that the content of the above-mentioned fault diagnosis method embodiments for electromechanical equipment is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above-mentioned fault diagnosis method embodiments for electromechanical equipment, and the beneficial effects achieved are also the same as those achieved in the above-mentioned fault diagnosis method embodiments for electromechanical equipment.

[0130] 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.

[0131] 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.

[0132] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A fault diagnosis method for electromechanical equipment, characterized in that, Includes the following steps: Acquire electromechanical equipment data; The electromechanical equipment data is input into the fault diagnosis model to obtain the diagnostic results of the electromechanical equipment; The fault diagnosis model includes a residual network, a weighted bidirectional feature pyramid network, and a feature representation module. The feature representation module adopts a cyclic stacked structure based on a scaled dot product attention mechanism. The weighted bidirectional feature pyramid network adopts a BiFPN network structure. The feature representation module adopts a lightweight Transformers module structure. The construction process of the fault diagnosis model includes: Obtain the training dataset; The training dataset is input into the residual network for feature extraction to obtain multi-layer first feature data; Based on a weighted bidirectional feature pyramid network, feature fusion is performed on the first feature data of multiple layers to obtain the second feature data of multiple layers; specifically, BiFPN is repeatedly used to perform feature fusion on the first feature data of multiple layers from top to bottom and from bottom to top; the weighted bidirectional feature pyramid network removes nodes with one input and adds a connection from the starting input to the output for the same level, and the feature output of each layer not only fuses the output features of the previous layer, but also fuses the output features of the next layer. Based on the feature representation module, feature mining is performed on the multi-layer second feature data to obtain the prediction results of the training dataset; specifically, it includes: convolving and labeling the second feature data of each layer to obtain a label embedding matrix; and inputting the label embedding matrix into a Transformer encoder to obtain the encoder output matrix. The importance weight value of each tag embedding matrix is ​​determined based on the encoder output matrix; The importance weights of each label embedding matrix are passed through a fully connected layer and then output by a softmax function to obtain the prediction result; the label embedding matrix is ​​represented as follows. , This represents the second feature data. This represents a two-dimensional convolution operation. For activation function, This is a max pooling operation; Based on the prediction results, backpropagation is performed to update the model parameters, resulting in a trained fault diagnosis model. A novel sequence pooling strategy is introduced to replace the traditional [class] tag, preserving relevant information about different parts of the input features contained in the output sequence, reducing the number of forwarding tokens. Its output sequence is mapped as follows: encoder output matrix as follows: ； in, It is an L-layer encoder f The output of () b It refers to the batch size. n It is the sequence length. d R is the total embedding dimension, and R represents the dimension tensor space; After obtaining the encoder output matrix, Passing to linear layer And output it using the softmax function: ； This leads to the importance weight z for each input token, as shown in the following formula: ； The result z is passed through a fully connected layer and then output by the softmax function.

2. The fault diagnosis method for electromechanical equipment according to claim 1, characterized in that, The process of constructing the fault diagnosis model also includes: Acquire historical data of electromechanical equipment, wherein the historical data includes normal data and abnormal data of different fault types; Perform a short-time windowed Fourier transform on the historical data to obtain data samples; The data samples are divided into a training dataset, a validation dataset, and a test dataset. The training dataset is used to train the fault diagnosis model, the test dataset is used to detect the generalization ability of the fault diagnosis model after training, and the validation dataset is used to detect whether the fault diagnosis model after training is overfitting.

3. The fault diagnosis method for electromechanical equipment according to claim 2, characterized in that, The process of performing a short-time windowed Fourier transform on the historical data to obtain data samples includes the following steps: The historical data over the entire time segment is divided into multiple short-time segments of historical data; Fourier transform is performed on historical data from multiple short time segments to obtain data samples.

4. The fault diagnosis method for electromechanical equipment according to claim 2, characterized in that, The formula for the short-time windowed Fourier transform is: ； in, For window functions, For the input historical data, It is a two-dimensional signal, containing both frequency and time dimensions. Indicates the current time.

