Automobile engine bearing fault diagnosis method based on multi-source heterogeneous data fusion

Abstract

The invention belongs to the field of fault diagnosis in vehicle engineering, and relates to an automobile engine bearing fault diagnosis method based on multi-source heterogeneous data fusion, which is obtained by training a multi-source heterogeneous data training data set. The method comprises the steps of data acquisition and marking, construction of a multi-source heterogeneous data fused automobile engine bearing fault diagnosis model and set hyper-parameters, multi-head attention mechanisms in a Transform architecture are utilized to perform coupling noise reduction on multi-source heterogeneous data, meanwhile, a self-attention mechanism is utilized to pay attention to the part with the most serious noise in a sample, and the noise is reduced. And in cooperation with a SimAM mechanism, multi-source heterogeneous data coupling noise reduction and single-sensor information independent noise reduction are realized, and the noise reduction effect of the network is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of fault diagnosis in vehicle engineering, and particularly relates to an automobile engine bearing fault diagnosis method based on multi-source heterogeneous data fusion. BACKGROUND

[0002] The automobile engine bearing, as a key component of the engine, directly affects the reliable and safe operation of the automobile. However, due to the precision characteristics of the engine, it is very difficult to directly observe the health status of the engine bearing. However, even bearings produced in the same batch will exhibit significantly different residual service life in actual use due to the influence of working environment (urban highway driving, mountain driving, and bumpy road driving), driving habits (rapid acceleration and rapid braking), and load change factors. This uncertainty leads to the complexity of engine maintenance strategies: regular detection of bearings that do not reach the expected life will consume a large amount of resources and increase unnecessary downtime; and checking when approaching the life limit may face serious failure risks, resulting in high repair costs and driving safety problems. In addition, since the evolution law of bearing residual life is closely related to early weak faults, once the bearing has early faults, its residual service life will often be sharply shortened. Therefore, timely identification and handling of these early faults are crucial to ensure stable engine operation.

[0003] Developing a new method for early fault diagnosis of bearings can take appropriate preventive maintenance measures through early fault diagnosis to avoid potential major failures and prolong the service life of bearings and even the entire engine. This not only helps to improve the overall reliability and safety of the automobile, reduce the risk of accidents caused by sudden failures, but also effectively reduces the high repair costs and downtime losses caused by sudden failures, thereby reducing the overall operating costs.

[0004] The prior art has the following defects:

[0005] 1. The early weak fault operating signals of the automobile engine are very similar to the normal operating signals, and the existing methods often have low diagnosis accuracy.

[0006] 2. The existing methods do not cover all fault characteristics, and the data from a single sensor cannot fully capture all information of the bearing operating signals. When using the current mainstream vibration sensor to collect operating signals, the vibration sensor will not capture such signals caused by increased bearing friction due to bearing lubrication problems, resulting in temperature rise. In the case of early small cracks in the bearing, the change of the vibration signal is very subtle and difficult to identify directly through vibration analysis.

[0007] 3. The feature fusion method is simple and lacks adaptability. Existing multi-source heterogeneous data fusion methods often use simple splicing or averaging methods without considering the speciality of different types of data, making it difficult to obtain high diagnostic accuracy.

[0008] 4. Existing methods often only conduct experiments on clean and ideal laboratory data, with little consideration for whether different types of noise interference exist. In fact, due to the different operating environments of cars, noise can have a significant impact on data, further reducing the accuracy of engine bearing fault diagnosis.

[0009] Therefore, in view of the above-mentioned deficiencies in actual production and implementation, the present application is modified and improved, and the spirit and concept of seeking good are followed, and the assistance of professional knowledge and experience, as well as the trial of many parties, are used to create the present application, which provides a fault diagnosis method for multi-source heterogeneous data fusion of automobile engine based on Transformer, to solve the problems. SUMMARY

[0010] The present application provides a fault diagnosis method for automobile engine bearing based on multi-source heterogeneous data fusion, which solves the problems of low diagnostic accuracy, insufficient coverage of fault features, and the inability of single sensor data to capture all information of bearing operation signals, making it difficult to obtain high diagnostic accuracy and early fault diagnosis of automobile engine bearing.

