A method and system for predicting automotive maintenance and repair requirements based on collaborative multi-perspectives

Abstract

The present invention discloses a method and system for predicting automotive maintenance and repair requirements based on collaborative multi-perspectives, belonging to the field of automotive maintenance and repair, including: performing temporal dependence learning on vehicle historical maintenance records through a multi-attention mechanism to obtain the temporal dependence between automotive maintenance items in different periods; based on the temporal dependence, performing dependence-aware temporal feature weighting on the historical maintenance records to obtain weighted temporal data; combining a long short-term memory network and an attention mechanism to learn key temporal information from the weighted temporal data to obtain a learning result of key temporal information; multiplying the weighted temporal data by the learning result of key temporal information to obtain a collaborative temporal modeling fusion result, and predicting future maintenance items based on the collaborative temporal modeling fusion result. The present invention can predict all automotive maintenance and repair requirements and eliminate the safety risks existing in the vehicle.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle maintenance, and particularly relates to a method and system for predicting vehicle maintenance requirements based on collaborative multi - perspectives. Background Art

[0002] Vehicle maintenance plays a crucial role in the vehicle life cycle. It is not only the basis for ensuring the safe operation of the vehicle but also an effective way to extend the vehicle's service life, optimize performance, and reduce usage costs. During the vehicle's use, as various components of the vehicle wear and age, the vehicle's maintenance requirements become more diverse and complex. In this context, accurately predicting vehicle maintenance requirements is particularly important, as it can not only prevent unnecessary maintenance, reduce the vehicle's downtime, but also quickly identify potential vehicle faults and extend the vehicle's service life. In the field of vehicle maintenance, the interdependence between maintenance items carried out at different times cannot be ignored. For example, if the engine overheating problem was not completely solved during the previous maintenance, then during a future maintenance, a more in - depth inspection of the engine cooling system may be required; another example is that if the vehicle's starting battery was replaced previously, then during the subsequent maintenance, it is particularly important to check the charging efficiency of the electrical system to ensure that no other potential electrical problems cause the new battery to be damaged prematurely.

[0003] In addition, during the vehicle maintenance process, certain key maintenance items have an important impact on future maintenance requirements. For example, regularly changing the engine oil is a key item in vehicle maintenance. New engine oil can effectively lubricate engine parts, reduce friction, and minimize part wear. If the engine oil is not changed regularly, it can lead to engine overheating, oil shortage, and even engine failure. Therefore, regularly changing the engine oil helps protect the engine and extend its service life, and has an important impact on the subsequent engine maintenance process. In addition, regularly checking components such as brake discs, brake pads, and brake fluid and keeping them in good condition play an irreplaceable role in the subsequent maintenance process of the braking system.

[0004] However, existing vehicle maintenance requirement prediction methods can only predict some of the vehicle's maintenance requirements and cannot predict all of the vehicle's maintenance requirements, resulting in certain safety risks for the vehicle and affecting its performance. Summary of the Invention

[0005] The present invention proposes a method and system for predicting vehicle maintenance requirements based on collaborative multi - perspectives 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 vehicle maintenance requirements based on collaborative multi - perspectives, including the following steps:

[0007] Perform temporal dependence learning on the vehicle's historical maintenance records through a multi-attention mechanism to obtain the temporal dependence between vehicle maintenance items in different periods;

[0008] Based on the temporal dependence, perform dependence-aware temporal feature weighting on the historical maintenance records to obtain weighted temporal data;

[0009] Combine the long short-term memory network and the attention mechanism to learn key temporal information from the weighted temporal data to obtain the learning result of key temporal information;

[0010] Multiply the weighted temporal data by the learning result of key temporal information to obtain the collaborative temporal modeling fusion result, and predict the items that need to be maintained in the future based on the collaborative temporal modeling fusion result.

[0011] Preferably, the temporal dependence learning includes:

[0012] Use two GRU neural networks to process the historical maintenance records to obtain two hidden state sequences; calculate the attention weights of the two hidden state sequences through the Softmax and Tanh functions to evaluate the importance of the historical maintenance records for future maintenance requirements.

