Automobile fault prediction method based on multi-attention mechanism

Abstract

The invention provides an automobile fault prediction method based on a multi-attention mechanism, and belongs to the technical field of automobile state monitoring, and the method comprises the steps: firstly collecting the historical fault data of an automobile, including the fault information in each maintenance record; and then, a global fault co-occurrence graph is constructed based on the data, and a graph attention mechanism is utilized to carry out fusion representation on the correlation between the faults, so that a fault correlation fusion representation result is obtained. A key fault information learning module is constructed in combination with a long short-term memory network and an attention mechanism, key fault information is extracted from weighted historical fault information, and a key fault information learning result is generated. And finally, comprehensively fusing the fault correlation fusion representation result and the key fault information learning result, and performing enhanced integration through an attention mechanism to obtain an automobile fault prediction result. According to the invention, the complex relationship between faults can be effectively captured, and the possible faults in the future can be accurately predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of automobile state monitoring, and in particular relates to an automobile fault prediction method based on a multi-attention mechanism. Background Art

[0002] Automobile fault prediction is of great significance in the modern automobile industry and service system. Fault prediction can greatly improve the safety and reliability of vehicles. By identifying potential problems in advance, it can prevent major failures and safety hazards that may occur, and protect the safety of users. In addition, fault prediction helps reduce maintenance costs. Traditional post-fault repairs often result in high costs. Through predictive maintenance measures, timely intervention can be carried out at the first sign of a fault, avoiding more serious damage and more expensive repair costs. Fault prediction can also increase the service life of the vehicle. Through scientific maintenance and care strategies, the aging and wear of automobile parts are delayed, thereby extending the service life of the entire vehicle.

[0003] In recent years, the research on fault prediction of automotive electronics and safety systems has received extensive attention and has made significant progress in many fields. A machine learning algorithm for truck compressor fault prediction using recorded vehicle data is proposed in the prior art. This study focuses on the deficiencies of standard attribute value knowledge representation and proposes a new rule induction algorithm based on Michalski's classic AQ method, which can consider the composite performance of each truck, allowing greater flexibility in predicting faults. There is also a data mining workflow for predictive maintenance of heavy truck pneumatic systems. The method includes feature selection, meta-feature engineering, bias sampling, and boosted tree models. Experiments show that meta-feature engineering and bias sampling are crucial to improving classifier performance. There is also an application of dynamic weighted loss functions in sensor-based prognosis and health management in the prior art. By using dynamic weighted loss functions in deep learning models, they have achieved significant improvements in prediction and fault detection rates. There is also a fault diagnosis framework for autonomous vehicles based on hybrid data analysis methods and fuzzy PID control. This study uses discrete wavelet transform for denoising and feature extraction, and uses an extreme learning machine-based autoencoder for anomaly detection. There is also support vector machine technology to achieve efficient fault detection and identification, and the effectiveness of the architecture is verified by MATLAB / IPG CarMaker. There is also a detection, identification, isolation and prediction architecture for multiple faults in multi-sensor systems such as self-driving cars. Using two deep neural network architectures, this method achieves efficient fault management in complex systems. Someone else proposed a high-precision health prediction model based on improved correlation vector machine integrated regression for sensor systems. By introducing uncertain deep soft sensors for performance variable extraction, this study significantly improved the prediction accuracy. The existing technology also introduces data-driven strategies for predictive maintenance, focusing on the prediction of clogging of oxygen sensors on diesel engines, and proposed the PREPIPE pipeline, which is comparable to advanced deep learning methods while maintaining interpretability.

