Automobile maintenance item prediction method based on multi-source data and attention mechanism

Abstract

The present invention relates to the field of automobile maintenance technology, and in particular to an automobile maintenance project prediction method based on multi-source data and an attention mechanism, comprising: obtaining historical maintenance project records, mileage, and basic vehicle information of a target automobile; inputting the historical maintenance project records, mileage, and basic vehicle information into a preset prediction model to predict the next automobile maintenance project of the target automobile, wherein the prediction model is obtained by training a training set, the training set includes historical maintenance project records, mileage, and basic vehicle information of several scrapped automobiles, and the prediction model is constructed based on a multi-scale deformation attention mechanism. The present invention can comprehensively predict the maintenance status of an automobile and improve traffic safety by integrating the historical maintenance project records, mileage, and basic vehicle information of an automobile.

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 multi-source data and an attention mechanism. Background Art

[0002] With the rapid development of the automobile industry, automobiles have become an indispensable part of many people's lives, and the automobile maintenance industry has become an important livelihood service industry that affects the quality of life of the people. The prediction of automobile maintenance items is of great significance for ensuring the safe and reliable driving of automobiles, reducing automobile maintenance costs and improving the quality of life of car owners. It has attracted the attention of more and more researchers and automobile maintenance industry practitioners. At present, automobiles only arrange maintenance items based on fixed time or mileage, which is a typical planned maintenance. Ignoring the major practical problem that the usage and driving environment of different vehicles are different, it not only reduces the flexibility of maintenance, but also easily causes excessive maintenance and wastes maintenance costs. In addition, the existing work mainly predicts the local maintenance items of automobiles based on the mileage and time of automobiles. However, it is not enough to predict only the local parts of the automobile. It is impossible to grasp the maintenance status of other parts of the automobile, which poses a great safety hazard. Therefore, there is an urgent need for a prediction method for automobile maintenance items based on multi-source data and attention mechanism. Summary of the invention

[0003] The purpose of this invention is to provide a method for predicting automobile maintenance items based on multi-source data and attention mechanism. By integrating the historical maintenance item records, mileage and basic vehicle information of the automobile, it can comprehensively predict the maintenance status of the automobile and improve traffic safety.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] Automobile maintenance project prediction method based on multi-source data and attention mechanism, including:

[0006] Obtain the historical maintenance records, mileage and basic vehicle information of the target vehicle;

[0007] The historical maintenance item records, mileage and basic vehicle information are input into a preset prediction model to predict the next vehicle maintenance item of the target vehicle, wherein the prediction model is obtained by training a training set, and the training set includes historical maintenance item records, mileage and basic vehicle information of several scrapped vehicles, and the prediction model is constructed based on a multi-scale deformation attention mechanism.

[0008] Optionally, the historical maintenance item record includes maintenance items and maintenance mileage information;

[0009] The basic vehicle information includes the vehicle license plate number, vehicle manufacturer, vehicle type, engine type, engine displacement, vehicle model, production year, and assembly manufacturer information.

[0010] Optionally, the prediction model includes a fusion processing module and a result prediction module, the fusion processing module is used to fuse the historical maintenance item records, mileage and vehicle basic information to obtain a car representation result; the result prediction module is used to predict the next car maintenance item based on the car representation result.

[0011] Optionally, the fusion processing module includes: a processing unit, a first fusion unit, and a second fusion unit; the processing unit is used to perform correlation learning based on the historical maintenance project records to obtain the correlation representation learning results of each historical maintenance project; the first fusion unit is used to use the attention mechanism to fuse the mileage with the correlation representation learning results of each historical maintenance project to obtain a fusion representation result; the second fusion unit is used to use the attention mechanism to fuse the vehicle basic information with the fusion representation result to obtain the vehicle representation result.

[0012] Optionally, the processing unit performs correlation learning based on the historical maintenance item records, and a method for obtaining a correlation representation learning result of each historical maintenance item is:

[0013]

[0014]

[0015] in, , represents the co-occurrence matrix of maintenance items, m and n are maintenance items m and maintenance items n respectively, is the co-occurrence frequency of maintenance item m and maintenance item n, is the frequency filtering threshold, For the types of maintenance projects, is the total frequency of maintenance item m meeting the requirements, For each historical maintenance project The correlation representation function.

[0016] Optionally, the first fusion unit fuses the driving mileage with the correlation representation learning result of each historical maintenance item by using an attention mechanism, and a method for obtaining the fusion representation result is:

[0017] Mapping the driving mileage to a latent space that is the same as the learning result of the correlation representation of each historical maintenance item;

[0018] The latent space is used as a query vector, and the correlation representation learning result of each historical maintenance project is used as a key vector and a value vector, which are input into an attention mechanism to output the fusion representation result.