[0013] Preferably, the dependence-aware temporal feature weighting of the historical maintenance records includes:

[0014] Calculate the dynamic correlation representation learning result of each historical maintenance record through the Kronecker product and element-wise multiplication;

[0015] Obtain the dependence-aware result of the historical maintenance records based on the dynamic correlation representation learning result;

[0016] Merge the dependence-aware result of the maintenance history records with each time step through the attention mechanism to obtain the dependence-aware temporal feature weighting result.

[0017] Preferably, the learning of key temporal information includes:

[0018] Use the long short-term memory network to process the historical maintenance records to obtain the hidden state of the most recent maintenance record; process the hidden state through the attention mechanism to extract key information.

[0019] Preferably, the expression for predicting the items that need to be maintained in the future based on the collaborative temporal modeling fusion result is:

[0020] ;

[0021] In the formula, represents the prediction output, W Y represents the weight, b YLet P represent paranoia, Y represent the collaborative time-series modeling fusion result, and dropout represent the dropout operation.

[0022] The present invention also provides a vehicle maintenance demand prediction system based on collaborative multi-perspectives, including:

[0023] A time-series dependence learning module, configured to perform time-series dependence learning on the vehicle historical maintenance records through a multi-attention mechanism to obtain the time-series dependence between vehicle maintenance items in different periods;

[0024] A dependence-aware time-series feature weighting module, configured to perform dependence-aware time-series feature weighting on the historical maintenance records based on the time-series dependence to obtain weighted time-series data;

[0025] A key time-series information learning module, configured to learn key time-series information from the weighted time-series data by combining a long short-term memory network and an attention mechanism to obtain a key time-series information learning result;

[0026] A prediction module, configured to multiply the weighted time-series data by the key time-series information learning result to obtain a collaborative time-series modeling fusion result, and predict the items that need to be maintained in the future based on the collaborative time-series modeling fusion result.

[0027] Preferably, the time-series dependence learning module includes:

[0028] Two gated recurrent unit neural networks, configured to process the historical maintenance records and generate a hidden state sequence;

[0029] An attention weight calculation unit, configured to calculate the attention weights of two hidden state sequences based on the hidden state sequence.

[0030] The present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory, where the processor executes the computer program to implement the steps of the method.

[0031] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method are implemented.

[0032] The present invention also provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of the method are implemented.

[0033] Compared with the prior art, the present invention has the following advantages and technical effects:

[0034] The present invention discloses a method and system for predicting automotive maintenance and repair requirements based on a collaborative multi - perspective approach, including: performing temporal dependence learning on vehicle historical maintenance records through a multi - attention mechanism to obtain the temporal dependence between automotive maintenance items in different periods; based on the temporal dependence, performing dependence - aware temporal feature weighting on the historical maintenance records to obtain weighted temporal data; combining a long short - term memory network and an attention mechanism to learn key temporal information from the weighted temporal data to obtain a learning result of key temporal information; multiplying the weighted temporal data by the learning result of key temporal information to obtain a collaborative temporal modeling fusion result, and predicting the items that need maintenance in the future based on the collaborative temporal modeling fusion result. The present invention can predict all automotive maintenance requirements and eliminate the safety risks existing in the vehicle. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0036] Figure 1 It is a collaborative multi - perspective time - series modeling framework diagram for vehicle maintenance requirement prediction according to an embodiment of the present invention;

[0037] Figure 2 It is a structure diagram of GRU according to an embodiment of the present invention;

[0038] Figure 3 It is a structure diagram of LSTM according to an embodiment of the present invention;

[0039] Figure 4 It is a result schematic diagram of w - F1, R@5, R@10, R@15 of different models according to an embodiment of the present invention;

[0040] Figure 5 It is a result schematic diagram of R@20, R@25, R@30, R@35 of different models according to an embodiment of the present invention;

[0041] Figure 6 It is a parameter sensitivity analysis diagram according to an embodiment of the present invention;

[0042] Figure 7 It is a performance schematic diagram under different attention weight reconstruction modeling combinations according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will refer to the drawings and combine the embodiments to detail this application.