[0004] Although existing studies have demonstrated the ability to accurately predict failures of specific vehicle components, these studies have focused on fault diagnosis and prediction of a single system or component and are insufficient in providing overall vehicle failure prediction. Summary of the invention

[0005] The present invention proposes a vehicle fault prediction method based on a multi-attention mechanism to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above object, the present invention provides a method for predicting automobile faults based on a multi-attention mechanism, comprising the following steps:

[0007] Collecting historical fault data of the vehicle, wherein the historical fault data includes fault information in each maintenance record;

[0008] Build a CoMAM model based on historical fault data and use the CoMAM model to predict faults;

[0009] Fault prediction using the CoMAM model includes the following steps:

[0010] A global fault co-occurrence graph is constructed based on historical fault data, and the correlation between faults is fused and represented using the graph attention mechanism to obtain the fault correlation fusion representation result.

[0011] A key fault information learning module is constructed based on the long short-term memory network and the attention mechanism. The key fault information learning module extracts key fault information from the historical fault information weighted by the graph attention mechanism to obtain the key fault information learning result.

[0012] The fault correlation fusion representation result and the key fault information learning result are comprehensively fused, and enhanced and integrated through the attention mechanism to obtain the automobile fault prediction result.

[0013] Preferably, constructing a global fault co-occurrence map includes:

[0014] The fault code pairs in each maintenance record in the historical fault data of the automobile are taken as nodes;

[0015] When a fault code pair appears simultaneously in the maintenance record of a vehicle, two identical weights are added to the global fault co-occurrence graph;

[0016] Calculate the co-occurrence frequency of fault code pairs in the maintenance records of all cars, and further calculate the edge weights based on the co-occurrence frequency.

[0017] Preferably, obtaining the fault correlation fusion representation result includes:

[0018] Based on the global project co-occurrence graph, a fault co-occurrence matrix is ​​constructed. Based on the fault co-occurrence matrix, the graph attention mechanism and the single-head attention mechanism are used to realize the fusion of the co-occurrence matrix and the successive historical fault information.

[0019] The fault correlation fusion representation result is obtained through the Stack method.

[0020] Preferably, obtaining key fault information learning results includes:

[0021] Use long short-term memory networks to process historical fault information, obtain the hidden state of fault records, and capture the long-term dependencies of historical fault sequences;

[0022] The output of the long short-term memory network is processed through the attention mechanism to strengthen the interaction between historical fault records and obtain key information that affects future faults.

[0023] Preferably, the process of comprehensive fusion prediction includes:

[0024] The fault correlation fusion representation result and the key fault information learning result are fused through the attention mechanism to obtain the fusion result;

[0025] Future faults are predicted based on the fusion results, and a dropout operation is performed on the fusion results before prediction.

[0026] Preferably, the method further comprises optimizing the CoMAM model, and the optimization step comprises:

[0027] Use the binary cross entropy loss function as the global objective function to train the model;

[0028] During the training process, model performance is optimized by adjusting hyperparameters and activation functions.

[0029] The present invention also proposes a computer device, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0030] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method are implemented.

[0031] The present invention also proposes a computer program product, comprising a computer program, which implements the steps of the method when executed by a processor.

[0032] Compared with the prior art, the present invention has the following advantages and technical effects:

[0033] The present invention proposes a method for predicting automobile faults based on a multi-attention mechanism. The method first collects historical fault data of automobiles, including fault information in each maintenance record. Subsequently, a global fault co-occurrence graph is constructed based on these data, and the correlation between faults is fused and represented using the graph attention mechanism, thereby obtaining a fault correlation fusion representation result. Furthermore, the present invention combines the long short-term memory network and the attention mechanism to construct a key fault information learning module, extracts key fault information from the weighted historical fault information, and generates a key fault information learning result. Finally, the fault correlation fusion representation result and the key fault information learning result are comprehensively fused, and enhanced and integrated through the attention mechanism to obtain the automobile fault prediction result. The present invention can effectively capture the complex relationship between faults and accurately predict possible future faults, thereby providing strong support for preventive maintenance of automobiles. The present invention aims to predict comprehensive faults of vehicles. This comprehensive prediction method can significantly improve the safety and availability of vehicles, ensure improved operating efficiency and extended service life. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 It is a model framework diagram of an embodiment of the present invention;