[0011] The technical scheme of the present application is as follows: a fault diagnosis method for automobile engine bearing based on multi-source heterogeneous data fusion, the fault diagnosis method is trained by a multi-source heterogeneous data training data set, characterized in that it comprises:

[0012] S1, data acquisition and labeling: the present method uses the bearing operation data collected by Anil Kumar and Rajesh Kumar precision measurement laboratory to train the model. The data uses three acceleration sensors to detect the vibration data of the bearing in X, Y and Z directions, and uses a microphone to collect sound data, to simulate multi-source heterogeneous data. The multi-source data used in the experiment contains 11 types of faults, which are divided into early slight fault, slight fault, medium fault and serious fault according to the degree of fault. The present method takes early slight fault for simulation experiment to verify the effectiveness of the method. In actual application, infrared sensor and acoustic emission sensor can be used for continuous collection. The method proposed in the experiment is also applicable.

[0013] S2, build a multi-source heterogeneous data fusion automobile engine bearing fault diagnosis model: the multi-source heterogeneous data fusion automobile engine bearing fault diagnosis model built includes a multi-modal data fusion noise reduction module, a multi-modal data feature extraction module, and a fault diagnosis result output module, wherein the multi-modal data noise reduction module is composed of the multi-source heterogeneous data fusion network proposed in the application, and the multi-modal data fault diagnosis module is composed of the feature enhancement convolutional neural network built in the application;

[0014] S3, the super parameter set by the application: the application adopts an Adam optimizer for network training, a Sigmoid function is selected as a noise reduction neural network activation function, a ReLU function is adopted as an activation function of the fault diagnosis module, a Drop-out is set to 0.12, a learning rate is set to 0.001, sample data includes 1024 sampling points, 32 sampling points are taken as a batch size, and 100 epochs of iterations are performed.

[0015] As a preferred embodiment, the data acquisition and labeling in S1 includes:

[0016] S101, data acquisition:

[0017] Data acquisition is the first step of building an automobile engine bearing fault diagnosis and is also a crucial step, and a large amount of real, multi-source heterogeneous, and automobile engine bearing operation related data containing different health state types need to be collected to ensure that the automobile engine bearing can learn rich fault diagnosis features. The main source of data acquisition is the bearing signal collected in the laboratory;

[0018] S102, data preprocessing:

[0019] (1) Vibration data: considering the data collected in the real running environment of the automobile engine, background vibration caused by road condition changes, uneven road surface, and the interaction between the wheels and the ground will cause background vibration. This vibration will be captured by the acceleration sensor, thereby mixing into the bearing vibration signal to produce noise signals. In order to simulate this signal, the laboratory signal is injected with Gaussian white noise in this paper to simulate real data containing noise.

[0020] (2) Audio data: considering that various external noises such as wind noise, road noise, and other vehicle noise interference will be encountered in real data, these noises will also be captured by the microphone to produce noise signals. The audio data collected in the experiment will be converted into digital signals after being converted by an analog-to-digital converter and stored. Therefore, this signal can also be injected with Gaussian white noise in this paper to simulate real data containing noise.

[0021] S103, data labeling:

[0022] Data labeling is to add meaningful labels to preprocessed data, so that the model can learn the intrinsic characteristics and relationships of the data and perform mathematical calculations;

[0023] Label division: according to the common three faults of bearings, inner ring fault, outer ring fault and ball fault, marked as IR-1, OR-1 and RO-1 respectively. These faults are artificial faults made by electric spark processing;

[0024] S104, data set construction:

[0025] In order to divide the labeled data into training set, validation set and test set, and prepare for the training, evaluation and test of the model, it is necessary to construct the data set: let the labeled data set be

[0026] (D, L), and divide the training set, validation set and test set in the ratio of 8:1:1.

[0027] As a preferred embodiment, the S2 of the application comprises:

[0028] S201, multi-modal data fusion denoising module:

[0029] Due to the characteristics of multi-source heterogeneous data acquisition and the design cost of sensors, the collected data must contain a large amount of noise, which will have a considerable impact on the downstream fault diagnosis task. Therefore, the first step of the method uses the Transformer architecture combined with the SimAM mechanism to build a neural network to perform overall denoising on four-dimensional data, and then uses a convolutional neural network to perform denoising on each dimension separately.

