Vehicle fault diagnosis and early warning method based on quasi-multi-scale relative principal component analysis

Abstract

The invention discloses a vehicle fault diagnosis and early warning method based on quasi-multi-scale relative principal component analysis. Comprising the following steps: acquiring and preprocessing vehicle-mounted CAN data, performing multi-scale decomposition on historical data, establishing a wavelet decomposition model of a finest scale, performing quasi-relative principal component analysis on each scale to screen wavelet coefficients, establishing a single-scale relative principal component model, performing multi-scale decomposition and relative transformation on online data, and judging whether a fault occurs or not according to a square error. According to the method, a preset expert rule is not needed, non-preset fault types can be recognized, the generalization ability and adaptability of the system are high, online data are processed in real time through the square error to judge the fault occurrence probability, and efficient real-time fault diagnosis is achieved. The discrete wavelet filtering is carried out on the CAN receipts, noise and abnormal values in the vehicle-mounted data can be effectively processed, dependence on high-quality historical data is reduced, and high prediction accuracy can be kept even under the condition that the data is unbalanced.

Description

Technical Field

[0001] The present invention relates to a vehicle fault diagnosis and warning method, specifically to a vehicle fault diagnosis and warning method based on pseudo multi-scale relative principal component analysis. Background Art

[0002] With the improvement of the automation level of modern industrial technology, the structure of the automotive system is becoming more and more complex, and each component is tightly coupled. Any fault may trigger a chain reaction, affecting the normal operation of the whole vehicle, and even may lead to serious traffic accidents. Therefore, the vehicle fault diagnosis and warning technology is very important. The existing vehicle fault diagnosis technologies are mainly based on knowledge bases and deep learning models, and are mostly used for fault classification and analysis of fault causes. Deep learning models usually require a large amount of computing resources and time, and it is difficult to achieve early prediction and prevention. The expert models based on knowledge bases usually require pre-set expert rules or fault entries, and have poor generalization under different models and working conditions, and it is difficult to be directly applied between different systems. In addition, shallow feature extraction methods are difficult to represent the tiny fault features in the vehicle system, resulting in the accuracy of processing tiny fault features being affected.

[0003] The Chinese patent application "Vehicle Fault Diagnosis Method, Device, Vehicle and Storage Medium" with the application number 202311348964.0 obtains the vehicle fault phenomenon and matches it with the expert rules in the knowledge base to obtain the corresponding fault cause, and pushes the fault information to the user terminal. Disadvantages: It is necessary to pre-set expert rules and specific fault phenomena, and it is impossible to identify fault types that are not pre-set, and it is impossible to diagnose and warn before the fault occurs.

[0004] The Chinese patent application "Vehicle Fault Analysis Method and Device, Electronic Equipment and Storage Medium" with the application number 202211639341.4 analyzes the vehicle controller signal and matches it with the preset fault entry table to obtain at least one target fault entry, and sends the fault cause and recommended treatment measures to the vehicle head unit. Disadvantages: It is necessary to pre-set the fault entry table, and it has poor generalization under different models and working conditions, and it belongs to post-mortem diagnosis.

[0005] The Chinese patent application "Vehicle Fault Diagnosis Method, Device, Equipment and Storage Medium" with the application number 202410367317.2 determines the vehicle suspected fault according to the vehicle fault description sent by the user and the fault dictionary library, locates the fault according to the vehicle fault code and the suspected fault, and determines the fault component by analyzing the status quantity of the suspected fault component. Disadvantages: It is necessary to pre-define the fault dictionary library, depends on the fault description actively sent by the user, and has strong data dependence.

[0006] The Chinese patent application "Fault Diagnosis Method, Device, Electronic Device and Storage Medium" with the application number 202310810136.8 solves the probability of each vehicle fault through the regression equation of the fault, and obtains the fault diagnosis result according to the fault probability and the electronic control unit corresponding to the diagnostic fault code. Disadvantages: The regression model belongs to the shallow feature extraction method, which is difficult to represent the tiny fault features in the vehicle system; and the regression equations for different faults are different, and the real-time performance of diagnosing multiple faults simultaneously is poor.

[0007] The Chinese patent application "Automobile Fault Diagnosis Method, System and Intelligent Vehicle Based on Cloud-Edge Collaboration" with the application number 202210962952.6 designs a long short-term memory stacked autoencoder neural network with a new architecture by obtaining the historical state parameters of the electric drive module, and screens the optimal model through the vehicle fault diagnosis model evaluation algorithm. Disadvantages: Deep learning models require a large amount of computing resources and time, and have high requirements for the quality of the historical data for training, and are strongly dependent on data.