[0019] Optionally, the second fusion unit fuses the vehicle basic information with the fusion representation result by using an attention mechanism, and obtaining the vehicle representation result includes:

[0020] The basic vehicle information is represented and spliced ​​using a representation function to obtain a splicing representation result;

[0021] The concatenated representation result is used as a query vector, and the fused representation result is used as a key vector and a value vector, which are input into an attention mechanism. Then, the output of the attention mechanism is concatenated with the concatenated representation result to obtain the automobile representation result.

[0022] Optionally, in the process of training the prediction model based on the training set, a binary cross entropy loss function is used as a global objective function, and the binary cross entropy loss function is:

[0023]

[0024] in, represents the binary cross entropy loss, Indicates maintenance items The prediction results, For the types of maintenance projects, Represents the transpose of a matrix.

[0025] The beneficial effects of the present invention are:

[0026] The present invention first constructs a co-occurrence matrix of maintenance items to obtain the correlation representation of each maintenance item, then uses the attention mechanism to integrate the correlation representation of historical maintenance items with the mileage to obtain a comprehensive representation result, and finally uses the attention mechanism again to incorporate the basic vehicle information into the final representation result; based on the final representation result, the maintenance status of the car can be comprehensively predicted, safety hazards can be reduced, and traffic safety can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] 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 work.

[0028] Figure 1 The figure is a schematic diagram of the prediction model structure and workflow of an embodiment of the present invention. DETAILED DESCRIPTION

[0029] 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.

[0030] 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.

[0031] This embodiment provides a method for predicting automobile maintenance items based on multi-source data and attention mechanism, including:

[0032] Obtain the historical maintenance records, mileage and basic vehicle information of the target vehicle;

[0033] The historical maintenance item records, mileage and basic vehicle information are input into the preset prediction model to predict the next vehicle maintenance item of the target vehicle. The prediction model is trained through a training set, which includes the historical maintenance item records, mileage and basic vehicle information of several scrapped vehicles. The prediction model is constructed based on a multi-scale deformation attention mechanism.

[0034] Furthermore, the historical maintenance item records include maintenance items and maintenance mileage information;

[0035] Basic vehicle information includes license plate number, manufacturer, type, engine type, engine displacement, model year, year of production, and assembly manufacturer information.

[0036] Furthermore, the prediction model structure and workflow are as follows Figure 1 As shown, the prediction model includes a fusion processing module and a result prediction module. The fusion processing module is used to fuse the historical maintenance project records, mileage and basic vehicle information to obtain the car representation results; the result prediction module is used to predict the next car maintenance project based on the car representation results.

[0037] The fusion processing module includes: a processing unit, a first fusion unit, and a second fusion unit; the processing unit is used to perform correlation learning based on historical maintenance project records to obtain the correlation representation learning results of each historical maintenance project; the first fusion unit is used to use the attention mechanism to fuse the mileage with the correlation representation learning results of each historical maintenance project to obtain the fusion representation results; the second fusion unit is used to use the attention mechanism to fuse the basic vehicle information with the fusion representation results to obtain the vehicle representation results.

[0038] Specifically, in this embodiment, the processing unit performs correlation learning based on the historical maintenance project records, and the method for obtaining the correlation representation learning result of each historical maintenance project is:

[0039] In view of the dynamic development characteristics of automobile maintenance items, the project correlation representation function is used to obtain the correlation representation learning results of each historical maintenance item.

[0040] During the maintenance of a car, multiple maintenance items are required for one maintenance. For example, replacing the oil filter and the air filter will appear in the maintenance record of the car at the same time, indicating that there is correlation between different maintenance items during the maintenance of the car. All maintenance items in the maintenance process of the car are combined to form a global item co-occurrence graph G. This structure reflects the correlation pattern between different maintenance items of the car. The nodes of G are the maintenance items in P. If a maintenance item is coded If both of them appear in the maintenance record of a vehicle, two equal weights will be added to G. and , then calculate Frequency of occurrence in the maintenance records of all vehicles For further calculation of edge weights, here, two weighted edges are used for each pair of nodes, because the influence of two items on each other should not be equal. One item may be a common combination of another item, but not vice versa. In addition, we also want to detect important and common item pairs. Therefore, for , define a threshold Filter out low-frequency combinations and get a qualified combination .make for The total frequency of meeting the conditions, using the adjacency matrix To represent G, The definition of is shown in formula (1):

[0041] (1)

