Anti-noise fault diagnosis method based on multi-channel Transform model

Abstract

The invention discloses an anti-noise fault diagnosis method based on a multi-channel Transform model. The method comprises the following steps: constructing different modal information for vibration signal data; the time sequence information is converted into picture data through wavelet transformation, and the picture data are input into a convolutional neural network to capture spatial feature information; different modal components are constructed through variational modal decomposition; aligning the feature dimensions of different modes and inputting the feature dimensions into an MCform network model; self-adaptive weighted fusion is carried out on the output of each channel by using a feed-forward layer; training a network for the predicted value and the true value by using a cross entropy loss function; establishing an overall fault diagnosis offline model; during fault diagnosis, sampling to obtain online process data X, processing the online process data X, converting the data into different modal representations, inputting the different modal representations into the model, and obtaining a posterior probability of an online data health state; and normalizing the posterior probability by using a normalization exponential function to obtain the health state of the current data. The method can provide scientific and reasonable decision support for production operators.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis in industrial production processes, and specifically relates to a noise-resistant fault diagnosis method based on a multi-channel Transformer model. Background Art

[0002] With the rapid development of Industry 4.0, industrial equipment is increasingly tending towards automation and complexity, and the scale of equipment operation monitoring data has also shown an explosive growth. These data not only include sensor data, control signals, and visual data, but also cover multi-dimensional information such as the operating state of the equipment, environmental changes, and operating conditions. With the continuous increase in the amount of data, traditional manual monitoring and maintenance methods are no longer sufficient to meet the need for efficient and timely fault detection. Therefore, the automation of equipment health monitoring is not only an inevitable trend in the industry's development but also a necessary means to ensure the long-term safe and stable operation of industrial equipment. In this context, fault diagnosis and prediction technologies based on industrial big data have emerged. By applying advanced technologies such as data-driven machine learning and deep learning, valuable features and information can be extracted from massive sensor data, and thus early identification and accurate positioning of equipment faults can be achieved. These technologies can help engineering and technical personnel timely discover potential problems during equipment operation, take preventive maintenance measures in advance, reduce the probability of faults, and thus maximize the operating efficiency and safety of the equipment. In addition, by means of these technical means, downtime can be effectively reduced, the service life of the equipment can be extended, and production interruptions or safety accidents caused by equipment failures can be avoided, thus providing strong guarantee for the efficient operation of industrial systems.

[0003] In the field of fault diagnosis and prediction, time series monitoring data collected by sensors is the core of analysis. These data include vibration signals, temperature data, sound signals, etc., and can well reflect the operating state of the equipment. By deeply analyzing these data, researchers and engineering and technical personnel can build a high-precision fault diagnosis model. Traditional fault diagnosis methods mostly rely on a single data modality (such as only using vibration data or temperature data) and extract features through classical signal processing techniques (such as Fourier transform, waveform analysis, etc.). However, the accuracy and reliability of such methods are usually limited by the singularity of the data modality used and it is difficult to comprehensively capture the multi-dimensional information of the equipment. Therefore, these methods often fail to meet the requirements for high precision and high reliability in actual industrial applications.

[0004] In recent years, with the rapid development of industrial Internet technology, devices can provide richer and more diverse data resources. For example, in addition to traditional vibration signals, temperature data, etc., devices can also provide multi-modal data such as images and texts. These multi-modal data usually contain more potential information and can more comprehensively reflect the health status of the device. For example, through visual image data, physical damages such as wear and cracks on the external part of the device can be identified; while text data (such as operation logs, maintenance records, etc.) can provide historical maintenance information of the device, providing important background knowledge for fault prediction and diagnosis. Nevertheless, in the actual industrial environment, there is still a large amount of environmental noise, seriously interfering with the effective fusion between different modal features, resulting in the failure of device fault diagnosis in real industrial scenarios. Therefore, there is an urgent need to provide an anti-noise fault diagnosis method based on a multi-channel Transformer model to effectively solve the problem of noise interference between multi-modal features. Summary of the Invention

[0005] Aiming at the problems existing in the above-mentioned prior art, the present invention provides an anti-noise fault diagnosis method based on a multi-channel Transformer model. This method can simultaneously process data information of multiple modalities and reduce noise interference between different modalities, can significantly improve the robustness and reliability of operation status evaluation, can provide scientific and reasonable decision-making support for production operators, is conducive to quickly locating the fault position, and can quickly and timely realize the maintenance of the device.