[0044] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0045] Embodiment 1

[0046] To facilitate a detailed description of the maintenance demand prediction task to be processed in this article, some symbols that run through the entire text are first given in Table 1. During the entire maintenance process of an automobile, a large number of maintenance items need to be carried out. All maintenance items are encoded to form a set of item code sets denoted as , where represents the type of maintenance item. Each vehicle has multiple records of maintenance items. The historical maintenance items of the vehicle are denoted as , and the maintenance item of the th maintenance is denoted as a multi-hot vector , indicating that the maintenance item appears in the th maintenance item record, where , , represents the number of vehicle maintenance times. The task of maintenance demand is to predict the maintenance item of the th time based on the historical maintenance items .

[0047] Table 1

[0048]

[0049] As Figure 1 shown, in this embodiment, a method for predicting automobile maintenance demand based on a collaborative multi-perspective is provided, including the following steps:

[0050] Perform temporal dependence learning on the vehicle historical maintenance records through a multi-attention mechanism to obtain the temporal dependence between automobile maintenance items in different periods;

[0051] Based on the temporal dependence, perform dependence-aware temporal feature weighting on the historical maintenance records to obtain weighted temporal data;

[0052] Combine the long short-term memory network and the attention mechanism to learn key temporal information from the weighted temporal data to obtain the learning result of key temporal information;

[0053] Multiply the weighted temporal data by the learning result of key temporal information to obtain a collaborative temporal modeling fusion result, and predict the items that need to be maintained in the future based on the collaborative temporal modeling fusion result.

[0054] Furthermore, the temporal dependency is learned as follows:

[0055] Throughout the life cycle of a car, there are complex dependencies between the maintenance items required at each stage. The maintenance items carried out in the early stage not only have a direct impact on the current operating status of the car, but also affect future maintenance needs. In order to deeply explore these dependencies, this embodiment designs a temporal dependency learning module based on attention weight modeling, which aims to learn from historical maintenance records and predict future maintenance needs.

[0056] First, two GRU neural networks GRU1 and GRU2 are used to process the historical maintenance records, and two hidden state sequences are obtained respectively: {N t} t=1,2,...,T and {O t} t=1,2,...,T These hidden states not only contain the original maintenance record information at each time step, but also integrate the information of all previous time steps, which helps to fully understand the maintenance history. The specific implementation process is shown in Formula 1.

[0057]

[0058] The calculation process of the GRU neural network in the above formula is shown in formula (2), and the structural diagram is shown in Figure 2 ,in is the weight, For bias.

[0059]

[0060] During the car maintenance process, the maintenance personnel will pay attention to the maintenance items that the car has undergone in the past, and decide the maintenance items that the car needs to undergo based on the historical maintenance information. To simulate the above process, this embodiment will include the original maintenance record information and the accumulated record information. and As the attention features of the previous maintenance records, and then use Softmax and tanh as activation functions to calculate the two attention weights of the previous maintenance records and , the specific implementation process is shown in equations (3) and (4);

[0061]

[0062]

[0063] in and denote weight and bias respectively, and A mechanism for evaluating the importance of historical maintenance records from different perspectives enhances the model's ability to capture complex temporal dependencies through the non-linearity of the activation function.

[0064] Furthermore, the dependency-aware temporal feature weighting is as follows:

[0065] After calculating the attention weights of each historical maintenance record, this embodiment proposes a temporal dependency-aware function F(Mt, Et, Ht) to obtain the dynamic correlation representation learning results {Ut}t=1,2,...,T of each historical maintenance record. The definition of F(Mt, Et, Ht) is shown in Equation 5.