[0030] S202, multi-modal data fault diagnosis module:

[0031] In order to effectively fuse the features of different modalities, the application uses a convolutional neural network to perform two-dimensional convolution operation on four-dimensional features, fuse the coupling information of four-dimensional features, and perform multi-source heterogeneous data fusion for automobile engine bearing fault diagnosis.

[0032] As a preferred embodiment, the S104 data set construction is to divide the labeled data into training set, validation set and test set in order to prepare for the training, evaluation and test of the model, and the data set (D, L) needs to be divided in the ratio of 8:1:1. The specific steps are as follows:

[0033] The labeled data set (D, L) is divided in the following ratio:

[0034] 80% of the data is used as the training set (Training Set)

[0035] 10% of the data is used as a validation set

[0036] 10% of the data is used as a test set

[0037] Such a division can ensure that there is sufficient data for parameter adjustment and optimization during model training, and evaluation of the validation set and final performance verification of the test set after model training, to ensure the generalization ability and effectiveness of the model.

[0038] As a preferred embodiment, when processing four-dimensional signal data (vibration signals in X, Y, Z directions and sound signals), the S201 multi-modal data fusion denoising module may have different noise levels in each dimension due to different acquisition methods. The sound signal collected by the microphone is usually more contaminated by noise than the vibration signal collected by the accelerometer. In order to effectively denoise and improve the accuracy of noise identification, a Transformer layer is introduced to extract the features of the four-dimensional data as a whole. Transformer can capture the complex relationship between these signals through its powerful self-attention mechanism without explicitly constructing complex mathematical models. Combined with the SimAm module for feature enhancement, more robust feature representations can be extracted from multi-dimensional signals, so as to better distinguish between real signals and noise, and finally obtain denoised bearing operation data, improving the accuracy and reliability of fault diagnosis.

[0039] As a preferred embodiment, the Transformer structure mentioned in the S201 multi-modal data fusion denoising module uses a self-attention mechanism to focus on the parts of the vibration signal that should be focused on, and is described by formula (1)

[0040]

[0041] Where Q, K, V represent query matrix, key matrix and value matrix respectively, representing the query information, key information and value information of the current sample; d k to place gradient vanishing or explosion; QK T is the dot product of the query matrix and the key matrix, used to calculate the relevance score, and Softmax is used to convert it into a probability distribution, so as to emphasize the positions with higher importance, and focus on the key parts

[0042] Then, the multi-head attention mechanism is used, so that the model can capture the features of multi-source heterogeneous data from different angles, improving the model's ability to learn complex patterns, which can be described by formula (2):

[0043] MultiHead(Q, K, V) = Concat(head1,..., head h )W O(2)

[0044] where each head i is:

[0045]

[0046] where h denotes the number of attention heads, which can allow the model to learn information in different heterogeneous data spaces in parallel, is the weight matrix of each attention head, and Concat is used to concatenate the outputs of multiple heads, W O can convert the concatenated features back to the appropriate dimension.

[0047] where the determination of whether the denoised data meets the requirements relies on the mean square error loss function, which can be described by formula (4):

[0048]

[0049] where L(θ) represents the loss function, which is used to measure the difference between the denoised data and the real data, θ is the parameter of the neural network, N is the number of training samples, y i represents the i-th noise signal sample, (y i ; θ) represents the residual signal obtained after processing by the neural network, x i represents the signal without noise, and the purpose of the loss function is to minimize the difference between the predicted signal and the real signal, thereby achieving noise reduction.

[0050] The SimAM mentioned in the application is a non-parametric attention mechanism, which can identify the optimal neuron by finding the optimal energy function, which can be described by formula 5:

[0051]

[0052] where e is the energy function, which is used to measure the error between the model output and the actual label, w t , b t are the weight vector and the bias term, respectively, x i represents the input feature, and y represents the label.