[0006] To achieve the above object, the present invention provides an anti-noise fault diagnosis method based on a multi-channel Transformer model, including the following steps:

[0007] Step 1: Convert offline vibration signal data into multi-modal data to construct a rich fault data set;

[0008] A1: Use vibration sensors installed in the rotating machinery system to collect industrial process data during operation and store and record it using a database; extract historical industrial process data from the database as an offline data set X=(x 1 ,x 2 ,...,x N ), and label the health categories Y=(y 1 ,y 2 ,...,y N ) of the offline data set X=(x 1 ,y 2 ,...,y N ), where N is the number of samples and the length of each sample is L;

[0009] A2: Perform data normalization preprocessing on the X data according to formula (1), and perform regularization preprocessing on the X data and Y data according to formula (2);

[0010]

[0011] Where X is the input data, min(X) is the minimum value of the input data, max(X) is the minimum value of the input data, μ represents the mean of the input data, and σ is the standard deviation of the input data;

[0012] A3: Use the variational mode decomposition signal processing technique according to formula (3) to decompose the data X into the components of the vibration signal, and select the first K components as the intrinsic modes, denoted as X vmd ;

[0013]

[0014] Where u k is the modal function, ω k is the modal center frequency, δ(t) represents the impulse function, f is the input signal, and k is the decomposition scale;

[0015] A4: Use the continuous wavelet transform according to formula (4) to construct the time-frequency diagram X of the data X f ;

[0016]

[0017] Where a is the scale parameter, b is the translation parameter, ψ(t) is the mother wavelet, and ψ * (t) is the complex conjugate of ψ(t);

[0018] A5: Use deep convolution to extract the high-dimensional features X of the time-frequency diagram cwt , and keep the feature dimensions unified and aligned with the intrinsic modes. In this process, obtain the mapping result h generated by the t-th convolutional kernel according to formula (5) t ;

[0019] h t = XW t + b t (5);

[0020] Where W t represents the weight parameter of the t-th convolutional kernel, and b t represents the bias parameter;

[0021] A6: Construct the multi-modal representation M = (X 1 , X 2 ,..., X m ) for each sample according to formula (6), where m is the number of modalities;

[0022] M = connect[X; X vmd ; X cwt (6);

[0023] In the formula, connect[·] represents the stacking operation;

[0024] Step 2: Input the multi-modal signals and train the MCformer model to establish an overall offline fault diagnosis model;

[0025] B1: Divide the input variable M of each input modality i i into patches of length l, that is, divide the signal into n = L / l segments, and this process generates a patch sequence Map the patch sequence through a linear transformation using formula (7) into a latent space of dimension d model

[0026]

[0027] In the formula, represents the high-dimensional representation after mapping, represents the linear transformation matrix;

[0028] B2: Introduce the positional encoding W pos using formula (8) and formula (9) to help the model understand the position information in the patch sequence;

[0029] PE (pos,2i) = sin(pos / 10000 2i / d )(8);

[0030] PE (pos,2i+1) = cos(pos / 10000 2i / d )(9);

[0031] B3: Embed the signals of each modality into the one represented by and W pos jointly as

[0032]

[0033] B4: Embed the multi-modal information into the encoder and learn the fault features of each modality i through the channel-independent multi-head self-attention mechanism. The attention output of a single attention head is shown in formula (11);

[0034]

[0035] ​Wherein, represents the query matrix, represents the key matrix, represents the value matrix, d k represents the dimension of the key vector, which is used to prevent gradient explosion;

[0036] B5: Concatenate multiple attention heads according to formula (12) to keep the input and output dimensions the same;

[0037]

[0038] Wherein, o i is the output result of the encoder, W o represents the linear projection after concatenating multiple heads;

[0039] B6: Construct the adaptive weight W according to formula (13) using the inverse information entropy i ;

[0040]

[0041] Wherein,

[0042] B7: Use formula (14) to perform weighted fusion representation on the features learned by each channel in the feed-forward layer;