[0066]

[0067] where ⊙ represents the Kronecker product, and ⊗ represents the element-wise multiplication (Hadamard product). The dynamic correlation representation learning results {U t} t=1,2,...,T not only contain the information of each historical original maintenance record but also contain the temporal dependency information of each historical maintenance record from different perspectives. Based on {U t}, the dependency-aware result of the historical maintenance record is obtained. t=1,2,...,T

[0068] To achieve a better weighting effect of the dependency-aware result of the historical maintenance record with each time step, this embodiment uses the attention mechanism to merge the dependency-aware result of the maintenance history with each time step to obtain the dependency-aware temporal feature weighting result D. The specific implementation process is shown in Equation 6.

[0069]

[0070] where C = [M1; M2;...; M T , and the implementation process of the above attention mechanism is shown in Equation 7.

[0071]

[0072] where d is the attention dimension, W q , W k ∈ R|P|×d, W v ∈ R|P|×|P| represents the attention weight.

[0073] Furthermore, the key information learning and prediction are as follows:

[0074] ​During the process of vehicle maintenance, certain key maintenance items have an important impact on future maintenance requirements. To better capture these crucial information in historical maintenance records, this embodiment proposes a method for learning key information of vehicle maintenance by combining long short-term memory network and attention mechanism. First, the long short-term memory network (LSTM) is used to process historical maintenance items to obtain the hidden states R1, R2,..., R of the most recent maintenance record T , and the specific implementation process is shown in Equation (8).

[0075] (8)

[0076] According to this procedure, R1, R2,..., R T encapsulate the original data and cumulative data in consecutive maintenance records, capturing the long-term dependencies in the historical maintenance sequence. Subsequently, R1, R2,..., R T are processed through the attention mechanism, as shown in Equation (7), and its specific implementation process is detailed in Equation (9).

[0077] (9)

[0078] where R = R1, R2,..., R T . Through this process, the interaction between historical maintenance records is enhanced, and B1, B2,..., B T obtain the key time information that affects future maintenance requirements. The comprehensive result S of key time information learning is obtained by summing up this information, as shown in Equation (10).

[0079] (10)

[0080] The element-wise multiplication method is adopted to promote the enhanced fusion of the dependency-aware time feature weighted result D and the comprehensive key information learning result S, thereby obtaining the fusion result Y, as shown in Equation (11).

[0081]

[0082] The maintenance item at the (T + 1)-th time is obtained based on the fusion result Y, and the specific implementation process is shown in Equation 12.

[0083]

[0084] where and are the weights and biases respectively, is the dimension of Before making predictions, dropout operation is performed to increase the robustness of the model. represents the predicted output, and dropout represents the dropout operation.

[0085] Furthermore, the model is optimized as follows:

[0086] The CoMTM model is trained to predict the maintenance items of each vehicle at the (T + 1)-th time, and the global objective function is the binary cross-entropy loss function, as shown in Equation 13.

[0087]

[0088] where represents the predicted result of the maintenance item , and represents the true label of the maintenance item .

[0089] The following experimental verification is also carried out in this embodiment:

[0090] To evaluate the performance of the method proposed in this embodiment, this embodiment uses the real vehicle maintenance data from 40 vehicle maintenance companies for verification. After screening, the data of vehicles with the number of maintenance times greater than or equal to two, as well as the complete data and basic information of each maintenance, are retained. The dataset includes the records of 10,256 vehicles maintained during the period from April 2011 to April 2023. The detailed dataset statistics are shown in Table 2.

[0091] Table 2

[0092] Number of vehicles 10256 Maximum maintenance times 69 Average maintenance times 3.92 Number of maintenance codes 2428 Maximum number of maintenance codes in one time 51 Average number of maintenance codes in one time 5.20

[0093] To strengthen the experimental process, this embodiment randomly divides the dataset into a training set, a validation set, and a test set. Specifically, these sets are composed of 6000 vehicles, 3256 vehicles, and 1000 vehicles respectively. In the method of this embodiment, this embodiment designates the maintenance items of the last maintenance record as the labels, and at the same time uses the remaining historical maintenance records as input features. The item co-occurrence graph is constructed based on the maintenance items in the training set.

