Vehicle maintenance reimbursement behavior abnormality detection method, device and storage medium

By combining isolated forest model and principal component analysis technology, abnormal detection of vehicle maintenance data is solved, and the problem of inefficient supervision in the existing technology is achieved, and efficient and accurate identification and supervision of abnormal behaviors is achieved.

CN110852860BActive Publication Date: 2024-11-08BEIJING YOUTU HAOCHENG AUTOMOBILE TECHNOLOGY DEVELOPMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201911116659.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2019-11-15
Publication Date
2024-11-08
Estimated Expiration
2039-11-15

AI Technical Summary

Technical Problem

It is difficult for the existing technology to effectively supervise and detect abnormal situations in vehicle maintenance and reimbursement behavior, resulting in inefficient supervision and property losses.

Method used

The fusion method of isolated forest model and principal component analysis (PCA) model is adopted to pre-process vehicle maintenance data, feature column combination, abnormality detection and filter, and abnormal behavior is automatically identified through unsupervised machine learning methods to avoid manual labeling and errors.

Benefits of technology

It improves the accuracy and efficiency of abnormal detection, enhances the analyticity of detection results, reduces the need for manual review, and improves regulatory efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN110852860B_ABST
    Figure CN110852860B_ABST
Patent Text Reader

Abstract

A method for detecting abnormal vehicle maintenance reimbursement behavior comprises the following steps: S1, preprocessing the collected vehicle maintenance data, and performing influencing factor elimination processing on the preprocessed data to obtain a first feature column group and a second feature column group; S2, performing isolation forest model and principal component analysis (PCA) model processing on the data in the first feature column group and the second feature column group respectively, and performing model fusion and intersection processing on the processing results to obtain a model fusion result, and sending the model fusion result to a business self-filtering module for filtering to obtain a vehicle ID result set with abnormal reimbursement behavior.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of big data processing technology, and in particular to a method, device and storage medium for detecting abnormal vehicle maintenance reimbursement behavior. Background Art

[0002] Vehicle maintenance reimbursement mainly refers to the maintenance of the company group or unit's public vehicles during daily driving. The maintenance records are imported into the database to form maintenance reimbursement data. As the amount of data increases, vehicle reimbursement management personnel cannot intuitively judge whether there are abnormalities in the vehicle reimbursement behavior. Driven by interests, there will be some illegal reimbursement behaviors such as false reporting, which will cause certain losses to the company or unit's property. The current supervision method is to judge and review through artificial sensory perception. It is often impossible to hold accountable and supervise due to the lack of strong evidence, resulting in low supervision efficiency. Summary of the invention

[0003] In view of this, the present invention provides a method, device and storage medium that can accurately and efficiently detect abnormal vehicle maintenance reimbursement behavior.

[0004] A method for detecting abnormal behavior of vehicle maintenance reimbursement comprises the following steps:

[0005] S1. Preprocessing the collected vehicle maintenance data, and performing influencing factor elimination processing on the preprocessed data to obtain a first feature column group and a second feature column group;

[0006] S2. Perform isolation forest model and principal component analysis (PCA) model processing on the data in the first feature column group and the second feature column group respectively, and perform model fusion and intersection processing on the processing results to obtain the model fusion result, and send the model fusion result to the business self-filtering module for filtering to obtain the vehicle ID result set with abnormal reimbursement behavior.

[0007] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0008] The step S2 also includes: S3, outputting the vehicle ID result set with abnormal reimbursement behavior to the result visualization module for visualization processing and then displaying.