[0008] The Chinese patent application "A Vehicle Fault Analysis Method, Device, Equipment and Storage Medium" with the application number 202410485042.2 identifies the fault occurrence time of the vehicle and sends the vehicle operation data in the time periods before and after the fault to the vehicle networking platform for analysis. Disadvantages: This fault diagnosis technology is for after-the-fact diagnosis and cannot achieve early prediction and prevention.

[0009] The Chinese patent application "A Method for Generating a Vehicle Fault Prediction Model, a Fault Prediction Method and a Device" with the application number 202011557530.8 collects the working condition data when the vehicle fails, extracts the training working condition feature vector representation set, and trains to obtain the vehicle fault prediction model. Disadvantages: A large amount of historical fault data needs to be obtained in advance to ensure the model accuracy, and the model is difficult to identify the fault types not included in the training set.

[0010] The Chinese patent application "A Vehicle Fault Monitoring and Warning Method, Equipment and Medium with Multi-Data Fusion" with the application number 202310141592.8 collects vehicle sound data and vehicle driving data through the Internet of Things, converts them into time-frequency data for splicing and fusion to obtain feature vectors, and uses a classifier to obtain the fault classification result. Disadvantages: Collecting data through the Internet of Things requires the deployment of additional hardware resources, and the deep features of the data are not extracted, making it difficult to accurately classify the tiny features.

[0011] In summary, the following deficiencies exist in the prior art:

[0012] 1. Preset rule limitations: Existing methods mostly rely on preset expert rules and fault entry tables. This method cannot identify fault types that are not preset in advance, restricting the generalization ability and adaptability of the system.

[0013] 2. Dependence on active user input: Some methods rely on users to actively provide fault descriptions, which requires a high level of initiative and accuracy from users. If the user's description is inaccurate or the information is not provided in a timely manner, it will directly affect the diagnostic effect.

[0014] 3. Poor real-time performance: The shallow features extracted by existing methods are difficult to represent the tiny fault features in complex vehicle systems. At the same time, the efficiency is low when dealing with multiple faults, and it cannot meet the high-efficiency requirements in practical applications.

[0015] 4. Focus on post-event diagnosis: Existing methods mainly focus on the cause diagnosis after a fault occurs, lacking the ability of early prediction and prevention, and unable to take measures in advance to avoid the expansion of faults.

[0016] 5. Strong data dependence: Existing methods mostly rely on a large amount of high-quality historical data or require additional hardware support, which seriously affects the accuracy of diagnosis and prediction in the case of data imbalance. Summary of the Invention

[0017] Object of the Invention: The object of the present invention is to provide a vehicle fault diagnosis and early warning method based on pseudo multi-scale relative principal component analysis to achieve real-time diagnosis and early warning of vehicle faults.

[0018] The method of the present invention includes: acquiring and preprocessing in-vehicle CAN data, performing multi-scale decomposition on historical data, establishing a wavelet decomposition model at the finest scale, performing pseudo relative principal component analysis on each scale to screen wavelet coefficients, establishing a single-scale relative principal component model, performing multi-scale decomposition and relativization transformation on online data, and judging whether a fault occurs according to the square error.

[0019] Technical Solution: The vehicle fault diagnosis and early warning method based on pseudo multi-scale relative principal component analysis according to the present invention.

[0020] Step 1: Acquire the basic parameters of a vehicle equipped with a CAN device.

[0021] Step 2: Acquire and preprocess CAN bus data.

[0022] Step 3: Based on the Discrete Wavelet Transformation (DWT), filter the normal historical data at different scales. Extract the signals from the fine scale to the coarse scale.

[0023] Step 4: Perform Principal Component Analysis (PCA) on the relativized normal historical data matrix.

[0024] Step 5: Perform relative principal component analysis (RPCA) at each scale to screen wavelet coefficients.

[0025] Step 6: Establish a quasi multi-scale relative principal component analysis model (multi-scale RPCA, MSRPCA).

[0026] Step 7: Perform real-time MSRPCA detection.

[0027] Furthermore, Step 2 is specifically as follows:

[0028] 2.1 Obtain in-vehicle CAN bus data.

