A prediction method for automobile maintenance items based on deep learning
By applying deep learning technology in the field of automotive maintenance, combined with LSTM and attention mechanism, a model for predicting automotive maintenance projects was built, which solved the shortcomings in the prevention of failures in traditional maintenance methods and significantly improved the prediction accuracy.
Patent Information
- Application Number
- CN202411677901.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-22
- Publication Date
- 2025-05-09
- Estimated Expiration
- 2044-11-22
AI Technical Summary
Traditional car maintenance methods rely on regular inspections and drivers’ subjective judgments, and cannot accurately and timely prevent failures, resulting in increased car downtime and increased maintenance costs.
A deep learning-based automobile repair project prediction method is adopted to build a car repair project prediction model by combining long-term short-term memory network (LSTM) and attention mechanisms to capture long-term dependencies in maintenance records and emphasize key historical information.
The ability to extract key features from historical sequence data is improved, the generalization ability and robustness of the model is enhanced, and the prediction accuracy of automobile repair projects is significantly improved.
Smart Images
Figure CN119180424B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of automobile maintenance technology, and in particular to an automobile maintenance project prediction method based on deep learning. Background Art
[0002] With the rapid growth of the number of cars in the world, cars play an increasingly important role in people's lives. Therefore, the car maintenance industry, which is closely related to cars, has become an important field that affects people's quality of life. In recent years, deep learning technology has made breakthrough progress in many fields such as image recognition, natural language processing, and complex strategy games. Among them, the application of deep learning in the automotive industry has attracted much attention, especially in car maintenance. As the automotive industry enters a new era of data-driven and intelligent services, deep learning technology is used to predict car maintenance;
[0003] As the automotive industry enters a new era of data-driven and intelligent services, using deep learning technology to predict automobile repair and maintenance has become a research focus.
[0004] Traditional car maintenance mainly relies on regular inspections and the driver's subjective judgment, but this method is often not accurate and timely enough to effectively prevent failures. As a result, this leads to increased car downtime and rising maintenance costs. In addition, with the increasing complexity of automotive systems, traditional prediction methods are difficult to accurately predict possible problems in the future. Summary of the invention
[0005] The purpose of this invention is to propose a deep learning-based automobile maintenance project prediction method to solve the problems of the above-mentioned prior art, capture the long-term dependencies in the maintenance records through LSTM, and emphasize key historical information through the attention mechanism. Dropout technology and customized activation functions are combined to enhance the generalization ability and robustness of the model.
[0006] To achieve the above object, the present invention provides the following solutions:
[0007] A method for predicting automobile maintenance items based on deep learning, comprising:
[0008] Obtain automobile maintenance data and build a data set;
[0009] Using the key information learning strategy, the long short-term memory network is combined with the attention mechanism to build the original prediction model;
[0010] Using the data set to train the original prediction model to obtain a prediction model for automobile maintenance items;
[0011] The automobile maintenance item prediction model is used to predict all maintenance items of the automobile.
[0012] Optionally, the automobile maintenance data is vehicle data with maintenance times greater than or equal to two times, and the automobile maintenance data includes: number of vehicles, maximum number of maintenance times, average number of maintenance times, number of maintenance item codes, maximum number of item codes for each maintenance, and average number of item codes for each maintenance.
[0013] Optionally, the attention mechanism is:
[0014]
[0015] Among them, Q is the query vector, K is the key vector, V is the value vector, and W q is the weight of the query vector, W k is the weight of the key vector, T is the matrix transpose operation, W v is the weight of the value vector, and d is the attention dimension.
[0016] Optionally, using the data set to train the original prediction model includes:
[0017] The last maintenance record in the data set is used as a label, and the remaining maintenance records are used as input features to train the original prediction model, so that the original prediction model can predict the next maintenance item based on the maintenance items of the previous T maintenances.
[0018] Optionally, using the data set to train the original prediction model further comprises: using a binary cross entropy loss function as a global objective function;
[0019] The binary cross entropy loss function is:
[0020]
[0021] Where L is the error result, is the type of maintenance item, i is the code number of the maintenance item, for Perform a transpose operation, is the prediction result of the i-th maintenance item, is the actual result of the i-th maintenance project.