[0053] After adopting the above technical scheme, the application has the following beneficial effects:

[0054] 1、The application utilizes the multi-head attention mechanism in the Transformer architecture to couple and denoise multi-source heterogeneous data, and utilizes the self-attention mechanism to focus on the part with the most serious noise in the sample, and cooperates with the SimAM mechanism to realize multi-source heterogeneous data coupling denoising and single sensor information independent denoising, improve the denoising effect of the network, and the method provided by the application is close to the actual automobile running background, and provides a solution for intelligent fault diagnosis of automobile engine bearings.

[0055] 2、The application utilizes the improved wide kernel convolutional neural network to couple feature extraction and classification of multi-source heterogeneous data, greatly expands the sample information, reduces the overfitting risk, and improves the fault diagnosis generalization performance. DETAILED DESCRIPTION

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0057] Figure 1 The figure is a multi-modal data fusion denoising module of the present application;

[0058] Figure 2 The figure is a SimAM architecture of the present application;

[0059] Figure 3 The figure is a Transformer architecture flowchart of the present application;

[0060] Figure 4 The figure is a multi-modal data fault diagnosis of the present application; DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0062] A fault diagnosis method for automobile engine bearings based on multi-source heterogeneous data fusion, the fault diagnosis method is trained by a multi-source heterogeneous data training data set, characterized in that, comprising:

[0063] S1, data acquisition and labeling: the method utilizes the bearing operation data collected by the precision measurement laboratory of Indian Sant Longowal Engineering Technology Institute, Anil Kumar and Rajesh Kumar, for model training. The data utilizes three acceleration sensors to detect the vibration data of the bearing in X, Y and Z directions, and the sound data collected by the microphone, to simulate multi-source heterogeneous data. The multi-source data used in the experiment contains 11 types of faults. According to the different fault degrees, it is divided into early slight fault, slight fault, medium fault and serious fault. The method takes the early slight fault for simulation experiment to verify the effectiveness of the method. In actual application, the infrared sensor and acoustic emission sensor can be used for continuous collection. The method proposed in the experiment is also applicable;

[0064] S2, construction of multi-source heterogeneous data fusion automobile engine bearing fault diagnosis model: the constructed multi-source heterogeneous data fusion automobile engine bearing fault diagnosis model includes a multi-modal data fusion noise reduction module, a multi-modal data feature extraction module and a fault diagnosis result output module. The multi-modal data noise reduction module is composed of the multi-source heterogeneous data fusion network proposed in the application, and the multi-modal data fault diagnosis module is composed of the feature enhancement convolutional neural network built in the application;

[0065] S3, the super parameter set by the application: the application adopts the Adam optimizer for network training, the noise reduction neural network activation function selects the Sigmoid function, the fault diagnosis module adopts the ReLU function as the activation function, the Drop-out is set to 0.12, the learning rate is set to 0.001, the sample data includes 1024 sampling points, and 32 sampling points are taken as the batch size for 100 epoch iterations.

[0066] Further, the data acquisition and labeling in S1 includes:

[0067] S101, data acquisition:

[0068] Data acquisition is the first step of building an automobile engine bearing fault diagnosis, and it is also a crucial step. A large amount of real, multi-source heterogeneous and sufficient different health state type automobile engine bearing operation related data need to be collected to ensure that the automobile engine bearing can learn rich fault diagnosis features. The main source of data acquisition is the bearing signal collected in the laboratory;

[0069] S102, data preprocessing:

[0070] (1) Vibration data: Considering the data collected in the real running environment of the automobile engine, the background vibration caused by the change of road conditions, the uneven road surface, and the interaction between the wheel and the ground will cause background vibration. This vibration will be captured by the acceleration sensor, mixed with the bearing vibration signal, and produce noise signals. In order to simulate this signal, this paper injects Gaussian white noise into the laboratory signal to simulate the real data containing noise;

[0071] (2) Audio data: Considering that in real data, various external noises such as wind noise, road noise, and other vehicle noise interference will be encountered. These noises will also be captured by the microphone to produce noise signals. The audio data collected in the experiment will be converted into digital signals after being converted by the analog-to-digital converter and stored. Therefore, this paper also injects Gaussian white noise into this signal to simulate real data containing noise.