[0043] y = W i (flatten(o 1 ,..., o m )·W 1 + b 1 )·W 2 + b 2 (14);

[0044] Wherein, y represents the prediction result, W 1 , W 2 respectively represent different mapping matrices, flatten(·) represents the flattening operation, b 1 , b 2 respectively represent different bias vectors;

[0045] B8: Normalize the posterior probability according to the normalization exponential function in formula (15);

[0046]

[0047] B9: Calculate the classification loss function L(p, y) according to formula (16);

[0048]

[0049] B10: Reverse fine-tune the model parameters until the classification loss error is minimized, and train to obtain the complete fault diagnosis model MCformer;

[0050] Step 3: Use online data for operating status evaluation;

[0051] C1: Use vibration sensors installed in the rotating machinery system to perform online sampling to obtain online process data X, and perform normalization and standardization preprocessing on X;

[0052] C2: Construct time-frequency diagrams and intrinsic modes for the online data to obtain multi-modal data;

[0053] C3: Use deep convolution to capture the fault feature representation of the time-frequency diagram, and align the feature dimensions with the intrinsic mode and the original signal;

[0054] C4: Input the multi-modal data into the trained fault diagnosis model MCformer, and output the posterior probability p of different samples;

[0055] C5: Determine the health state with the maximum probability according to the posterior probability, which is the final online fault diagnosis result of the sample.

[0056] The present invention provides a method combining a multi-channel method and a Transformer model. By using channel-independent Tranformer networks to capture fault features of different modalities respectively, it effectively prevents the mutual interference between different modality noises. For a single vibration signal, a multi-modal expression is constructed to enrich its fault feature representation, and at the same time, the problem of noise mutual interference between different modalities is considered. The present invention innovatively designs a weighted algorithm based on inverse information entropy for the feature outputs of different channels. This method encourages the model to focus on more robust modality features in the feed-forward layer, thereby significantly improving the reliability of fault diagnosis. During the training process, the output result of the model is combined with the fault labels for supervised learning, and different health states are accurately divided according to the category labels of each fault. The model can learn the key feature expressions related to these faults. Calculate the final loss according to the predicted label and the true label of the model, and use the loss function to perform reverse fine-tuning on the model to optimize the model parameters until the classification loss error is minimized, thereby obtaining a high-precision fault diagnosis model, effectively ensuring the accuracy of fault diagnosis. This model can convert the vibration signal of the rotating machinery into multi-modal data, and adopt a multi-channel method to prevent the mutual interference of noise information, improving the accuracy and robustness of fault diagnosis.

[0057] The advantage of this method is that it can process multi-modal data information simultaneously and reduce the noise interference between different modalities, especially suitable for industrial production environments with high noise levels and multi-source data. At the same time, this method utilizes channel independence to independently process the signal features of each modality, effectively reducing the noise interference between multi-modal features, greatly improving the quality and efficiency of feature extraction, and thus significantly enhancing the robustness and reliability of the operating state evaluation. By accurately judging the types of equipment failures, this method can provide scientific and reasonable decision-making support for production operators, facilitate the rapid location of fault positions, and enable the rapid and timely maintenance of equipment, thereby significantly improving the overall operating efficiency of the industrial process, reducing energy and resource waste, facilitating the efficient fault detection operation of rotating machinery, and effectively ensuring the safe, stable, and efficient operation of the industrial process. Brief Description of the Drawings

[0058] Figure 1 is the overall flowchart of the fault diagnosis of the rotating machinery system in the present invention;

[0059] Figure 2 is a schematic diagram of the experimental results of the optimal decomposition scale of variational mode decomposition in the present invention;

[0060] Figure 3 is a schematic diagram of the experimental results of the comparison of the adaptive weighting method in the present invention;

[0061] Figure 4 is a schematic diagram of the visualization results of the features of the rotating machinery fault diagnosis based on the MCformer model in the present invention;

[0062] Figure 5 is a schematic diagram of the confusion matrix of the rotating machinery fault diagnosis based on the MCformer model in the present invention;

[0063] Figure 6 is a schematic diagram of the visualization structure of the multi-head self-attention of the rotating machinery fault diagnosis based on the MCformer model in the present invention;

[0064] Figure 7 is a schematic diagram of the structure of the rotating machinery test platform in the present invention. Detailed Embodiments

[0065] The present invention will be further described below with reference to the accompanying drawings.