[0022] In the original prediction model, the long-term dependencies in the maintenance records are captured through the long short-term memory network, so that the model can learn the long-term rules in the development process of maintenance projects, and emphasize key historical information through the attention mechanism, so that the model can learn the impact of historical key maintenance projects on future maintenance projects, thereby improving the prediction accuracy of the model.
[0023] The beneficial effects of the present invention are:
[0024] In order to improve the ability to extract key features from historical sequence data, the present invention introduces an attention mechanism that enables the model to automatically prioritize important historical maintenance records. In order to improve the generalization and robustness of the prediction model, a customized classifier is developed by combining hidden layer output and dropout technology, while the activation function enhances the nonlinear representation ability of the model. Comprehensive experiments using real-world data show that the model of the present invention has excellent performance compared with state-of-the-art benchmarks. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0026] Figure 1 This is a schematic diagram of the LSTM-Att model structure of an embodiment of the present invention;
[0027] Figure 2 This is a schematic diagram of the LSTM model structure of an embodiment of the present invention;
[0028] Figure 3 A schematic diagram of predicted w-F1 according to an embodiment of the present invention;
[0029] Figure 4 This is a schematic diagram of predicted R@5 according to an embodiment of the present invention;
[0030] Figure 5 This is a schematic diagram of predicted R@10 according to an embodiment of the present invention;
[0031] Figure 6 This is a schematic diagram of predicted R@15 according to an embodiment of the present invention;
[0032] Figure 7 This is a schematic diagram of predicted R@20 according to an embodiment of the present invention;
[0033] Figure 8 This is a schematic diagram of predicted R@25 according to an embodiment of the present invention;
[0034] Fig. 9 This is a schematic diagram of the predicted value R@30 of an embodiment of the present invention;
[0035] Fig.10 Schematic diagram of predicted value R@35 of an embodiment of the present invention. DETAILED DESCRIPTION
[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0037] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0038] This embodiment proposes a method for predicting automobile maintenance items based on deep learning, including:
[0039] Obtain automobile maintenance data and build a data set;
[0040] The original prediction model is trained using the data set to obtain a prediction model for automobile maintenance items;
[0041] Use the automobile maintenance item prediction model to predict all automobile maintenance items.
[0042] Furthermore, the automobile maintenance data is vehicle data with maintenance times greater than or equal to two times, and the automobile maintenance data includes: number of vehicles, maximum number of maintenance times, average number of maintenance times, number of maintenance item codes, maximum number of item codes for each maintenance, and average number of item codes for each maintenance.
[0043] Furthermore, using the data set to train the original prediction model includes:
[0044] The last maintenance record in the data set is used as a label, and the remaining maintenance records are used as input features to train the original prediction model, so that the original prediction model can predict the next maintenance item based on the maintenance items of the previous T maintenances.
[0045] Furthermore, using the data set to train the original prediction model also includes: using a binary cross entropy loss function as a global objective function.
[0046] In this example, real car maintenance data from 40 car maintenance companies are used to verify it. After screening, vehicle data with maintenance times greater than or equal to two are selected. The dataset includes records of 10,256 cars maintained from April 2011 to April 2023. The detailed statistics of the dataset are shown in Table 1.
[0047] Table 1: Experimental dataset statistics
[0048]
[0049] Randomly divide into training set, validation set and test set. The training set includes 4,000 vehicles, the validation set includes 5,000 vehicles, and the test set includes 5,000 vehicles. The last maintenance record is used as the label, and the rest of the maintenance records are used as input features.
[0050] Using the key information learning strategy, the long short-term memory network is combined with the attention mechanism to build the original prediction model;
[0051] In the field of vehicle maintenance, specific key items can have a significant impact on future maintenance activities. In order to more effectively capture key information in historical maintenance records, this embodiment proposes a key information learning strategy that combines a long short-term memory network (LSTM) with an attention mechanism. The structure of the proposed LSTM-Attention (LSTM-Att) learning for vehicle maintenance item prediction is presented in Figure 1 middle.
[0052] Initially, the history maintenance items are processed using LSTM to derive hidden states R1, R2, ..., R T Recent maintenance records, as shown in Equation 1.