[0029] 2.2 Dynamically divide the data into historical data and online data according to the time window granularity.

[0030] 2.3 And divide the historical data obtained in Step 2.2 into normal data and fault data according to the fault code.

[0031] Furthermore, Step 3 is specifically as follows:

[0032] 3.1 Construct a DWT operator, specifically as follows:

[0033]

[0034] Where W Y is the wavelet transform operator matrix; γ i is the wavelet transform coefficient vector.

[0035] 3.2 Quantize the wavelet transform coefficients, specifically as follows:

[0036]

[0037] Where is the observed signal of the i-th variable at the L-th scale of the wavelet transform; the subscripts V and D are the smoothed and corresponding detailed parts of the signal at a coarser scale; N and L are the finest and coarsest scales of the wavelet decomposition.

[0038] 3.3 Transform the data into the multi-scale space to obtain the data matrices at each scale, specifically as follows:

[0039]

[0040] 3.4 Perform relativization transformation on the results obtained in Step 3.3 and respectively, specifically as follows:

[0041]

[0042]

[0043] In the formula, is the wavelet transform coefficient matrix after relativization transformation at the j scale; R is the relativization transformation operator.

[0044] Furthermore, step 4 is specifically as follows:

[0045] 4.1 Establish a PCA model at the N scale, specifically as follows:

[0046]

[0047] In the formula, R i ∈R p×1 is the loading vector; V i ∈R 1×n is the score vector; E is the residual matrix.

[0048] 4.2 Assume that the confidence level of the hypothesis test is α, and determine the control limit α of the squared prediction error (SPE), specifically as follows: j as follows:

[0049]

[0050] Furthermore, step 5 is specifically as follows:

[0051] 5.1 Perform multi-scale decomposition on the online data Y based on step 3.

[0052] 5.2 Based on the established finest-scale PCA model, solve its SPE control limit.

[0053] 5.3 Project the results obtained in step 3.4 at each scale and onto the principal component space defined by the PCA model obtained in step 4.1 to obtain the relative principal component approximate decomposition formulas of and specifically as follows:

[0054]

[0055] In the formula, ν is the number of key principal components; B υ is the matrix composed of the first ν columns of the loading matrix.

[0056] 5.4 Solve the corresponding residual matrix, specifically as follows:

[0057]

[0058] 5.5 Calculate the SPE statistic at time k (k = 1, 2, …, n) online, specifically as follows:

[0059]

[0060] 5.6 Compare the SPE statistic obtained in step 5.5 with the SPE control limit obtained in step 5.2, and select the wavelet coefficient reconstruction that exceeds the control limit Specifically as follows:

[0061]

[0062] Furthermore, step 6 is specifically as follows:

[0063] 6.1 Determine the number v of key principal components of the data matrix obtained in step 5.6 based on RPCA, and solve the loading vector of

[0064] 6.2 Construct a single-scale PCA model of the data matrix specifically as follows:

[0065]

[0066] 6.3 Project onto the principal component subspace and the residual subspace of the reconstructed single-scale PCA model, and calculate the corresponding SPE, specifically as follows:

[0067]

[0068] Furthermore, step 7 is specifically as follows:

[0069] 7.1 Perform multi-scale decomposition and relativization transformation of step 5 on the online data in real time.

[0070] 7.2 Solve the SPE statistic in real time based on step 6.3.

[0071] 7.3 Compare it with the control limit in step 4, judge the probability of the existence of a fault risk, and give an early warning if the confidence level is higher than the threshold.

[0072] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0073] 1. The present invention performs fault diagnosis and early warning through multi-scale decomposition and relative principal component analysis of in-vehicle CAN data, without preset expert rules, can identify fault types that are not preset in advance, and the generalization ability and adaptability of the system are relatively high.

[0074] 2. Through multi-scale decomposition and relativization transformation, the present invention can conduct detailed fault analysis at each scale, ensuring the accurate characterization of tiny fault features. The square error is used to process online data in real time to judge the probability of fault occurrence, realizing efficient real-time fault diagnosis.

[0075] 3. The present invention conducts discrete wavelet filtering on CAN receipts, which can effectively process the noise and outliers in in-vehicle data, reducing the dependence on high-quality historical data and maintaining high prediction accuracy even in the case of unbalanced data. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0077] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be further described below.