[0036] Figure 2 Parameter diagram for sensitivity analysis of dropout rate parameters according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0038] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] Embodiment 1

[0040] This embodiment provides a method for predicting automobile faults based on a multi-attention mechanism, comprising the following steps:

[0041] Collecting historical fault data of the vehicle, wherein the historical fault data includes fault information in each maintenance record;

[0042] Build a CoMAM model based on historical fault data and use the CoMAM model to predict faults;

[0043] Fault prediction using the CoMAM model includes the following steps:

[0044] A global fault co-occurrence graph is constructed based on historical fault data, and the correlation between faults is fused and represented using the graph attention mechanism to obtain the fault correlation fusion representation result.

[0045] A key fault information learning module is constructed based on the long short-term memory network and the attention mechanism. The key fault information learning module extracts key fault information from the historical fault information weighted by the graph attention mechanism to obtain the key fault information learning result.

[0046] The fault correlation fusion representation result and the key fault information learning result are comprehensively fused, and enhanced and integrated through the attention mechanism to obtain the automobile fault prediction result.

[0047] Furthermore, the model framework is as follows:

[0048] Figure 1 The proposed Collaborative Multiple Attention Mechanisms for Automotive Fault Prediction (CoMAM) framework for automobile maintenance demand prediction is given in the paper. Specifically, the framework consists of three modules:

[0049] Perform correlation fusion representation on historical faults. This embodiment proposes a fault correlation representation learning module based on a graph attention mechanism, which uses the graph attention mechanism to fuse the correlation features of faults.

[0050] Key fault information learning. This embodiment proposes a key fault learning module that combines a long short-term memory network with an attention mechanism to learn key fault information from weighted historical fault information.

[0051] Comprehensive fusion prediction. The correlation information fusion representation learning results and key timing information learning are fused using the attention mechanism, and the comprehensive fusion results are then used to predict future faults.

[0052] Furthermore, the correlation fusion is expressed as follows:

[0053] All the fault combinations in the process of automobile fault development form a global item co-occurrence map ,This structure reflects the pattern of correlation between different faults of the car. The node is Failure in If a fault code is If both appear in the maintenance record of a vehicle, Add two equal weights and , then calculate Frequency of occurrence in the maintenance records of all vehicles It is used to further calculate the weight of the edge. One item may be a common combination of another item, but not vice versa. In addition, this embodiment also wants to detect important and common item pairs. , get the combination .make for The total frequency of this embodiment uses the co-occurrence matrix To express , The definition of is shown in formula (1).

[0054] (1)

[0055] Based on the above co-occurrence matrix, the graph attention mechanism and the single-head attention mechanism are used to realize the fusion of the co-occurrence matrix and the successive historical fault information. The specific process is shown in formula (2).

[0056] (2)

[0057] The attention mechanism in the above formula is defined as shown in formula (3):

[0058] (3)

[0059] in, is the dimension of attention, is the attention weight.

[0060] Then the Stack method is used to obtain the fault phase fusion representation result. .

[0061] Furthermore, the key fault information is learned as follows:

[0062] In the process of automobile failure development, some critical failures will have a significant impact on the future development of failures. In order to more effectively capture these critical information in historical failure records, this embodiment proposes a critical failure information learning method that combines a long short-term memory network (LSTM) with an attention mechanism. First, use LSTM to process historical failure information and obtain the hidden state of the failure record. , as shown in formula (4).

[0063] (4)

[0064] Through this process, The previous and cumulative information in the continuous fault records are learned, capturing the long-term dependencies of the historical fault sequences. It is processed through the attention mechanism, as shown in formula (5).

[0065] (5)

[0066] in ,Through this process, the interaction between the historical fault records is enhanced, Critical information that affects future failures is obtained.

[0067] Then the comprehensive result of key fault information learning is obtained by summing , as shown in formula (6).