[0009] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0010] In step S1, the pre-processed data is subjected to an influence factor elimination process to obtain a first feature column group and a second feature column group, which includes:

[0011] S11, for each vehicle, statistics are given on the values ​​of each feature column based on the vehicle type, ownership unit, service life, and repair shop combined segmentation dimensions;

[0012] S12, converting the vehicle type, the affiliated unit, the service life, and the repair shop joint segmentation dimensions into dummy variable columns; if the affiliated unit dimension is included, jump to step S13, otherwise jump to step S14;

[0013] S13, for each feature column obtained in S11, take the feature column as the dependent variable, and the dummy variable column corresponding to the vehicle type, the affiliated unit, the service life, and the repair shop obtained in S12 as the independent variable, establish a linear regression equation, and take the residual column after fitting to replace the dependent variable feature column, to obtain the first feature column group;

[0014] S14. For each feature column obtained in step S11, take the feature column as the dependent variable, and the dummy variable column corresponding to the vehicle type, age, and repair shop obtained in step S12 as the independent variable, establish a linear regression equation, and take the residual column after fitting to replace the dependent variable feature column to obtain the second feature column group.

[0015] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0016] In step S2, performing isolation forest model processing on the first feature column group and the second feature column group respectively includes:

[0017] The feature column group is an N-dimensional vehicle repair reimbursement feature vector data set. Uniform sampling without replacement is performed to select m samples from n data in the feature column group data set as training samples for building the first decision tree.

[0018] In the m samples, a vehicle maintenance reimbursement feature is randomly selected, and a value is randomly selected from the range of all values ​​of this vehicle maintenance reimbursement feature. The samples are divided into a binary tree, and the samples with values ​​less than the value are divided to the left of the node, and the samples with values ​​greater than or equal to the value are divided to the right of the node, thereby completing a data division, obtaining a splitting condition and the left and right data sets, and then repeating the above steps on the left and right data sets until the preset termination condition is met;

[0019] Repeat the above steps until t decision trees are obtained, and use the generated isolation forest model to evaluate the test data and the newly added sample data to determine the abnormal vehicle ID information.

[0020] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0021] The generated isolation forest model is used to evaluate the test data and the newly added sample data including:

[0022] A training sample data x is made to traverse each decision tree of the above process, and then calculate the layer of each decision tree where x finally falls, that is, the path length h(x) of the sample point x is the number of edges passed from the root node to the leaf node of the decision tree; then the average height of x in each tree E(h(x)) can be obtained;

[0023] The model output anomaly score calculation formula is as follows:

[0024]

[0025] Where c(n) is a dataset of n samples, and the average path length of the tree is calculated as follows:

[0026]

[0027] The closer the anomaly score output by the model training is to 1, the more likely the data is anomaly, and the closer it is to 0, the more likely the data is normal. If the anomaly score of the entire data sample is around 0.5, it means that there are no abnormal samples in the data sample set.

[0028] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0029] The principal component analysis (PCA) model processing of the data in the first feature column group and the second feature column group in S2 includes:

[0030] The feature column group is an N-dimensional vehicle maintenance reimbursement feature vector data set, and the vehicle maintenance reimbursement features in the N-dimensional vehicle maintenance reimbursement feature vector data set are subjected to dimensionality reduction processing, so that the N-dimensional features are presented in the form of vectors in space, and then projected so that the variance of the projected data is maximized, thereby minimizing the information loss after the projection;

[0031] The data after dimensionality reduction is thresholded, and the threshold is set according to the TOP principle of the data value after dimensionality reduction. After the threshold is set, the abnormal vehicle ID information determined by the PCA model is output.

[0032] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0033] The N-dimensional vehicle maintenance reimbursement features are reduced to 3 dimensions, and each dimension of data represents a business direction, forming a semantic output result.

[0034] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0035] In step S2, the isolation forest model results of the first feature column group and the second feature column group are respectively fused with the principal component analysis PCA model results, and the fusion results are then fused with intersection and union to output the fusion results.

[0036] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0037] The model fusion result is sent to the business self-filtering module for filtering in step S2, including:

[0038] Determine whether the proportion of annual vehicle consumption corresponding to the vehicle ID information to total consumption is greater than a first preset value, and if so, confirm it as abnormal vehicle ID information; otherwise, further determine whether the proportion of parts consumption to total consumption is greater than a second preset value, and if so, confirm it as abnormal vehicle ID information; otherwise, further determine whether the number of parts is greater than a third preset value, and if so, confirm it as abnormal vehicle ID information; otherwise, confirm it as normal vehicle ID information.