[0078] Figure 1 The following shows the specific process of the vehicle fault diagnosis and early warning method based on pseudo multi-scale relative principal component analysis of this embodiment. The following combines Figure 1 to describe the relevant steps in detail:

[0079] Step 1: Obtain the basic parameters of the vehicle equipped with CAN devices.

[0080] Step 2: Obtain and preprocess the CAN bus data, specifically as follows:

[0081] 2.1 Obtain the in-vehicle CAN bus data.

[0082] 2.2 Dynamically divide the data into historical data and online data according to the time window granularity, where the data before 5 minutes belongs to historical data, and the data in the most recent 5 minutes is online data, and it is dynamically updated over time.

[0083] 2.3 Divide the historical data obtained in step 2.2 into normal data and fault data according to the fault codes.

[0084] Step 3: Based on the Discrete Wavelet Transformation (DWT), filter the normal historical data at different scales to extract the signals from the fine scale to the coarse scale, specifically as follows:

[0085] 3.1 Construct the DWT operator, specifically as follows:

[0086]

[0087] In the formula, W Yis the wavelet transform operator matrix; γ i is the wavelet transform coefficient vector.

[0088] 3.2 Quantize the wavelet transform coefficients as follows:

[0089]

[0090] In the formula, is the observed signal of the i-th variable at the wavelet transform coefficient on the L-th scale; the subscripts V and D are the smoothed and corresponding detailed parts of the signal at a coarser scale; N and L are the finest and coarsest scales of wavelet decomposition.

[0091] 3.3 Transform the data into the multi-scale space to obtain the data matrices at each scale as follows:

[0092]

[0093] 3.4 Perform relativization transformation on and respectively as follows:

[0094]

[0095] In the formula, is the wavelet transform coefficient matrix after relativization transformation at the j-th scale; R is the relativization transformation operator.

[0096] Step 4. Perform principal component analysis (PCA) on the relativized normal historical data matrix as follows:

[0097] 4.1 Establish the PCA model at the N-th scale as follows:

[0098]

[0099] In the formula, B i ∈R p×1 is the loading vector; V i ∈R 1×n is the score vector; E is the residual matrix.

[0100] 4.2 Assume that the confidence level of the hypothesis test is α = 95%, and determine the control limit α j of the squared prediction error (SPE) as follows:

[0101]

[0102] Step 5. Perform relative principal component analysis (RPCA) at each scale to screen wavelet coefficients, specifically as follows:

[0103] 5.1 Perform multi-scale decomposition on the online data Y based on Step 3.

[0104] 5.2 Establish the finest-scale PCA model and solve its SPE control limit.

[0105] 5.3 Project the results obtained in Steps 3.4 and onto the principal component space defined by the PCA model obtained in Step 4.1 to obtain the relative principal component approximate decomposition expressions of and , specifically as follows:

[0106]

[0107] where v is the number of key principal components; B υ is the matrix formed by the first v columns of the loading matrix.

[0108] 5.4 Solve the corresponding residual matrix, specifically as follows:

[0109]

[0110] 5.5 Calculate the SPE statistic at time k (k = 1, 2,..., n) online, specifically as follows:

[0111]

[0112] 5.6 Compare the SPE statistic obtained in Step 5.5 with the SPE control limit obtained in Step 5.2, and select the wavelet coefficients that exceed the control limit for reconstruction specifically as follows:

[0113]

[0114] Step 6. Establish a multi-scale relative principal component analysis model (MSRPCA), specifically as follows:

[0115] 6.1 Determine the number of key principal components ν of the data matrix obtained in Step 5.6 based on RPCA, and solve the loading vector

[0116] 6.2 Construct a single-scale PCA model for the data matrix , specifically as follows:

[0117]

[0118] 6.3 Project onto the principal component subspace and the residual subspace of the reconstructed single-scale PCA model, and calculate the corresponding SPE as follows:

[0119]

[0120] Step 7. Perform MSRPCA detection in real time as follows:

[0121] 7.1 Perform the multi-scale decomposition and relativization transformation of Step 5 on the online data in real time.

[0122] 7.2 Solve the SPE statistic in real time based on Step 6.3.

[0123] 7.3 Compare it with the control limit in Step 4 to judge the probability of the risk of failure. If the confidence level is higher than the 90% threshold, give an early warning.

[0124] The above is only the preferred embodiment of the present invention and does not impose any limitation on the present invention. Any person skilled in the art within the technical field of the present invention, without departing from the technical solution of the present invention, makes any form of equivalent replacement or modification and other changes to the technical solution and technical content disclosed by the present invention, all of which belong to the content of not departing from the technical solution of the present invention and still fall within the protection scope of the present invention.