[0042] In this embodiment, It is designed to be asymmetrical to show the different impacts between the two projects. As a static diagram, It measures the frequency of global maintenance items appearing at the same time. However, different maintenance items appear and disappear at different stages. Even if a certain maintenance item does not appear in the current maintenance record, due to the dynamic development process of maintenance items, related maintenance items may appear in the future. Therefore, in order to consider the impact of maintenance item relevance, the historical maintenance items of each vehicle are ,use Get the correlation representation learning result of each historical maintenance project, and take out the relevant fault submatrix of each fault record in A, which is defined as the fault correlation representation learning result , The definition of is shown in formula (2):

[0043] (2)

[0044] Specifically, in this embodiment, the first fusion unit uses the attention mechanism to fuse the mileage and the correlation representation learning results of each historical maintenance project, and the method for obtaining the fusion representation result is:

[0045] In view of the important impact of vehicle mileage on the development process of automobile maintenance projects, the attention mechanism is used to fuse the mileage and project correlation representation learning results.

[0046] The mileage of a car is an important factor related to the condition of the car. Generally speaking, the greater the mileage of a car, the higher the types and frequency of maintenance items that the car needs to undergo. When evaluating car maintenance work, maintenance staff need to first understand the current condition of the car, such as historical maintenance items, mileage, etc. Based on this information, maintenance staff can roughly infer the maintenance items that the car currently needs to undergo. However, due to the differences in the driving patterns of different cars, the mileage for maintenance items of different cars is inconsistent. In response to the above situation, a mileage-aware key query attention mechanism is proposed to learn the key mileage steps in the development process of car failures. It uses the learning result of each item correlation representation as the query vector, and the mileage of each failure as the key vector and value vector.

[0047] Specifically, due to the original mileage Relevant learning results with maintenance items Not in the same latent space, so the original mileage data needs to be use Map to and The same potential space, The specific definition is shown in formula (3):

[0048] (3)

[0049] in, , , , are all parameters. According to the maintenance process of the car, it is determined that the most recent mileage has a greater impact on the maintenance items, and the mileage close to the last time should be activated. In order to achieve the above process, the square operation is used, that is, the square of each element. Only when and When the activation of is close to 0, the corresponding position will be activated. The same potential space Afterwards, each As the query vector, As the key vector and value vector, they are input into the attention mechanism. The specific implementation is shown in formula (4):

[0050] (4)

[0051] The attention mechanism in the above formula is defined as shown in formula (5):

[0052] (5)

[0053] Among them, a is the dimension of attention, , is the weight of attention. For each maintenance item and mileage of the car, the attention mechanism is fused one by one, and the historical mileage of the car is combined to obtain the fused attention representation result. .

[0054] Specifically, in this embodiment, the second fusion unit uses the attention mechanism to fuse the basic vehicle information with the fusion representation result, and obtaining the vehicle representation result includes:

[0055] In view of the impact of basic automobile information on automobile maintenance projects, the basic information is fused and represented, and then the basic information and the fused representation result are fused using the attention mechanism, and finally merged with the basic information to obtain the representation result of the automobile.

[0056] The basic information of a car includes the region, manufacturer, vehicle type, engine type, engine displacement, car model, production year and assembly plant, which are closely related to the development of the car's maintenance projects. The driving environment of cars in different regions is different, which affects the development process of car maintenance projects; different manufacturers and different assembly plants manufacture and assemble different types of cars of different quality, and the development process of car maintenance projects is also different; the car's model year, engine type and displacement and production year also affect the development process of car maintenance projects. Therefore, integrating the above basic car information into the car maintenance project prediction model can effectively improve the accuracy of the prediction. The specific implementation process is as follows;

[0057] First, the basic information Using the representation function and To express , the embedded representation functions and results of each basic information are shown in Table 1. Assume that there is a variable X containing n categories, where the i-th category can be expressed as , Each category is converted into a binary vector of length n, where only one element is 1 and the rest are 0. Specifically, for the i-th category, its The result is shown in formula (6):

[0058] (6)

[0059] The length of the vector is n, the i-th element is 1, and the rest are 0.

[0060] (7)

[0061] exist Based on the above, the discrete features are mapped to a more meaningful low-dimensional space, where is a parameter matrix, and d represents the dimension of the low-dimensional space.

[0062] use and The basic information representation results are concatenated in order. .

[0063] Table 1

[0064]

[0065] In order to integrate the basic information of the car into the maintenance project prediction task, U is used as the query vector, E is used as the key vector and value vector, and input into the attention mechanism to obtain the representation result Q of the basic information of the car U and the historical mileage fusion attention representation E. The specific implementation method is shown in formula (4). Then Q and U are spliced ​​again to obtain the representation result of the car , and maintenance items are predicted based on the expression result H.