[0053] R1, R2, . . . , R T = LSTM(M1, M2, . . . , M T )(1)
[0054] The equation for the LSTM neural network calculation process here is as follows:
[0055] (2)
[0056] The structural diagram of the LSTM neural network is as follows Figure 2 shown.
[0057] According to this procedure, R1, R2, ..., R T Encapsulates the raw and accumulated data from continuous maintenance records to capture long-term correlations in historical maintenance sequences. T Use the attention mechanism for processing. The specific process is detailed in the following formula:
[0058] B1, B2, . . . , B T = Atten (R, R, R) (3)
[0059] Here, R = R1, R2, ..., R T Through this process, the interaction between historical maintenance records B1, B2, ..., B TCapture critical temporal information - options that influence future maintenance projects.
[0060] The implementation of the above attention mechanism is shown as follows:
[0061] (4)
[0062] Among them, Q is the query vector, K is the key vector, V is the value vector, and W q is the weight of the query vector, W k is the weight of the key vector, T is the matrix transpose operation, W v is the weight of the value vector, d is the attention dimension;
[0063] The method of emphasizing key historical information through the attention mechanism is to first use the long short-term memory network (LSTM) to process the historical maintenance records to extract the hidden states in these records. These hidden states represent the maintenance information at different time points. Subsequently, the attention mechanism is applied, which calculates the importance weight of each hidden state and uses these weights to highlight the key historical information. The specific implementation of the attention mechanism is shown in Formula 4. In this way, the attention mechanism calculates the attention score at each moment, and weighted sums these scores to obtain the final output result.
[0064] Key historical information refers to time information that has an important impact on future maintenance projects. By using the attention mechanism, this information can be given a higher weight, allowing the model to focus more on the data that affects future maintenance decisions.
[0065] The synthesis result Y is shown in the following formula:
[0066] (5)
[0067] The maintenance items for the (T+1)th maintenance are obtained based on the correlation perception fusion result Y. The specific implementation process is shown in Formula 6.
[0068] (6)
[0069] in, and are weight and bias respectively, yes In order to increase the robustness of the model, A dropout operation was performed before making predictions.
[0070] B. Model Optimization
[0071] This example trains the LSTM-Att model to predict the maintenance items of each vehicle (T+ 1), using a binary cross entropy loss function as the global objective function, as shown below:
[0072] (7)
[0073] Where L is the error result, is the type of maintenance item, i is the code number of the maintenance item, for Perform a transpose operation, is the prediction result of the i-th maintenance item, is the actual result of the i-th maintenance project.
[0074] The main task of the experiment is to predict the (T +1)-time maintenance record, based on the previous T times of maintenance;
[0075] This is a multi-label classification problem. For this task, the evaluation indicators are weighted F1 score (w-F1) and R@k. w-F1 calculates the F1 score and reports its weighted average. R@k is the average ratio of the top k. It measures the accuracy of the prediction. In order to compare the method proposed in this embodiment with the most advanced model, this embodiment selects the following method as a comparative experiment.
[0076] Implementation details:
[0077] To verify the performance of the method in this embodiment, this embodiment compares it with typical machine learning and deep learning time series prediction methods, including MLP, RNN, LSTM, GRU and Transformer. In addition, it is also compared with typical methods in the field of medical diagnosis prediction, such as maintenance project prediction and Dipole.
[0078] In this example, the model parameters are randomly initialized. Hyperparameters and activation functions are adjusted based on the validation set. Specifically, this example sets the threshold λ to 0.01 and the batch size to 32. When training the model, this example uses 100 epochs and the Adam optimizer. The learning rate is set to 0.001. All projects are implemented using Python 3.7.0 and Python 1.10.0 and CUDA 11.4 on a device with 64GB memory and NVIDIA-SMI 472.39 GPU. To ensure robustness, this example repeats the experiment five times with different random seeds.
[0079] Predictive Performance:
[0080] Table 2 lists the experimental results. Since the average number of maintenance items per maintenance is 5.20, this embodiment sets k=[5, 10, 15, 20, 25, 30, 35]. R@k. Obviously, the performance of the model proposed in this embodiment is better than all baseline models. From w-F1, LSTM-Att is 0.67% higher than the best basic RNN. Because F 一 is the harmonic mean of precision and recall, which shows that the LSTM-Att model performs well in both metrics, proving that the model is able to accurately predict most maintenance items while minimizing false predictions. From the perspective of R@k, with the change of k, the LSTM-Att model always maintains its performance advantage. This shows that the LSTM-Att model is able to accurately predict the required maintenance items in both short and long lists, showing strong generalization ability.