[0039] The present invention also provides a computer device, comprising:

[0040] Memory for storing computer programs;

[0041] A processor is used to execute the computer program to implement the steps of any of the above methods.

[0042] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of any of the above-mentioned methods.

[0043] Beneficial technical effects: Compared with the prior art, the vehicle maintenance reimbursement behavior anomaly detection method, device and storage medium of the present invention adopts anomaly detection by integrating isolation forest and PCA models, avoiding the errors caused by anomaly detection using the isolation forest model alone; the PCA semantic parsing link is added to the unsupervised machine learning method to enhance the parsability of the detection results; and the unsupervised machine learning method is adopted to avoid manual labeling work. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of a method for detecting abnormal vehicle maintenance reimbursement behavior provided by an embodiment of the present invention;

[0045] Figure 2 It is a schematic diagram of data classification at the model input end;

[0046] Figure 3 It is a flow chart of model construction;

[0047] Figure 4 It is a visualization of PCA dimensionality reduction data;

[0048] Figure 5 This is a schematic diagram of business self-filtering rules. DETAILED DESCRIPTION

[0049] like Figure 1 As shown, in an embodiment of the present invention, a method for detecting abnormal vehicle maintenance reimbursement behavior includes the following steps:

[0050] S1. Preprocessing the collected vehicle maintenance data, and performing influencing factor elimination processing on the preprocessed data to obtain a first feature column group and a second feature column group;

[0051] Optionally, preprocessing the collected vehicle maintenance data includes performing ETL (extraction, transformation, loading) processing on the vehicle maintenance reimbursement data.

[0052] Taking into account the differences between vehicles, such as vehicle type, age, repair shop and affiliated unit, direct data comparison calculation will cause serious deviations in the results, so it is necessary to eliminate the influencing factors of the data.

[0053] S2. Perform isolation forest model and principal component analysis (PCA) model processing on the data in the first feature column group and the second feature column group, respectively, perform model fusion and intersection processing on the processing results to obtain the model fusion result, and send the model fusion result to the business self-filtering module for filtering to obtain the vehicle ID result set with abnormal reimbursement behavior. Performing isolation forest model and principal component analysis (PCA) model processing on the data in the first feature column group and the second feature column group respectively means performing isolation forest module processing on the data in the first feature column group and the second feature column group; performing principal component analysis (PCA) model processing on the data in the first feature column group and the second feature column group.

[0054] In view of the different frequencies of vehicle use by the units or departments to which the vehicles belong, and the possible overall violation of regulations by the units or departments to which the vehicles belong, the input of the model is classified into two data sets, one with the influencing factors of the units being eliminated and the other with the influencing factors of the units not being eliminated.

[0055] Eliminating the attribution unit can avoid the difference between the units or departments with high frequency of vehicle use and the units or departments with low frequency of vehicle use. Not eliminating the attribution unit can allow the model to locate the overall violation of the attribution unit or department. The vivid description is as follows: Figure 2 shown.

[0056] The overall idea of ​​model construction in step S2 is as follows Figure 3As shown in the figure, the core model construction idea is to fuse the classified data at the input end with the isolation forest model and the PCA (principal component analysis) model respectively. After processing, the model fusion results enter the business self-filtering module and are finally output to the result visualization module.

[0057] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0058] The step S2 also includes: S3, outputting the vehicle ID result set with abnormal reimbursement behavior to the result visualization module for visualization processing and then displaying.