[0072] S103, data labeling:

[0073] Data labeling is to add meaningful labels to the preprocessed data, so that the model can learn the internal features and relationships of the data and perform mathematical calculations.

[0074] Label division: According to the three common faults of bearings, inner ring fault, outer ring fault, and ball fault, they are marked as IR-1, OR-1, and RO-1 respectively. These faults are artificial faults made by electric spark processing. The specific information of the fault is shown in Table 1.

[0075] Table 1 Faults of bearings at a speed of 2050 revolutions per minute and a load of 200N:

[0076]

[0077] S104, data set construction:

[0078] In order to divide the labeled data into training set, validation set and test set, and prepare for model training, evaluation and testing, it is necessary to construct the data set: the labeled data set is denoted as

[0079] (D, L), and the training set, validation set and test set are divided in the ratio of 8:1:1.

[0080] Further, the S2 of the application comprises:

[0081] S201, multi-modal data fusion noise reduction module:

[0082] Due to the characteristics of multi-source heterogeneous data collection and the design cost of sensors, the collected data inevitably contains a large amount of noise, which will have a considerable impact on the downstream fault diagnosis task. Therefore, the first step of the method uses the Transformer architecture to build a neural network with the SimAM mechanism to perform overall noise reduction on four-dimensional data. Then, a convolutional neural network is used to individually reduce noise in each dimension. The structure of the network is shown in the following figure. Figure 1

[0083] S202, a multi-modal data fault diagnosis module,

[0084] In order to effectively fuse the features of different modalities, the application uses a convolutional neural network to perform two-dimensional convolution operations on four-dimensional features, fuse the coupling information of four-dimensional features, and perform multi-source heterogeneous data fusion for automobile engine bearing fault diagnosis. The structure diagram is shown in the following figure. Figure 4

[0085] Among them, the mentioned convolution module adopts a 31*31 convolution kernel, which can greatly improve the precision of fault diagnosis under high noise background, although it will bring certain computational complexity. The experimental results can be shown in Table 2.

[0086] Table 2 Experimental results of different methods using different convolution kernels:

[0087]

[0088] Among them, WDCNN and DRSNCW are two typical noise reduction fault diagnosis neural networks.

[0089] Further, the S104 data set is constructed to divide the labeled data into training set, validation set and test set, so as to prepare for the training, evaluation and testing of the model. The data set (D, L) needs to be divided according to the ratio of 8:1:1. The specific steps are as follows:

[0090] The labeled data set (D, L) is divided according to the following ratio:

[0091] 80% of the data is used as the training set (Training Set)

[0092] 10% of the data is used as the validation set (Validation Set)

[0093] 10% of the data is used as the test set (Test Set)

[0094] Such division can ensure that there is sufficient data for parameter adjustment and optimization during model training, and the validation set is evaluated and the final performance of the test set is verified after model training, to ensure the generalization ability and effectiveness of the model. ​​

[0095] Further, when processing four-dimensional signal data (vibration signals in X, Y, Z directions and sound signals), the S201 multi-modal data fusion noise reduction module may have different noise levels in each dimension due to different collection means. The sound signal collected by the microphone is usually more polluted by noise than the vibration signal collected by the accelerometer. In order to effectively reduce noise and improve the accuracy of noise identification, a Transformer layer is introduced to extract the features of the four-dimensional data as a whole. Transformer can capture the complex relationship between these signals through its powerful self-attention mechanism without explicitly constructing complex mathematical models. Combined with the SimAm module for feature enhancement, more robust feature representations can be extracted from multi-dimensional signals, so as to better distinguish real signals and noise. Finally, the bearing running data after noise reduction is obtained, and the accuracy and reliability of fault diagnosis are improved. The SimAM architecture flowchart is shown in the following figure. Figure 2

[0096] Further, the Transformer structure mentioned in the S201 multi-modal data fusion noise reduction module uses a self-attention mechanism to focus on the parts of the vibration signal that should be focused on. The Transformer architecture can be described by Figure 3 and described by formula (1).