[0094] The baseline model and evaluation metrics are as follows:

[0095] The main task of this experiment is to predict the (T + 1)-th maintenance record based on the previous T maintenance records of the vehicle, which is a multi-label classification problem. For this task, the evaluation metrics are the weighted F1 score (w-F1) and R@k. w-F1 calculates the F1 score for each item code and reports its weighted average. R@k is the ratio of the average ratio of the expected item codes in the top k predictions for each maintenance to the total number of all expected item codes in each maintenance. It measures the accuracy of the prediction. To compare with the latest models, the following methods are selected as comparative experiments in this embodiment.

[0096] To verify the performance of the method in this embodiment, it is compared with typical machine learning and deep learning time series prediction methods in this embodiment, including MLP, CNN, RNN, LSTM, and Transformer. In addition, it is also compared with typical methods for similar medical diagnosis predictions in the field of maintenance demand prediction, including RETAIN, Dipole, and Chet.

[0097] Table 3

[0098] Model w-F 1 R@5 R@10 R@15 R@20 R@25 R@30 R@35 MLP 30.95±0.14 52.90±0.32 59.35±0.21 64.51±0.27 68.38±0.11 71.81±0.20 75.01±0.38 77.46±0.31 CNN 39.70±0.54 56.07±0.27 63.58±0.27 68.54±0.44 72.15±0.36 74.93±0.30 77.26±0.39 79.16±0.35 Transformer 30.93±0.28 52.90±0.23 58.98±0.39 63.74±0.71 67.90±0.57 71.47±0.21 74.23±0.41 76.76±0.36 RNN 30.93±0.17 52.90±0.16 59.40±0.19 64.53±0.07 68.45±0.18 71.96±0.12 74.98±0.40 77.56±0.13 LSTM 31.00±0.15 52.90±0.34 59.27±0.16 64.31±0.20 68.41±0.22 72.08±0.29 75.04±0.28 77.30±0.31 Dipole 31.05±0.18 52.90±0.14 59.35±0.16 64.46±0.18 68.60±0.09 71.91±0.28 74.75±0.40 77.25±0.17 RETAIN 40.12±0.13 56.23±0.31 63.67±0.30 68.64±0.27 72.32±0.18 75.25±0.11 77.71±0.18 79.76±0.24 Chet 39.26±0.46 56.06±0.19 63.54±0.11 68.38±0.17 72.04±0.28 75.00±0.27 77.56±0.20 79.60±0.21 CoMTM 41.06±0.11 56.76±0.03 64.54±0.06 69.58±0.11 73.29±0.07 76.26±0.07 78.78±0.10 80.89±0.07

[0099] In the experiment of this embodiment, the model parameters are randomly initialized. The hyperparameters and activation functions are tuned on the validation set. Specifically, when training the model of this embodiment, the batchsize is set to 32, 100 epochs are used, and the Adam (Kingma and Ba 2015) optimizer is used. The learning rate is set to 0.001. All programs are implemented on a device with Python 3.7.0, PyTorch 1.10.0, CUDA 11.4, 64GB of memory, and an NVIDIA-SMI 472.39 GPU. To ensure robustness, the experiment is repeated 5 times with different random seeds in this embodiment.

[0100] According to the experimental results shown in Table 3, in this embodiment, it can be seen that the CoMTM model is significantly superior to other comparison models in the task of maintenance project prediction. Considering that the average number of maintenance projects per maintenance is 5.20, different k values are set in this study, including [5, 10, 15, 20, 25, 30, 35], to comprehensively evaluate the performance of R@k. The data shows that the CoMTM model outperforms all the listed baseline models whether in a short or long prediction list. From the perspective of the w-F1 index, the CoMTM model leads with a score of 41.06±0.11%, which is 2.33% higher than the existing best RETAIN model. Generally, the higher the F1 score, the better the balance between precision and recall of the model. Therefore, the CoMTM model can effectively predict more correct maintenance projects while suppressing the number of false predictions, which is very important in practical scenarios. In terms of multiple evaluation indicators of R@k, the CoMTM model also shows obvious advantages. For example, the scores of R@5, R@10, R@15, R@20, R@25, R@30, and R@35 are all the highest for this model, indicating that CoMTM can maintain high performance in prediction lists of different lengths, which is particularly crucial for practical applications because users may select different numbers of prediction results according to actual needs. Generally speaking, the CoMTM model not only demonstrates its high accuracy and generalization ability in the vehicle maintenance project prediction task, but also its consistent performance across different prediction list lengths further verifies the practicality and forward-looking nature of this model. These experimental results provide valuable references and a basis for future research and practical applications.