[0059] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0060] In step S1, the pre-processed data is subjected to an influence factor elimination process to obtain a first feature column group and a second feature column group, which includes:

[0061] S11, for each vehicle, statistics are given on the values ​​of each feature column based on the vehicle type, ownership unit, service life, and repair shop combined segmentation dimensions;

[0062] S12, converting the vehicle type, the affiliated unit, the service life, and the repair shop joint segmentation dimensions into dummy variable columns; if the affiliated unit dimension is included, jump to step S13, otherwise jump to step S14;

[0063] S13, for each feature column obtained in S11, take the feature column as the dependent variable, and the dummy variable column corresponding to the vehicle type, the affiliated unit, the service life, and the repair shop obtained in S12 as the independent variable, establish a linear regression equation, and take the residual column after fitting to replace the dependent variable feature column, to obtain the first feature column group;

[0064] S14. For each feature column obtained in step S11, take the feature column as the dependent variable, and the dummy variable column corresponding to the vehicle type, age, and repair shop obtained in step S12 as the independent variable, establish a linear regression equation, and take the residual column after fitting to replace the dependent variable feature column to obtain the second feature column group.

[0065] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0066] In step S2, performing isolation forest model processing on the first feature column group and the second feature column group respectively includes:

[0067] The feature column group is an N-dimensional vehicle repair reimbursement feature vector data set. In the data set of the feature column group, m samples are uniformly sampled without replacement to serve as training samples for building the first decision tree; the feature column groups here are the first feature column group and the second feature column group.

[0068] The input data of the isolation forest model is an N-dimensional vehicle repair reimbursement feature vector data set. Assuming that the data set has a total of n data, when building the first decision tree, m samples are uniformly sampled without replacement from the n data in the data set, and these samples are used as training samples for this decision tree.

[0069] Among the m samples, a vehicle maintenance reimbursement feature is randomly selected, and a value is randomly selected within the range of all values ​​of this vehicle maintenance reimbursement feature (between the minimum and maximum values), and the samples are divided into a binary tree, and the samples with values ​​less than the value are divided to the left side of the node, and the samples with values ​​greater than or equal to the value are divided to the right side of the node, thereby completing a data partition, obtaining a splitting condition and the data sets on the left and right sides, and then repeating the above steps on the data sets on the left and right sides until the preset termination condition is met; optionally, there are two termination conditions, one is that the data set itself cannot be divided any further (only one sample is included, or all samples are the same), and the other is that the height of the tree reaches log2(m). Limiting the height of the tree is mainly to consider the efficiency of the algorithm and to avoid falling into a state of continuous partitioning due to excessive data volume.

[0070] Repeat the above steps until t decision trees are obtained, and use the generated isolation forest model to evaluate the test data and the newly added sample data to determine the abnormal vehicle ID information.

[0071] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0072] The generated isolation forest model is used to evaluate the test data and the newly added sample data including:

[0073] A training sample data x is made to traverse each decision tree of the above process, and then calculate the layer of each decision tree where x finally falls, that is, the path length h(x) of the sample point x is the number of edges passed from the root node to the leaf node of the decision tree; then the average height of x in each tree E(h(x)) can be obtained;

[0074] The model output anomaly score calculation formula is as follows:

[0075]

[0076] Where c(n) is a dataset of n samples, and the average path length of the tree is calculated as follows:

[0077]

[0078] The closer the anomaly score output by the model training is to 1, the more likely the data is anomaly, and the closer it is to 0, the more likely the data is normal. If the anomaly score of the entire data sample is around 0.5, it means that there are no abnormal samples in the data sample set.

[0079] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0080] The principal component analysis (PCA) model processing of the data in the first feature column group and the second feature column group in S2 includes:

[0081] The feature column group is an N-dimensional vehicle maintenance reimbursement feature vector data set. The vehicle maintenance reimbursement features in the N-dimensional vehicle maintenance reimbursement feature vector data set are reduced in dimension, so that the N-dimensional features are presented in the form of vectors in space, and then projected to maximize the variance of the projected data, thereby minimizing the information loss after projection; the feature column groups here are the first feature column group and the second feature column group.