[0068] (6)

[0069] Furthermore, the comprehensive fusion prediction is as follows:

[0070] In order to fully integrate the correlation fusion representation results and the key fault information learning comprehensive results to predict future faults, this embodiment uses the attention mechanism to achieve enhanced integration, thereby obtaining the fusion result , the specific process is shown in formula (7).

[0071] (7)

[0072] No. The failure is based on the correlation perception fusion result The specific implementation process is shown in formula (8).

[0073] (8)

[0074] in, and is a parameter, yes In order to improve the robustness of the model, A dropout operation is performed before making predictions.

[0075] Furthermore, the model is optimized as follows:

[0076] This example trains the CoMAM model to predict the first For each failure, the global objective function is the binary cross entropy loss function, as shown in formula (9).

[0077] (9)

[0078] in, Indicates a fault The prediction results, Indicates a fault The real label.

[0079] The following experiments were also conducted in this embodiment:

[0080] In order to evaluate the performance of the method proposed in this embodiment, this embodiment uses real automobile maintenance data from 30 automobile maintenance companies for verification. After screening, the data of cars with maintenance times greater than or equal to two times and the fault information of each maintenance are retained. The data set includes the fault records of 7932 cars maintained from April 2011 to December 2020. The detailed data set statistics are shown in Table 1.

[0081] Table 1

[0082] Number of cars 7932 Maximum maintenance times 45 Average maintenance times 3.89 Maintenance code quantity 2240 Maximum number of maintenance codes per time 51 Average number of maintenance codes per time 5.19

[0083] In order to enhance the experimental process, this embodiment randomly divides the data set into a training set and a test set. Specifically, these data consist of 5500 and 2432 car failures, respectively. In the method of this embodiment, this embodiment specifies the failure of the last maintenance record as a label, and uses the remaining historical maintenance records as input features.

[0084] Benchmark model and evaluation indicators: The main task of the experiment is to The failure prediction in this maintenance record This is a multi-label classification problem. For this task, the evaluation metric is weighted Fraction( )and , The score is all faults The weighted sum of scores, It is before The average ratio of predicted faults to the total number of real faults. It measures the accuracy of prediction. In order to compare the method proposed in this embodiment with the most advanced models, this embodiment selects the following methods as comparative experiments. Including MLP, RNN, LSTM, GRU and Transformer. In addition, it is also compared with Dipole, a typical method in the field of medical diagnosis prediction.

[0085] In the experiment of this embodiment, this embodiment randomly initializes the model parameters. Hyperparameters and activation functions are tuned on the validation set. When training the model of this embodiment, this embodiment uses 100 epochs and the Adam optimizer. The learning rate is set to 0.001. All programs are implemented using Python 3.10.0 and PyTorch 1.10.0, CUDA 11.4 on a device with 64GB memory and NVIDIA-SMI 472.39 GPU. To ensure robustness, this embodiment repeats the experiment 5 times using different random seeds.

[0086] The model experiment results are shown in Table 2. Since the average number of failures is 5.20, this embodiment sets k=[5,10,15,20,25,30,35] as The experimental results show that the performance of the CoMAM model proposed in this example is significantly better than all baseline models. From the above, we can see that the CoMAM model performs well in both precision and recall. Compared with the best performing baseline model Dipole, CoMAM improves the F1 score by 1.03%, which shows that the CoMAM model has better classification balance ability and can identify faults to the maximum extent while maintaining high accuracy, thus effectively reducing the false alarm rate. From the perspective of k, the CoMAM model always maintains its performance advantage as the k value changes. This shows that the CoMAM model can accurately predict the required faults in both short and long prediction lists, showing strong generalization ability. This performance stability is crucial for practical applications, especially in highly dynamic and complex fault environments.