[0097]

[0098] Where Q, K, V represent query matrix, key matrix and value matrix respectively, Q represents query information, K represents key information and V represents value information of the current sample; d k is used to place gradient disappearance or explosion; QK T is the dot product of the query matrix and the key matrix, used to calculate the relevance score, and Softmax is used to convert it into a probability distribution, so as to emphasize the positions with higher importance and focus on the key parts

[0099] Then, the multi-head attention mechanism is used, so that the model can capture the features of multi-source heterogeneous data from different angles and improve the learning ability of the model to complex patterns, which can be described by formula (2):

[0100] MultiHead(Q, K, V) = Concat(head1,..., head h )W O (2)

[0101] Where each head i is:

[0102]

[0103] ​wherein h represents the number of attention heads, which can allow the model to learn information of different heterogeneous data spaces in parallel, is the weight matrix of each attention head, and Concat is to concatenate the outputs of multiple heads, W O The concatenated features can be converted back to the appropriate dimension.

[0104] wherein whether the noise-reduced data meets the requirements is determined by the mean square error loss function, which can be described by formula (4):

[0105]

[0106] wherein L(θ) represents the loss function, which is used to measure the difference between the noise-reduced data and the real data, θ is the parameter of the neural network, N is the number of training samples, y i represents the i-th noise signal sample, (y i ;θ) represents the residual signal obtained after processing by the neural network, x i represents the signal without noise, and the purpose of the loss function is to minimize the difference between the predicted signal and the real signal, thereby achieving noise reduction.

[0107] The SimAM mentioned in the application is a non-parametric attention mechanism, which can identify the optimal neuron by finding the optimal energy function, which can be described by formula 5:

[0108]

[0109] wherein e is the energy function, which is used to measure the error between the model output and the actual label, w t , b t are the weight vector and the bias term, respectively, x i represents the input feature, and y represents the label.

[0110] In the application, the SimAM mechanism is used to explore the local relationship between the real signals in the time series, and this relationship is used to identify and separate the noise signals from the real signals in each dimension, thereby achieving the noise reduction function.

[0111] The application simulates multi-source heterogeneous data by detecting vibration data of the bearing in X, Y and Z directions through three acceleration sensors and collecting sound data by using a microphone, and takes early slight faults to simulate experiments to verify the effectiveness of the method. In actual application, the method proposed in the experiment is also applicable by relying on infrared sensors and acoustic emission sensors to continue to collect. Through multi-source heterogeneous data fusion of automobile engine bearing fault diagnosis, the traditional bearing fault diagnosis method often only uses vibration signals for fault diagnosis. However, due to the running information of the bearing, slight cracks or temperature rise, accurate detection cannot be achieved. However, by using the method, multi-modal data can be used for fault diagnosis, fully utilizing the complex coupling information contained in multi-modal data, improving the accuracy of fault diagnosis, and constructing a multi-source heterogeneous data fusion automobile engine bearing fault diagnosis model, including a multi-modal data fusion noise reduction module, a multi-modal data feature extraction module and a fault diagnosis result output module. Through the multi-source heterogeneous data fusion noise reduction module, due to the characteristics of multi-source heterogeneous data acquisition and sensor configuration problems, the signals collected by the engine bearing during actual operation often contain a large amount of noise signals, which greatly affects the accuracy of fault diagnosis. The application uses the multi-head attention mechanism in the Transformer architecture, cooperates with the SimAM mechanism, and couples and reduces noise of the overall heterogeneous data. At the same time, the independent signals of each sensor are independently detected, which greatly reduces the noise problem in the data. The Adam optimizer is used for network training, the Sigmoid function is selected as the activation function of the noise reduction neural network, the ReLU function is selected as the activation function of the fault diagnosis module, and the improved convolutional neural network is used to extract the overall features of the multi-source heterogeneous data through the multi-source heterogeneous data fusion fault diagnosis module. The wide kernel convolution has the advantage of large receptive field, and the overall modeling of multi-source heterogeneous data is realized, which greatly improves the fault diagnosis accuracy of multi-source heterogeneous data fusion in a strong noise environment.