5. The fault diagnosis method for electromechanical equipment according to claim 1, characterized in that, The residual network uses ResNet, which is represented as: ； in, Features extracted from the current layer The features extracted from the previous layer are the input features of the current layer. This represents the output feature of the current layer.

6. The fault diagnosis method for electromechanical equipment according to claim 1, characterized in that, Bi FPN is represented as: ； ； in, It is the first x The input features of the layer, i.e., the residual network of the 1st layer x The first feature data output by the layer , Indicates the first x+ Input features of layer 1; It is the first one on the top-down path x Intermediate features of the layer; It is the first one on the bottom-up path x The output features of the layer; Indicates the first [number]th ... x- Output features of layer 1; , , , and All of these represent learnable weight parameters. This represents a constant used for zero protection. This represents a function that matches the feature sizes of each layer through upsampling and subsampling.

7. The fault diagnosis method for electromechanical equipment according to claim 1, characterized in that, The step of inputting the marker embedding matrix into the Transformer encoder to obtain the encoder output matrix includes the following steps: A linear transformation is performed on the input tag embedding matrix to obtain the query matrix, key matrix, and value matrix; Based on the multi-head attention mechanism, the encoder output matrix is ​​determined according to the query matrix, the key matrix, and the value matrix.

8. A fault diagnosis system for electromechanical equipment, characterized in that, include: The first module is used to acquire data from electromechanical equipment; The second module is used to input the electromechanical equipment data into the fault diagnosis model to obtain the diagnostic results of the electromechanical equipment. The fault diagnosis model includes a residual network, a weighted bidirectional feature pyramid network, and a feature representation module. The feature representation module adopts a cyclic stacked structure based on a scaled dot product attention mechanism. The weighted bidirectional feature pyramid network adopts a BiFPN network structure. The feature representation module adopts a lightweight Transformers module structure. The construction process of the fault diagnosis model includes: Obtain the training dataset; The training dataset is input into the residual network for feature extraction to obtain multi-layer first feature data; Based on a weighted bidirectional feature pyramid network, feature fusion is performed on the first feature data of multiple layers to obtain the second feature data of multiple layers; specifically, BiFPN is repeatedly used to perform feature fusion on the first feature data of multiple layers from top to bottom and from bottom to top; the weighted bidirectional feature pyramid network removes nodes with one input and adds a connection from the starting input to the output for the same level, and the feature output of each layer not only fuses the output features of the previous layer, but also fuses the output features of the next layer. Based on the feature representation module, feature mining is performed on the multi-layer second feature data to obtain the prediction result of the training dataset; specifically, this includes: convolving and labeling the second feature data of each layer to obtain a label embedding matrix; inputting the label embedding matrix into a Transformer encoder to obtain an encoder output matrix; determining the importance weight value of each label embedding matrix based on the encoder output matrix; and passing the importance weight value of each label embedding matrix through a fully connected layer and outputting it by a softmax function to obtain the prediction result; the label embedding matrix is ​​represented as follows. , This represents the second feature data. This represents a two-dimensional convolution operation. For activation function, This is a max pooling operation; Based on the prediction results, backpropagation is performed to update the model parameters, resulting in a trained fault diagnosis model. A novel sequence pooling strategy is introduced to replace the traditional [class] tag, preserving relevant information about different parts of the input features contained in the output sequence, reducing the number of forwarding tokens. Its output sequence is mapped as follows: encoder output matrix as follows: ； in, It is an L-layer encoder f The output of () b It refers to the batch size. n It is the sequence length. d R is the total embedding dimension, and R represents the dimension tensor space; After obtaining the encoder output matrix, Passing to linear layer And output it using the softmax function: ； This leads to the importance weight z for each input token, as shown in the following formula: ； The result z is passed through a fully connected layer and then output by the softmax function.

Images