[0087] Table 2

[0088] Model <![CDATA[F1]]> R@5 R@10 R@15 R@20 R@25 R@30 R@35 MLP 31.57+0.42 52.64+0.15 57.65+0.11 62.57+0.13 66.51+0.19 69.73+0.14 70.81+0.10 72.23+0.18 RNN 32.31+0.66 52.94+0.09 58.37+0.20 63.04+0.17 67.34+0.16 70.23+0.15 72.33+0.12 73.53+0.10 LSTM 32.09+0.80 53.30+0.18 57.58+0.17 62.38+0.11 66.82+0.14 70.25+0.16 73.82+0.19 75.66+0.15 GRU 32.20+0.43 53.11+0.20 57.70+0.12 62.89+0.14 67.44+0.13 70.46+0.11 72.76+0.17 74.16+0.09 Dipole 32.59+0.31 52.88+0.21 57.73+0.16 62.66+0.10 67.45+0.18 70.61+0.15 73.66+0.12 75.49+0.13 Transformer 31.63+0.19 52.74+0.12 57.37+0.14 62.64+0.20 66.71+0.16 70.03+0.13 71.34+0.19 73.23+0.11 CoMAM 33.62+0.47 54.38+0.58 60.62+1.18 65.87+1.02 69.66+0.89 73.11+0.37 78.16+0.13 78.16+0.23

[0089] In order to study the impact of CoMAM's key parameters on its performance, this example evaluates the sensitivity of the dropout rate. In this example, the dropout rate is set in the range of 0.3 to 0.7. Figure 2 The results shown in Figure 3 show that the model works best when the dropout rate is 0.5, and the performance remains relatively stable.

[0090] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A vehicle fault prediction method based on multi-attention mechanism, characterized in that: The following steps are involved: Collecting historical fault data of the vehicle, wherein the historical fault data includes fault information in each maintenance record; Build a CoMAM model based on historical fault data and use the CoMAM model to predict faults; Fault prediction using the CoMAM model includes the following steps: A global fault co-occurrence graph is constructed based on historical fault data, and the correlation between faults is fused and represented using the graph attention mechanism to obtain the fault correlation fusion representation result. A key fault information learning module is constructed based on the long short-term memory network and the attention mechanism. The key fault information learning module extracts key fault information from the historical fault information weighted by the graph attention mechanism to obtain the key fault information learning result. The fault correlation fusion representation result and the key fault information learning result are comprehensively fused, and enhanced and integrated through the attention mechanism to obtain the automobile fault prediction result.

2. The method according to claim 1, characterized in that Constructing a global fault co-occurrence map includes: The fault code pairs in each maintenance record in the historical fault data of the automobile are taken as nodes; When a fault code pair appears simultaneously in the maintenance record of a vehicle, two identical weights are added to the global fault co-occurrence graph; Calculate the co-occurrence frequency of fault code pairs in the maintenance records of all cars, and further calculate the edge weights based on the co-occurrence frequency.

3. The method according to claim 1, characterized in that Obtaining the fault correlation fusion representation results includes: Based on the global project co-occurrence graph, a fault co-occurrence matrix is ​​constructed. Based on the fault co-occurrence matrix, the graph attention mechanism and the single-head attention mechanism are used to realize the fusion of the co-occurrence matrix and the successive historical fault information. The fault correlation fusion representation result is obtained through the Stack method.

4. The method according to claim 1, characterized in that Obtain key fault information and learning results include: Use long short-term memory networks to process historical fault information, obtain the hidden state of fault records, and capture the long-term dependencies of historical fault sequences; The output of the long short-term memory network is processed through the attention mechanism to strengthen the interaction between historical fault records and obtain key information that affects future faults.

5. The method according to claim 1, characterized in that The process of comprehensive fusion prediction includes: The fault correlation fusion representation result and the key fault information learning result are fused through the attention mechanism to obtain the fusion result; Future faults are predicted based on the fusion results, and a dropout operation is performed on the fusion results before prediction.

6. The method according to claim 1, characterized in that The method further comprises optimizing the CoMAM model, wherein the optimization step comprises: Use the binary cross entropy loss function as the global objective function to train the model; During the training process, model performance is optimized by adjusting hyperparameters and activation functions.

7. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images