[0101] To examine the impact of data sufficiency on prediction, in this embodiment, by fixing the size of the validation set at 3,256, the numbers of the training set are respectively set to 3,000, 4,000, 5,000, and 6,000, and the remaining data is used as the test set. From Figure 4 - Figure 5 it can be seen that compared with other benchmark models, CoMTM still performs excellently in the case of insufficient data.

[0102] The ablation experiments are as follows:

[0103] To further analyze the effectiveness of each module proposed in this embodiment, three ablation variants of the model are also compared in this embodiment. The specific settings of each model are as follows:

[0104] CoMTM-TDL1: To prove the role of the first temporal dependence learning in the temporal dependence learning part, this model cancels the input of {E t} t=1,2,...,T in the temporal dependence perception function;

[0105] Table 4

[0106] Model w-F 1 R@5 R@10 R@15 R@20 R@25 R@30 R@35 CoMTM-TDL1 40.57±0.24 56.63±0.07 64.33±0.08 69.47±0.13 73.12±0.14 76.24±0.16 78.63±0.10 80.70±0.12 CoMTM-TDL2 41.03±0.08 56.75±0.08 64.48±0.09 69.54±0.08 73.26±0.04 76.04±0.11 78.59±0.07 80.75±0.11 CoMTM-Atten 40.58±0.05 56.68±0.09 64.41±0.09 69.50±0.11 73.07±0.08 75.90±0.08 78.27±0.05 80.47±0.05 CoMTM 41.06±0.11 56.76±0.03 64.54±0.06 69.58±0.11 73.29±0.07 76.26±0.07 78.78±0.10 80.89±0.07

[0107] CoMTM-TDL2: To prove the role of the second temporal dependence learning in the temporal dependence learning part, this model cancels the input of {H t} t=1,2,...,T in the temporal dependence perception function;

[0108] CoMTM-Atten: To prove the role of the attention mechanism in the key temporal information learning and prediction part, the attention mechanism is cancelled in this model, making S = B.

[0109] The results of all ablation experiments are shown in Table 4. Several key conclusions can be drawn from this embodiment. First, the performance of each model variant has decreased compared to the complete CoMTM model, indicating that the proposed model components - the two learning modules in the temporal dependence learning part and the attention mechanism - all play important roles in improving the prediction performance of the model.

[0110] Especially in terms of the performance of w-F1 score and R@k, the CoMTM model outperforms the other three variants in all evaluation metrics, indicating that each part introduced in the model is necessary for the improvement of the overall performance. The results of CoMTM-TDL1 and CoMTM-TDL2 are close, but CoMTM-TDL2 is slightly better in most metrics, which may indicate that the second temporal dependence learning is slightly more important for capturing and utilizing temporal dependence information. In addition, although the performance of CoMTM-Atten lags behind the other two ablation variants, it is still relatively good, which proves that the role of the attention mechanism in the model is to improve the prediction accuracy by more effectively focusing on key temporal information.

[0111] Through the systematic ablation study, this embodiment verifies the importance and effectiveness of each component in the CoMTM model. Future work can further explore the optimization direction of the model structure based on these findings, such as improving the strategy of temporal dependence learning, introducing more advanced attention mechanisms, or exploring new model fusion methods to improve the accuracy and robustness of the model. In addition, considering applying the model to a wider range of automotive fault prediction and maintenance scenarios is also a meaningful direction for future work, in order to achieve more comprehensive and in-depth optimization of vehicle maintenance strategies.