Claims

1. A vehicle fault diagnosis and early warning method based on quasi-multiscale relative principal component analysis, characterized in that: The steps include: Step 1: Obtain basic parameters of a vehicle equipped with a CAN device; Step 2: Acquire and preprocess CAN bus data; Step 3: Use different scales to transform normal historical data based on discrete wavelet transform Perform filtering to extract signals from fine scale to coarse scale; Step 4: Relativize the normal historical data array Conduct principal component analysis; Step 5: Perform relative principal component analysis at each scale to screen wavelet coefficients; Step 6: Establish a quasi-multiscale relative principal component analysis model; Step 7: Perform MSRPCA detection in real time.

2. The vehicle fault diagnosis and early warning method based on quasi-multiscale relative principal component analysis according to claim 1 is characterized in that: The method of step 2: 2.1 Obtain vehicle CAN bus data; 2.2 Data is dynamically divided into historical data and online data according to the time window granularity, of which data before 5 minutes is historical data, and data within the last 5 minutes is online data, which is dynamically updated over time; 2.3 The historical data obtained in step 2.2 is divided into normal data and fault data according to the fault code.

3. The vehicle fault diagnosis and early warning method based on quasi-multiscale relative principal component analysis according to claim 2 is characterized in that: The method of step 3: 3.1 Construct the DWT operator as follows: Where W Y is the wavelet transform operator matrix; γ i is the wavelet transform coefficient vector, 3.2 Quantize the wavelet transform coefficients, as follows: In the formula, is the observed signal of the i-th variable The wavelet transform coefficients at the Lth scale; the subscripts V and D are the smoothing and corresponding details of the signal at a coarser scale; N and L are the finest and coarsest scales of the wavelet decomposition; 3.3 Convert the data to multi-scale space, and obtain the data arrays at each scale, as follows: 3.4 Step 3.3 and Relativize the transformations respectively, as follows: In the formula, is the wavelet transform coefficient matrix after relativization transformation on scale j; R is the relativization transformation operator.

4. The vehicle fault diagnosis and early warning method based on quasi-multi-scale relative principal component analysis according to claim 3 is characterized in that: The method of step 4: 4.1 Establish a PCA model on the N scale, as follows: In the formula, B i ∈R p×1 is the load vector; V i ∈R 1×n is the score vector; E is the residual matrix; 4.2 Assume that the confidence level of the test is α and determine the control limit α of the squared prediction error j , as follows:

5. The vehicle fault diagnosis and early warning method based on quasi-multiscale relative principal component analysis according to claim 4 is characterized in that: The method of step 5: 5.1 Based on step 3, perform multi-scale decomposition on the online data Y; 5.2 Based on establishing the finest scale PCA model and solving its SPE control limits; 5.3 Apply the results from step 3.4 to each scale and Projecting onto the principal component space defined by the PCA model obtained in step 4.1, we obtain and The relative pivot approximate decomposition formula is as follows: In the formula, υ is the number of key components; B υ is the matrix consisting of the first v columns of the loading matrix; 5.4 Solve the corresponding residual matrix as follows: 5.5 Online calculation of SPE statistics at time k (k = 1, 2, ..., n), as follows: 5.6 Compare the SPE statistics obtained in step 5.5 with the SPE control limits obtained in step 5.2, and select the wavelet coefficients that exceed the control limits to reconstruct The details are as follows:

6. The vehicle fault diagnosis and early warning method based on quasi-multi-scale relative principal component analysis according to claim 5, characterized in that the method of step 6: 6.1 Determine the data array obtained in step 5.6 based on RPCA The number of key principal components ν, solving the load vector 6.2 Constructing the data array The single-scale PCA model is as follows: 6.3 Project to the principal component subspace and residual subspace of the reconstructed single-scale PCA model and calculate the corresponding SPE as follows:

7. The vehicle fault diagnosis and early warning method based on quasi-multi-scale relative principal component analysis according to claim 6 is characterized in that: The method of step 7: 7.1 Perform multi-scale decomposition and relativization transformation of step 5 on online data in real time; 7.2 Solve the SPE statistics in real time based on step 6.3; 7.3 and compared with the control limits in step 4 to determine the probability of failure risk, and issue a warning if the confidence level is higher than the threshold.

Images