[0066] As Figures 1 to 7 shown, the present invention provides a noise-resistant fault diagnosis method based on a multi-channel Transformer model (MCformer, Multi-channel Transformer), including the following steps:

[0067] Step 1: Convert the offline vibration signal data into multi-modal data and construct a rich fault data set;

[0068] A1: Use vibration sensors installed in the rotating machinery system to collect industrial process data during operation and store and record it using a database; Extract historical industrial process data from the database as the offline data set X = (x 1 , x 2 ,..., x N ), and label the health category Y = (y 1 , y 2 ,..., y N ) of the offline data set X = (x 1 , y 2 ,..., y N ), where N is the number of samples and the length of each sample is L;

[0069] A2: Perform data normalization preprocessing on the X data according to formula (1), and perform regularization preprocessing on the X data and Y data according to formula (2);

[0070]

[0071] In the formula, X is the input data, min(X) is the minimum value of the input data, max(X) is the minimum value of the input data, μ represents the mean value of the input data, and σ is the standard deviation of the input data;

[0072] A3: Use the variational mode decomposition signal processing technology according to formula (3) to decompose the data X into the components of the vibration signal, and select the first K components as the intrinsic modes, denoted as X vmd ; In this way, the multi-scale information of the signal can be enriched;

[0073]

[0074] In the formula, u k is the modal function, ω k is the modal center frequency, δ(t) represents the impulse function, f is the input signal, and k is the decomposition scale;

[0075] A4: Use continuous wavelet transform according to formula (4) to construct the time-frequency diagram X f of the data X to convert the time series information into picture data;

[0076]

[0077] In the formula, a is the scale parameter, b is the translation parameter, ψ(t) is the mother wavelet, and ψ * (t) is the complex conjugate of ψ(t);

[0078] A5: Use depth convolution to extract the high-dimensional features X of the time-frequency map cwt , and keep it unified and aligned with the feature dimension of the intrinsic mode. In this process, the mapping result h generated by the t-th convolution kernel is obtained according to formula (5) t ;

[0079] h t = XW t + b t (5);

[0080] In the formula, W t represents the weight parameter of the t-th convolution kernel, and b t represents the bias parameter;

[0081] A6: Construct the multi-modal representation M = (X 1 , X 2 ,..., X m ) for each sample according to formula (6), where m is the number of modes;

[0082] M = connect[X; X vmd ; X cwt (6);

[0083] In the formula, connect[·] represents the stacking operation;

[0084] Step 2: Input the multi-modal signal and train the MCformer model to establish an overall fault diagnosis offline model;

[0085] B1: Divide the input variable M of each input mode i i into patches of length l, that is, divide the signal into n = L / l segments, and this process generates a patch sequence Map the patch sequence into a latent space of dimension d model through linear transformation using formula (7)

[0086]

[0087] In the formula, represents the high-dimensional representation after mapping, and represents the linear transformation matrix;

[0088] B2: Introduce the positional encoding W pos using formula (8) and formula (9) to help the model understand the position information in the patch sequence;

[0089]

[0090] B3: Embed the signals of each modality according to formula (10) into what is jointly represented by and W pos as

[0091]

[0092] B4: Embed the multi-modal information into the encoder, and learn the fault features of each modality i through the channel-independent multi-head self-attention mechanism. The attention output of a single attention head is shown in formula (11);

[0093]

[0094] In the formula, represents the query matrix, represents the key matrix, represents the value matrix, and d k represents the dimension of the key vector, which is used to prevent gradient explosion;

[0095] B5: Concatenate multiple attention heads according to formula (12) to keep the input and output dimensions the same;

[0096]

[0097] In the formula, o i is the output result of the encoder, and W o represents the linear projection after concatenating the multi-heads;

[0098] B6: Construct the adaptive weight W according to formula (13) using the inverse information entropy i ;

[0099]

[0100] In the formula,

[0101] B7: Use formula (14) to perform weighted fusion representation of the features learned by each channel in the feed-forward layer;