[0112] The above is only a preferred embodiment of the application and is not intended to limit the application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A method for automobile engine bearing fault diagnosis based on multi-source heterogeneous data fusion, wherein the fault diagnosis method is trained with a multi-source heterogeneous data training dataset, and is characterized in that: include: S1. Data collection and labeling: This method uses the bearing operation data collected by the Anil Kumar and Rajesh Kumar Precision Metrology Laboratory to train the model. The data uses three accelerometers to detect the vibration data of the bearing in the X, Y, and Z directions respectively, and the sound data collected by the microphone to simulate multi-source heterogeneous data. The multi-source data used in the experiment contains 11 types of faults, which are divided into early minor faults, minor faults, medium faults, and severe faults according to the degree of fault. This method uses early minor faults for simulation experiments to verify the effectiveness of the method. In actual applications, infrared sensors and acoustic emission sensors can be used to continue collecting data, and the method proposed in the experiment is also applicable. S2. Constructing a multi-source heterogeneous data fusion automobile engine bearing fault diagnosis model: The constructed multi-source heterogeneous data fusion automobile engine bearing fault diagnosis model includes a multimodal data fusion noise reduction module, a multimodal data feature extraction module, and a fault diagnosis result output module. The multimodal data noise reduction module adopts the multi-source heterogeneous data fusion network proposed in the present invention, and the multimodal data fault diagnosis module adopts the feature enhancement convolutional neural network constructed in the present invention. S3. Hyperparameters set by the invention: The present invention uses the Adam optimizer for network training, the Sigmoid function is selected as the activation function of the denoising neural network, the fault diagnosis module uses the ReLU function as the activation function, the Drop-out is set to 0.12, the learning rate is set to 0.001, the sample data includes 1024 sampling points, and 32 sampling points are used as the batch size, and 100 epochs are iterated.

2. The automobile engine bearing fault diagnosis method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: The data collection and marking in S1 include: S101, Data Collection: Data collection is the first and most crucial step in establishing automotive engine bearing fault diagnosis. It requires collecting a large amount of real, multi-source, heterogeneous data related to automotive engine bearing operation, containing sufficient data in different health states, to ensure that automotive engine bearings can learn rich fault diagnosis features. The main source of data collection is bearing signals collected in the laboratory. S102, data preprocessing: (1) Vibration data: Considering that the data collected by the automobile engine in the real operating environment will cause background vibration due to changes in road conditions, uneven road surface, and background vibration caused by the interaction between the wheels and the ground, this vibration will be captured by the acceleration sensor and mixed with the vibration signal of the bearing to generate a noise signal. In order to simulate this signal, this paper injects Gaussian white noise into the laboratory signal to simulate the real data containing noise; (2) Audio data: Considering that real data may encounter various external noises, such as wind noise, road noise, and other vehicle noise interference, these noises will also be captured by the microphone to generate noise signals. The audio data collected in the experiment will be converted into digital signals after being converted by an analog-to-digital converter and stored. Therefore, this paper can also inject Gaussian white noise into this signal to simulate real data containing noise; S103, data tag: Data labeling is to add meaningful labels to preprocessed data so that the model can learn the intrinsic characteristics and relationships of the data and perform mathematical calculations; Label classification: According to the three common bearing faults, inner ring fault, outer ring fault, and ball fault, they are marked as IR-1, OR-1, and RO-1 respectively. These faults are all artificial faults created by EDM; S104. Dataset construction: In order to divide the labeled data into training set, validation set and test set, and prepare for model training, evaluation and testing, it is necessary to construct a dataset: the labeled dataset is recorded as (D, L), and the training set, validation set and test set are divided into 8:1:1 ratio.