[0082] The input data of the PCA (Principal Component Analysis) model is also an N-dimensional vehicle maintenance reimbursement feature vector data set. It mainly reduces the dimension of the N-dimensional vehicle maintenance reimbursement features, presents the N-dimensional features in the form of vectors in space, and then projects them. To minimize the information loss after projection, it is necessary to maximize the variance of the projected data (the looser the data).

[0083] The data after dimensionality reduction is thresholded, and the threshold is set according to the TOP principle of the data value after dimensionality reduction. After the threshold is set, the abnormal vehicle ID information determined by the PCA model is output. PCA dimensionality reduction data visualization is as follows Figure 4 shown.

[0084] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0085] The N-dimensional vehicle maintenance reimbursement features are reduced to 3 dimensions, and each dimension of data represents a business direction, forming a semantic output result.

[0086] Optionally, in the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0087] In step S2, the isolation forest model results of the first feature column group and the second feature column group are respectively fused with the principal component analysis PCA model results, and the fusion results are then fused with intersection and union to output the fusion results.

[0088] In the vehicle maintenance reimbursement behavior abnormality detection method of the present invention,

[0089] The model fusion result is sent to the business self-filtering module for filtering in step S2, including:

[0090] Determine whether the proportion of the annual consumption of the vehicle corresponding to the vehicle ID information to the total consumption is greater than a first preset value. If so, confirm it as abnormal vehicle ID information. Otherwise, further determine whether the proportion of parts consumption to the total consumption is greater than a second preset value. If so, confirm it as abnormal vehicle ID information. Otherwise, further determine whether the number of parts is greater than a third preset value. If so, confirm it as abnormal vehicle ID information. Otherwise, confirm it as normal vehicle ID information. In order to further improve the recall rate of abnormal detection results of vehicle maintenance reimbursement behavior, this application proposal designs a business rule self-filtering module, which mainly performs further filtering operations on the model output result set to avoid the phenomenon of "seemingly abnormal, but actually normal". Optionally, the filtering rules can be as follows: Figure 5 shown.

[0091] The present invention also provides a computer device, comprising:

[0092] Memory for storing computer programs;

[0093] A processor is used to execute the computer program to implement the steps of any of the above methods.

[0094] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of any of the above-mentioned methods.

[0095] Beneficial technical effects: Compared with the prior art, the vehicle maintenance reimbursement behavior anomaly detection method, device and storage medium of the present invention adopts anomaly detection by integrating isolation forest and PCA models, avoiding the errors caused by anomaly detection using the isolation forest model alone; the PCA semantic parsing link is added to the unsupervised machine learning method to enhance the parsability of the detection results; and the unsupervised machine learning method is adopted to avoid manual labeling work.

[0096] It is understandable that a person skilled in the art can make other corresponding changes and modifications based on the technical concept of the present invention, and all these changes and modifications should fall within the protection scope of the claims of the present invention.