[0102]

[0103] In the formula, y represents the prediction result, and W 1 , W 2 represent different mapping matrices respectively, flatten(·) represents the flattening operation, b 1 , b 2 represent different bias vectors respectively;

[0104] B8: Normalize the posterior probability according to the normalization exponential function in formula (15);

[0105]

[0106] B9: Calculate the classification loss function \(L(p, y)\) according to formula (16);

[0107]

[0108] B10: Backward fine-tune the model parameters until the classification loss error is minimized, and train to obtain the complete fault diagnosis model MCformer;

[0109] Step 3: Use online data for operating status evaluation;

[0110] C1: Use the vibration sensors installed in the rotating machinery system to perform online sampling to obtain online process data \(X\), and perform normalization and standardization preprocessing on \(X\);

[0111] C2: Construct time-frequency diagrams and intrinsic modes for the online data to obtain multi-modal data;

[0112] C3: Use deep convolution to capture the fault feature representation of the time-frequency diagram, and align the feature dimensions with the intrinsic mode and the original signal;

[0113] C4: Input the multi-modal data into the trained fault diagnosis model MCformer, and output the posterior probability \(p\) of different samples;

[0114] C5: Determine the health state with the highest probability according to the posterior probability, which is the final online fault diagnosis result of the sample.

[0115] The technical solution of the present invention will be described in detail below in conjunction with embodiments, and its feasibility will be verified.

[0116] Embodiment:

[0117] In this embodiment, a rotating machinery system is preferably used as the research object. Rotating machinery is widely used in various industries and daily life, including electric motors, generators, pumps, compressors, fans, turbines, conveyor belt systems, etc. These devices play a crucial role in the modern industrial system, and their safe and stable operation can effectively ensure the continuity and efficiency of the production process. However, various faults may occur during the long-term operation of rotating machinery, such as bearing damage, gear wear, imbalance, abnormal vibration, etc. Fault diagnosis is a key measure to ensure the reliability of rotating machinery and extend its service life. In view of the characteristics that the working environment of rotating machinery equipment is harsh and industrial noise affects the actual sensor data acquisition, the present invention effectively reduces the noise interference between multi-modal features by using a channel-independent method, and more accurately and robustly diagnoses the current equipment health state.

[0118] In the present invention, the rotating equipment test platform is as follows Figure 7 shown, including a 2-horsepower (1.5KW) motor, a torque sensor / decoder, a vibration sensor, a power tester and an electronic controller. The bearings to be detected are the drive-end bearing and the fan-end bearing. This dataset contains a total of four equipment health conditions: normal, inner ring fault, outer ring fault, and rolling element fault. The data sampling frequency used in the experiment is 12 kHz, and single-point damage is processed on the bearing by electrical discharge machining. The drive-end bearing dataset is selected for the experiment, and the damage diameters are 0.007 inches, 0.014 inches, and 0.021 inches. The experimental data are respectively collected under four conditions of motor speeds of 1797 rpm, 1772 rpm, 1750 rpm, and 1730 rpm. Therefore, 10 types of equipment health states are divided for comparative tests. The specific settings are shown in Table 1.

[0119] Table 1: Rotating Machinery Test Platform

[0120]

[0121] 1) Establish an off-line state evaluation model

[0122] Collect off-line data. The data processing uses the sliding window sampling method. The sliding window size is 300, the sampling length is 1024, there are 5600 groups of training set samples, and 1390 groups of test set samples.

[0123] Combined with expert experience and comparative experiments, the hyperparameter settings of the model are shown in Table 2.

[0124] Table 2: Hyperparameter Settings of MCformer

[0125]

[0126]

[0127] In this embodiment, the length of the input time series is set to 1024, and the size of the input batch is set to 32. During the process of training the model, the loss function uses the cross-entropy loss function, and the Adam optimization algorithm is used to minimize this loss.