[0081] Table 2: Comparison of experimental results between the model of this embodiment and the best existing model
[0082] Model <![CDATA[w-F1]]> R@5 R@10 R@15 R@20 R@25 R@30 R@35 MLP 45.79±0.14 64.13±0.32 70.00±0.21 74.19±0.27 76.98±0.11 79.25±0.20 79.43±0.38 80.21±0.31 RNN 46.53±0.17 64.43±0.16 70.72±0.19 74.66±0.07 77.81±0.18 79.75±0.12 80.95±0.40 81.51±0.13 LSTM 46.31±0.15 64.79±0.34 69.93±0.16 74.00±0.20 77.29±0.22 79.77±0.29 82.44±0.28 83.64±0.31 GRU 46.42±0.13 64.60±0.31 70.05±0.30 74.51±0.27 77.91±0.18 79.98±0.11 81.38±0.18 82.14±0.24 Dipole 45.81±0.18 64.37±0.14 70.08±0.16 74.28±0.18 77.92±0.09 80.13±0.28 82.28±0.40 83.47±0.17 transformer 45.81±0.28 64.23±0.23 69.72±0.39 74.26±0.71 77.18±0.57 79.55±0.21 79.96±0.41 81.21±0.36 LSTM-Att 46.84±0.34 64.87±0.34 71.97±0.16 76.49±0.07 79.13±0.19 81.63±0.23 83.61±0.19 85.14±0.28
[0083] Performance Evaluation of Data Adequacy
[0084] To evaluate the impact of data sufficiency on prediction accuracy, this example keeps the size of the validation set at 3,256 entries. This example changes the size of the training set to 3,000, 3,500, 4,000, and 4,500, respectively, and uses the remaining data for the test set. Figures 3 to 10 demonstrate the superior performance of LSTM-Att over other baseline models, even when data is limited.
[0085] In view of the limitation that existing methods can only predict some maintenance items of automobiles, this embodiment develops a prediction model based on LSTM and attention to predict all maintenance items of automobiles. Experimental analysis confirms that the model performance of this embodiment is excellent.
[0086] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.
Claims
1. A method for predicting automobile maintenance items based on deep learning, characterized in that: include: Obtain automobile maintenance data and build a data set; Using the key information learning strategy, the long short-term memory network is combined with the attention mechanism to build the original prediction model; Using the data set to train the original prediction model to obtain a prediction model for automobile maintenance items; Using the automobile maintenance item prediction model, predict all maintenance items of the automobile; The automobile maintenance data is data of vehicles whose maintenance times are greater than or equal to two times, and the automobile maintenance data includes: number of vehicles, maximum number of maintenance times, average number of maintenance times, number of maintenance item codes, maximum number of item codes for each maintenance, and average number of item codes for each maintenance; The attention mechanism is: Among them, Q is the query vector, K is the key vector, V is the value vector, and W q is the weight of the query vector, W k is the weight of the key vector, T is the matrix transpose operation, W v is the weight of the value vector, d is the attention dimension; Using the data set to train the original prediction model includes: The last maintenance record in the data set is used as a label, and the remaining maintenance records are used as input features to train the original prediction model, so that the original prediction model can predict the next maintenance item based on the maintenance items of the previous T maintenances; Using the data set to train the original prediction model further includes: using a binary cross entropy loss function as a global objective function; The binary cross entropy loss function is: Where L is the error result, is the type of maintenance item, i is the code number of the maintenance item, for Perform a transpose operation, is the prediction result of the i-th maintenance item, is the actual result of the i-th maintenance project; In the original prediction model, the long-term dependencies in the maintenance records are captured through the long short-term memory network, so that the model can learn the long-term rules in the development process of maintenance projects, and emphasize key historical information through the attention mechanism, so that the model can learn the impact of historical key maintenance projects on future maintenance projects.
Citation Information
Patent Citations
Method and system for predicting fatigue life durability of automobile
CN116629078A