[0112] In the field of automotive maintenance, the prediction of new maintenance items has important practical significance. In this study, this embodiment defines new maintenance items as those that have not appeared in the previous maintenance history of the vehicle. Such predictions are particularly crucial for discovering potential vehicle problems and expanding the scope of maintenance services. Based on the assumption that there is a dependence between maintenance records at different times, the model proposed in this embodiment should have better performance in the task of predicting new maintenance items.

[0113] The experimental results of each model's prediction for the new maintenance project are shown in Table 5. It can be analyzed that CoMTM has surpassed all baseline models in terms of F1 score, and the F1 score of the closest baseline model, RETAIN, is about 2 percentage points lower than that of CoMTM. In addition, CoMTM also performs excellently in the R@k metric (i.e., the ratio of correctly predicted new maintenance projects among the top k predictions), especially under higher k values. For example, in R@35, CoMTM is about 1.73 percentage points higher than the closest competitor, Chet, which further verifies the effectiveness and superiority of CoMTM in dealing with the prediction of new maintenance projects.

[0114] Table 5

[0115] Model w-F 1 R@5 R@10 R@15 R@20 R@25 R@30 R@35 MLP 14.88±0.30 22.58±0.13 31.77±0.41 38.80±0.17 44.92±0.35 50.63±0.14 54.61±0.28 57.92±0.23 CNN 19.64±0.53 24.93±0.41 34.88±0.46 42.31±0.45 48.06±0.35 53.08±0.29 57.52±0.32 60.99±0.20 Transformer 14.79±0.31 22.32±0.55 31.74±0.24 38.62±0.46 44.52±0.62 50.01±0.39 54.29±0.58 57.72±0.44 RNN 14.85±0.27 22.64±0.16 31.93±0.15 38.80±0.17 45.27±0.29 50.43±0.41 54.50±0.22 58.10±0.21 LSTM 14.89±0.18 22.62±0.18 31.80±0.19 38.65±0.21 45.15±0.29 50.54±0.24 54.38±0.33 57.94±0.32 Dipole 14.96±0.25 22.48±0.05 31.76±0.30 38.91±0.21 45.09±0.27 50.59±0.38 54.66±0.33 57.99±0.38 RETAIN 19.90±0.51 25.19±0.57 34.85±0.31 41.85±0.40 47.81±0.48 52.73±0.47 56.87±0.51 60.49±0.67 Chet 19.25±0.35 25.23±0.35 35.23±0.23 42.63±0.27 48.43±0.29 53.22±0.39 57.47±0.43 61.12±0.57 CoMTM 21.09±0.03 26.34±0.08 36.89±0.10 43.80±0.12 49.31±0.08 55.21±0.13 59.12±0.09 62.85±0.12

[0116] These results indicate that the CoMTM model significantly improves the prediction accuracy of new maintenance projects by fully utilizing the time series dependence and complex feature interactions between maintenance records. Compared with traditional machine learning methods and simple deep learning structures, CoMTM successfully captures the subtle patterns in maintenance records by combining the representation ability of deep learning and the advantages of complex model architectures, and effectively applies this knowledge to the prediction task.

[0117] The parameter sensitivity analysis is as follows:

[0118] To study the impact of the main hyperparameters involved in CoMTM on the model performance, this embodiment examines the sensitivity of some hyperparameters, including the dropout rate in Formula 12 and the model groups used for the two attention weight reconstructions in the time series dependence learning part. This embodiment sets the dropout rate in the range of 0.44, 0.45, 0.46, 0.47, 0.48, 0.49, 0.50. According to Figure 6 the results in, when the dropout rate = 0.47, the overall performance remains relatively stable and reaches the best performance. Figure 7 shows the performance of the proposed CoMTM model under different combinations of attention weight reconstruction models. The experimental results show that the combination of GRU and GRU used in this embodiment has the best performance.