[0128] The overall process of multi-channel Transformer anti-noise fault diagnosis is as follows Figure 1 shown, which includes three parts: the data processing stage, the MCformer model, and the fault prediction classifier. Figure 2 For the experimental results of the optimal decomposition scale of variational mode decomposition, the average accuracy rate is 99.64% when the decomposition scale is 4, which is 0.21% higher than when the decomposition scale is 3 and 0.13% higher than when the decomposition scale is 5. Therefore, the decomposition scale 4 is selected to decompose the vibration signal. Figure 3For the comparative experiment of the adaptive weighted method, the maximum accuracy, minimum accuracy, and average accuracy of the inverse information entropy algorithm are 99.90%, 99.41%, and 99.64% respectively. It can be seen that the proposed inverse information entropy weighted algorithm can better select stable fault features, so the average value is relatively stable.

[0129] Figure 4 The visualization results of the fault diagnosis features of rotating machinery based on the MCformer model are shown. Each category has high distinctiveness, and the decision boundaries between categories are clear, indicating that Mcformer can extract relevant fault features for each category. Thus, it can be seen that the present invention has high robustness to different health categories of rotating machinery. Combining Figure 5 Based on the confusion matrix of the fault diagnosis of rotating machinery by the MCformer model, the diagnostic accuracy of each health state can be seen. The diagnostic accuracy of each health category reaches above 99%, indicating that this method can effectively identify each fault category.

[0130] Figure 6 The multi-head self-attention visualization of the fault diagnosis of rotating machinery based on the MCformer model is shown. It can be observed that each attention head has a unique focus of attention or "perspective", so it can capture information in different representation subspaces of the sequence. The model tends to identify the relationships between different positions in the sequence during the training process, rather than just focusing on a single position. This means that the MCformerr model becomes better at identifying and focusing on those specific signals that can characterize faults, which is crucial for improving the accuracy and efficiency of fault detection.

[0131] From the above simulation examples, it can be seen that the present invention has a certain effectiveness. By implementing timely and robust fault diagnosis for rotating machinery, it can provide an important guiding basis for the production adjustment of the actual industrial process.

[0132] The present invention provides a method combining a multi-channel method with a Transformer model, which can be effectively applied to the online health status evaluation operation of a rotating mechanism system. By using a channel-independent Transformer network to capture fault features of different modalities respectively, the mutual interference between different modality noises is effectively prevented. For a single vibration signal, a multi-modal expression thereof is constructed to enrich the representation of its fault features, and at the same time, the problem of noise mutual interference between different modalities is considered. The present invention innovatively designs a weighted algorithm based on inverse information entropy for the feature outputs of different channels. This method encourages the model to pay attention to more robust modality features in the feed-forward layer, thereby significantly improving the reliability of fault diagnosis. During the training process, the output result of the model is combined with the fault label for supervised learning, and different health states are accurately divided according to the category label of each fault. The model can learn the key feature expressions related to these faults. The final loss is calculated according to the predicted label and the true label of the model, and the loss function is used to perform back fine-tuning on the model to optimize the model parameters until the classification loss error is minimized, thereby obtaining a high-precision fault diagnosis model, effectively ensuring the accuracy of fault diagnosis. This model can convert the vibration signal of a rotating machine into multi-modal data and adopt a multi-channel method to prevent the mutual interference of noise information, improving the accuracy and robustness of fault diagnosis.

[0133] The advantages of this method are that it can process data information of multiple modalities simultaneously and reduce the noise interference between different modalities, especially suitable for industrial production environments with a large noise environment and multi-source data. At the same time, this method utilizes channel independence to independently process the signal features of each modality, effectively reducing the noise mutual interference between multi-modal features, greatly improving the quality and efficiency of feature extraction, and thus significantly enhancing the robustness and reliability of the operating state evaluation. By accurately judging the types of equipment faults, this method can provide scientific and reasonable decision-making support for production operators, facilitate quickly locating the fault position, and can quickly and timely achieve the maintenance of the equipment, thereby significantly improving the overall operating efficiency of the industrial process, reducing the occurrence of energy and resource waste, facilitating the efficient fault detection operation of rotating mechanical equipment, and effectively ensuring the safe, stable and efficient operation of the industrial process.