[0066] Furthermore, the automobile representation result is input into the result prediction module to predict the next automobile maintenance item.

[0067] Furthermore, the prediction model training includes:

[0068] Collect historical maintenance records, mileage and basic vehicle information of several scrapped cars, build a data set, and randomly divide the data set into a training set, a validation set, and a test set. Save the prediction model that has been trained with the training set and tested with the test set to reach the preset accuracy.

[0069] In the process of training the prediction model, the global objective function uses the binary cross entropy loss function, as shown in formula (8):

[0070] (8)

[0071] in, represents the binary cross entropy loss, Indicates maintenance items The prediction results, For the types of maintenance projects, Represents the transpose of a matrix.

[0072] Next, based on the test set, the prediction model (MsdAtt model) constructed in this embodiment is compared and verified with the existing model. The existing model uses MLP, CNN, RNN, LSTM, RETAIN, Dipole, and Chet.

[0073] Taking R@k as the evaluation index, after multiple experiments, the experimental results are shown in Table 2.

[0074] Table 2

[0075]

[0076] It can be seen that the prediction model proposed in this embodiment is better than all baseline models. Compared with the best baseline RETAIN, when k = 10, k = 20, k = 30, and k = 40, the accuracy of MsdAtt in maintenance item prediction is improved by 1.90%, 2.50%, 2.88%, and 2.50%, respectively.

[0077] 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. Automobile maintenance project prediction method based on multi-source data and attention mechanism, characterized by: include: Obtain the historical maintenance records, mileage and basic vehicle information of the target vehicle; Inputting the historical maintenance item records, mileage and basic vehicle information into a preset prediction model to predict the next vehicle maintenance item of the target vehicle, wherein the prediction model is obtained by training a training set, the training set includes historical maintenance item records, mileage and basic vehicle information of several scrapped vehicles, and the prediction model is constructed based on a multi-scale deformation attention mechanism; The prediction model includes a fusion processing module and a result prediction module. The fusion processing module is used to perform fusion processing on the historical maintenance item records, mileage and basic vehicle information to obtain a vehicle representation result. The result prediction module is used to predict the next vehicle maintenance item based on the vehicle representation result. The fusion processing module includes: a processing unit, a first fusion unit, and a second fusion unit; the processing unit is used to perform correlation learning based on the historical maintenance item records to obtain the correlation representation learning result of each historical maintenance item; the first fusion unit is used to use the attention mechanism to fuse the mileage with the correlation representation learning result of each historical maintenance item to obtain the fusion representation result; the second fusion unit is used to use the attention mechanism to fuse the vehicle basic information with the fusion representation result to obtain the vehicle representation result; The processing unit performs correlation learning based on the historical maintenance item records, and a method for obtaining the correlation representation learning result of each historical maintenance item is as follows: ； ； in, , represents the co-occurrence matrix of maintenance items, m and n are maintenance items m and maintenance items n respectively, is the co-occurrence frequency of maintenance item m and maintenance item n, is the frequency filtering threshold, For the types of maintenance projects, is the total frequency of maintenance item m meeting the requirements, For each historical maintenance project The correlation representation function of ; The first fusion unit fuses the mileage and the correlation representation learning results of each historical maintenance item by using an attention mechanism, and the method for obtaining the fusion representation result is: Mapping the driving mileage to a latent space that is the same as the learning result of the correlation representation of each historical maintenance item; The latent space is used as a query vector, and the correlation representation learning result of each historical maintenance project is used as a key vector and a value vector, which are input into the attention mechanism, and the fusion representation result is output; The second fusion unit fuses the vehicle basic information with the fusion representation result by using an attention mechanism, and obtaining the vehicle representation result includes: The basic vehicle information is represented and spliced ​​using a representation function to obtain a splicing representation result; The concatenated representation result is used as a query vector, and the fused representation result is used as a key vector and a value vector, which are input into an attention mechanism. Then, the output of the attention mechanism is concatenated with the concatenated representation result to obtain the automobile representation result.

2. The automobile maintenance project prediction method based on multi-source data and attention mechanism according to claim 1 is characterized in that: The historical maintenance item records include maintenance items and maintenance mileage information; The basic vehicle information includes the vehicle license plate number, vehicle manufacturer, vehicle type, engine type, engine displacement, vehicle model, production year, and assembly manufacturer information.

3. The automobile maintenance project prediction method based on multi-source data and attention mechanism according to claim 1 is characterized in that: In the process of training the prediction model based on the training set, a binary cross entropy loss function is used as the global objective function, and the binary cross entropy loss function is: ； in, represents the binary cross entropy loss, Indicates maintenance items The prediction results, For the types of maintenance projects, Represents the transpose of a matrix.

Images