[0119] This embodiment also provides a vehicle maintenance demand prediction system based on collaborative multi-perspectives, including:

[0120] A time series dependence learning module for performing time series dependence learning on the vehicle's historical maintenance records through a multi-attention mechanism to obtain the time series dependence between vehicle maintenance projects in different periods;

[0121] A dependency-aware temporal feature weighting module, configured to perform dependency-aware temporal feature weighting on historical maintenance records based on the temporal dependency, so as to obtain weighted temporal data;

[0122] A key temporal information learning module, configured to learn key temporal information from the weighted temporal data by combining a long short-term memory network and an attention mechanism, so as to obtain a key temporal information learning result;

[0123] A prediction module, configured to multiply the weighted temporal data by the key temporal information learning result to obtain a collaborative temporal modeling fusion result, and predict items that need to be maintained in the future based on the collaborative temporal modeling fusion result.

[0124] Further, the temporal dependency learning module includes:

[0125] Two gated recurrent unit neural networks, configured to process historical maintenance records and generate a hidden state sequence;

[0126] An attention weight calculation unit, configured to calculate attention weights of two hidden state sequences based on the hidden state sequence.

[0127] This embodiment further provides a computer device, including a memory, a processor, and a computer program stored on the memory, where the processor executes the computer program to implement the steps of the method.

[0128] This embodiment further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method are implemented.

[0129] This embodiment further provides a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of the method are implemented.

[0130] The above is only a preferred specific embodiment 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 those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A collaborative multi-perspective automobile maintenance demand prediction method, characterized in that: The following steps are involved: Through the multi-attention mechanism, the temporal dependency of historical vehicle maintenance records is learned to obtain the temporal dependency between vehicle maintenance items in different periods; The timing dependency learning includes: Use two GRU neural networks to process historical maintenance records and obtain two hidden state sequences. Use Softmax and Tanh functions to calculate the attention weights of the two hidden state sequences to evaluate the importance of historical maintenance records to future maintenance needs. Based on the time series dependency, weighting the time series features of the historical maintenance records with dependency awareness to obtain weighted time series data; The time series feature weighting for dependency perception of historical maintenance records includes: The dynamic correlation of each historical maintenance record is calculated through Kronecker product and element-by-element multiplication to represent the learning result; Obtain dependency perception results of historical maintenance records based on dynamic correlation representation learning results; The dependency-aware results of the maintenance history records are merged with each time step through the attention mechanism to obtain the weighted results of the dependency-aware temporal features. Combining the long short-term memory network with the attention mechanism to learn key time series information from weighted time series data, and obtain key time series information learning results; The weighted time series data is multiplied by the key time series information learning result to obtain a collaborative time series modeling fusion result, and the projects that need maintenance in the future are predicted based on the collaborative time series modeling fusion result.

2. The method according to claim 1, characterized in that: The learning key timing information includes: The long short-term memory network is used to process historical maintenance records to obtain the hidden state of the most recent maintenance record; the hidden state is processed through the attention mechanism to extract key information.

3. The method according to claim 1, characterized in that The expression for predicting the items that need maintenance in the future based on the fusion results of the collaborative time series modeling is: ； Where W Y represents the weight, b Y represents bias, Y represents the collaborative time series modeling fusion result, Represents the predicted output, and dropout represents the discarding operation.

4. A collaborative multi-perspective automobile maintenance demand prediction system, characterized in that: include: The temporal dependency learning module is used to learn the temporal dependency of historical vehicle maintenance records through a multi-attention mechanism to obtain the temporal dependency between vehicle maintenance items in different periods; A dependency-aware time series feature weighting module, used to perform dependency-aware time series feature weighting on historical maintenance records based on the time series dependency to obtain weighted time series data; The key time series information learning module is used to combine the long short-term memory network and the attention mechanism to learn the key time series information from the weighted time series data and obtain the key time series information learning results; The prediction module is used to multiply the weighted time series data with the key time series information learning result to obtain the collaborative time series modeling fusion result, and predict the projects that need maintenance in the future based on the collaborative time series modeling fusion result.

5. The system according to claim 4, characterized in that The timing dependency learning module includes: Two gated recurrent unit neural networks to process historical maintenance records and generate hidden state sequences; The attention weight calculation unit is used to calculate the attention weights of two hidden state sequences based on the hidden state sequence.

6. 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 3.

7. 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 3 are implemented.

8. 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 3 are implemented.

Images