3. The automobile engine bearing fault diagnosis method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: The automobile engine bearing fault diagnosis model constructed by multi-source heterogeneous data fusion in S2 includes: S201, multimodal data fusion noise reduction module: Due to the characteristics of multi-source heterogeneous data collection and the design cost of sensors, the collected data inevitably contains a large amount of noise, which will have a significant impact on downstream fault diagnosis tasks. Therefore, in the first step of this method, a neural network is built using the Transformer architecture in conjunction with the SimAM mechanism to perform overall noise reduction on the four-dimensional data. Then, a convolutional neural network is used to perform noise reduction on each dimension separately. S202, multimodal data fault diagnosis module: In order to effectively fuse the features of different modes, the present invention uses convolutional neural networks to perform two-dimensional convolution operations on four-dimensional features, fuse the coupling information of four-dimensional features, and perform automobile engine bearing fault diagnosis by multi-source heterogeneous data fusion.

4. The automobile engine bearing fault diagnosis method based on multi-source heterogeneous data fusion according to claim 2 is characterized in that: The S104 dataset construction is to divide the labeled data into training set, validation set and test set to prepare for model training, evaluation and testing. The dataset (D, L) needs to be divided in a ratio of 8:1:

1. The specific steps are as follows: Divide the labeled dataset (D, L) into the following proportions: 80% of the data is used as training set 10% of the data is used as a validation set 10% of the data is used as a test set (Test Set) Such a division can ensure that there is sufficient data for parameter adjustment and optimization during the model training process, and perform evaluation of the validation set and final performance verification of the test set after model training to ensure the generalization ability and effectiveness of the model.

5. The automobile engine bearing fault diagnosis method based on multi-source heterogeneous data fusion according to claim 3 is characterized in that: When the S201 multimodal data fusion noise reduction module processes four-dimensional signal data (vibration signals and sound signals in the X, Y, and Z directions), the noise levels in each dimension may be significantly different due to different acquisition methods. The sound signals collected by the microphone are usually more subject to noise pollution than the vibration signals collected by the accelerometer. In order to effectively reduce noise and improve the accuracy of noise recognition, the Transformer layer is introduced to extract the features of the four-dimensional data as a whole. Through its powerful self-attention mechanism, the Transformer can capture the complex relationship between these signals without explicitly constructing a complex mathematical model. Combined with the SimAm module for feature enhancement, it can extract more robust feature representations from multidimensional signals, thereby better distinguishing between real signals and noise, and finally obtaining the denoised bearing operation data, thereby improving the accuracy and reliability of fault diagnosis.

6. The automobile engine bearing fault diagnosis method based on multi-source heterogeneous data fusion according to claim 3 is characterized in that: The Transformer structure mentioned in the S201 multimodal data fusion denoising module uses the self-attention mechanism to focus on the part of the vibration signal that should be paid attention to, and is described by formula (1) Among them, Q, K, and V represent the query matrix, key matrix, and value matrix, respectively, indicating the query information, key information, and value information of the current sample; d k Used to prevent gradient disappearance or explosion; QK T The dot product of the query matrix and the key matrix is ​​used to calculate the relevance score, and Softmax converts it into a probability distribution, thereby emphasizing the more important positions and focusing on the key parts. Afterwards, the multi-head attention mechanism is used to enable the model to capture the features of multi-source heterogeneous data from different perspectives, improving the model's ability to learn complex patterns, which can be described by formula (2): MultiHead(Q,K,V)=Concat(head1,...,head h )W O (2) Each of the heads i for: Where h represents the number of attention heads, which allows the model to learn information from different heterogeneous data spaces in parallel. is the weight matrix of each attention head, and Concat concatenates the outputs of multiple heads. O The concatenated features can be converted back to the appropriate dimension; Among them, whether the denoised data meets the requirements depends on the mean square error loss function, which can be described by formula (4): Among them, L(θ) represents the loss function, which is used to measure the difference between the denoised data and the real data, θ is the parameter of the neural network, N is the number of training samples, and y i represents the i-th noise signal sample, (y i ; θ) represents the residual signal obtained after neural network processing, x i Represents a signal without noise. The purpose of this loss function is to minimize the gap between the predicted signal and the true signal, thereby reducing the noise. The SimAM mentioned in this invention is a parameter-free attention mechanism that can identify the optimal neuron by finding the optimal energy function, which can be described by Formula 5: Among them, e is the energy function, which is used to measure the error between the model output and the actual label, w t , b t are weight vector and bias term respectively, x i represents the input feature, and y represents the label.

Images