Claims

1. A method for detecting abnormal behavior of vehicle maintenance reimbursement, characterized in that: The steps include: S1. Preprocessing the collected vehicle maintenance data, and performing influencing factor elimination processing on the preprocessed data to obtain a first feature column group and a second feature column group; S2. Perform isolation forest model and principal component analysis (PCA) model processing on the data in the first feature column group and the second feature column group respectively, perform model fusion and intersection processing on the processing results to obtain model fusion results, and send the model fusion results to the business self-filtering module for filtering to obtain the vehicle ID result set with abnormal reimbursement behavior; S3, outputting the result set of vehicle IDs with abnormal reimbursement behavior to the result visualization module for visualization and display; In step S1, the pre-processed data is subjected to an influence factor elimination process to obtain a first feature column group and a second feature column group, which includes: S11, for each vehicle, statistics are given on the values ​​of each feature column based on the vehicle type, ownership unit, service life, and repair shop combined segmentation dimensions; S12, converting the vehicle type, the affiliated unit, the service life, and the repair shop joint segmentation dimensions into dummy variable columns; if the affiliated unit dimension is included, jump to step S13, otherwise jump to step S14; S13, for each feature column obtained in S11, take the feature column as the dependent variable, and the dummy variable column corresponding to the vehicle type, the affiliated unit, the service life, and the repair shop obtained in S12 as the independent variable, establish a linear regression equation, and take the residual column after fitting to replace the dependent variable feature column, to obtain the first feature column group; S14, for each feature column obtained in step S11, take the feature column as the dependent variable, and the dummy variable column corresponding to the vehicle type, service life, and repair shop obtained in step S12 as the independent variable, establish a linear regression equation, and take the residual column after fitting to replace the dependent variable feature column, to obtain the second feature column group; In step S2, performing isolation forest model processing on the first feature column group and the second feature column group respectively includes: The feature column group is an N-dimensional vehicle repair reimbursement feature vector data set. Uniform sampling without replacement is performed to select m samples from n data in the feature column group data set as training samples for building the first decision tree. Among the m samples, a vehicle maintenance reimbursement feature is randomly selected, and a value is randomly selected from the range of all values ​​of this vehicle maintenance reimbursement feature. The samples are divided into a binary tree, and the samples with values ​​less than the value are divided to the left of the node, and the samples with values ​​greater than or equal to the value are divided to the right of the node, thereby completing a data division, obtaining a splitting condition and the left and right data sets, and then repeating the above steps on the left and right data sets until the preset termination condition is met; Repeat the above steps until t decision trees are obtained, and use the generated isolation forest model to evaluate the test data and the newly added sample data to determine the abnormal vehicle ID information; The generated isolation forest model is used to evaluate the test data and the newly added sample data including: A training sample data x is made to traverse each decision tree of the above process, and then calculate the layer of each decision tree where x finally falls, that is, the path length h(x) of the sample point x is the number of edges passed from the root node to the leaf node of the decision tree; then the average height of x in each tree E(h(x)) can be obtained; The model output anomaly score calculation formula is as follows: Where c(n) is a dataset of n samples, and the average path length of the tree is calculated as follows: The closer the anomaly score output by model training is to 1, the greater the possibility of data anomaly, and the closer it is to 0, the greater the possibility of data normality. If the anomaly score of the entire data sample is around 0.5, it means that there are no abnormal samples in the data sample set; The principal component analysis (PCA) model processing of the data in the first feature column group and the second feature column group in S2 includes: The feature column group is an N-dimensional vehicle maintenance reimbursement feature vector data set, and the vehicle maintenance reimbursement features in the N-dimensional vehicle maintenance reimbursement feature vector data set are subjected to dimensionality reduction processing, so that the N-dimensional features are presented in the form of vectors in space, and then projected so that the variance of the projected data is maximized, thereby minimizing the information loss after the projection; The data after dimension reduction is thresholded, and the threshold is set according to the TOP principle of the data value after dimension reduction. After the threshold is set, the abnormal vehicle ID information determined by the PCA model is output; Reduce the N-dimensional vehicle maintenance reimbursement features to 3 dimensions, with each dimension representing a business direction, to form a semantic output result; The model fusion result is sent to the business self-filtering module for filtering in step S2, including: Determine whether the proportion of annual vehicle consumption corresponding to the vehicle ID information to total consumption is greater than a first preset value, and if so, confirm it as abnormal vehicle ID information; otherwise, further determine whether the proportion of parts consumption to total consumption is greater than a second preset value, and if so, confirm it as abnormal vehicle ID information; otherwise, further determine whether the number of parts is greater than a third preset value, and if so, confirm it as abnormal vehicle ID information; otherwise, confirm it as normal vehicle ID information.

2. A computer device, characterized in that: include: Memory for storing computer programs; A processor is used to execute the computer program to implement the steps of the method according to claim 1.

3. 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 claim 1 are implemented.