Claims

1. A noise-resistant fault diagnosis method based on a multi-channel Transformer model, characterized in that: The following steps are involved: Step 1: Convert offline vibration signal data into multimodal data to construct a rich fault data set; A1: Use vibration sensors installed in rotating machinery systems to collect industrial process data during operation and store and record them in a database; extract historical industrial process data from the database as an offline data set X = (x1, x2, ..., x N ), and annotate the offline dataset X = (x1, x2, ..., x N ) of health category Y=(y1,y2,...,y N ), where N is the number of samples and the length of each sample is L; A2: Perform data normalization preprocessing on X data according to formula (1), and perform regularization preprocessing on X data and Y data according to formula (2); In the formula, X is the input data, min(X) is the minimum value of the input data, max(X) is the minimum value of the input data, μ represents the mean of the input data, and σ is the standard deviation of the input data; A3: Using formula (3) and variational mode decomposition signal processing technology, decompose the data X into components of the vibration signal, and select the first K components as eigenmodes, expressed as X vmd ; In the formula, u k is each mode function, ω k is the center frequency of each mode, δ(t) represents the impulse function, f is the input signal, and k is the decomposition scale; A4: Use formula (4) to use continuous wavelet transform to construct the time-frequency graph X of data X f ; Where a is the scale parameter, b is the translation parameter, ψ(t) is the mother wavelet, and ψ * (t) is the complex conjugate of ψ(t); A5: Use deep convolution to extract high-dimensional features of time-frequency graphs X cwt , and keep it unified and aligned with the characteristic dimension of the intrinsic mode. In this process, the mapping result h generated by the t-th convolution kernel is obtained according to formula (5): t ; h t =XW t +b t (5); Where W t represents the weight parameter of the tth convolution kernel, b t represents the bias parameter; A6: According to formula (6), construct a multimodal representation M = (X1, X2, ..., X m ), where m is the number of modes; M=connect[X;X vmd ;X cwt ] (6); In the formula, connect[·] represents the stacking operation; Step 2: Input the multimodal signal and train the MCformer model to establish an overall fault diagnosis offline model; B1: The input variable M of each input mode i i Divide into patches of length l, that is, divide the signal into n = L / l segments. This process generates a patch sequence Using formula (7) to transform the patch sequence into Mapped to dimension d model Potential space In the formula, express The high-dimensional representation after mapping, represents a linear transformation matrix; B2: Using formula (8) and formula (9) to introduce position coding W pos , to help the model understand the position information in the patch sequence; ON (pos,2i) =sin(pos / 10000 2i / d ) (8); ON (pos,2i+1) =cos(pos / 10000 2i / d ) (9); B3: According to formula (10), the signal of each mode is embedded into and W pos Commonly expressed as B4: Embed the multimodal information into the encoder and learn the fault features of each modality i through a channel-independent multi-head self-attention mechanism. The attention output of a single attention head is shown in formula (11); In the formula, represents the query matrix, represents the key matrix, represents the value matrix, d k Represents the dimension of the key vector, used to prevent gradient explosion; B5: Concatenate multiple attention heads according to formula (12) to keep the input and output dimensions the same; In the formula, o i is the output of the encoder, W o It represents the linear projection after splicing multiple heads; B6: According to formula (13), the adaptive weight W is constructed using the inverse information entropy i ; In the formula, B7: Use formula (14) to perform weighted fusion representation on the features learned by each channel in the feedforward layer; Where y represents the prediction result, W1 and W2 represent different mapping matrices, flatten(·) represents the flattening operation, and b1 and b2 represent different bias vectors. B8: Normalize the posterior probability according to the normalized exponential function in formula (15); B9: Calculate the classification loss function L(p,y) according to formula (16); B10: Reversely fine-tune the model parameters until the classification loss error is minimized, and the complete fault diagnosis model MCformer is trained; Step 3: Use online data to evaluate the operation status; C1: Use the vibration sensor installed in the rotating mechanical system to obtain the online process data X through online sampling, and perform normalization and standardization preprocessing on X; C2: construct time-frequency diagram and eigenmode of online data to obtain multimodal data; C3: Use deep convolution to capture fault feature representation in time-frequency graph and align feature dimensions with intrinsic modes and original signals; C4: Input the multimodal data into the trained fault diagnosis model MCformer and output the posterior probability p of different samples; C5: Determine the health state with the highest probability based on the posterior probability, which is the final online fault diagnosis result of the sample.

Images