A method for onboard quality testing of commercial vehicles

Abstract

This invention discloses a method for detecting the on-board quality of commercial vehicles, comprising the following steps: Step S1, data preprocessing, including data parsing, cleaning, and noise reduction of the original vehicle terminal data; Step S2, operating condition identification and analysis, further analyzing the operating conditions of the preprocessed data; Step S3, data filtering, determining whether the data from the operating condition analysis belongs to test data or training data. If it belongs to test data, proceed to Step S5; if it belongs to training data, proceed to Step S4; Step S4, model construction and fusion, constructing multiple different basic models using the training data from Step S2, and finally fusing the models through training and filtering. This invention enables real-time monitoring of the on-board quality of commercial vehicles without adding additional hardware, ensuring the accuracy of the results while reducing the instability of the results caused by the randomness of statistical data.

Description

Technical Field

[0001] This invention belongs to the field of intelligent vehicle networking and logistics transportation technology. Specifically, this invention relates to a method for on-board quality detection of commercial vehicles. Background Technology

[0002] The vehicle's overall load weight has a significant impact on energy consumption and vehicle wear and tear. Severe overloading poses a major safety hazard. With the current development of vehicle intelligence, the vehicle's overall load weight is a fundamental data point for the application of some new technologies. How to monitor the load weight of customer vehicles in real time is a problem faced by various car manufacturers and also a problem faced by the country in terms of controlling overloading.

[0003] Currently, there are four real-time monitoring principles and methods for vehicle mass: strain principle, displacement principle, air pressure principle, and dynamic principle. Some measuring devices have also begun to appear on the market for data collection on vehicles. This method requires the installation of additional hardware sensors, which is costly and has high subsequent maintenance costs, and it also affects the measurement accuracy. The dynamic principle, due to its high accuracy, ability to perform dynamic measurements, and long lifespan, has become a hot topic and trend in recent years.

[0004] However, current solutions based on dynamic principles all use a single model for prediction. The problem with this approach is that the accuracy of the model prediction results is very unstable, making it difficult to apply to scenarios with relatively strict load prediction requirements. Summary of the Invention

[0005] This invention provides a method for detecting the on-board mass of commercial vehicles, which solves the problems mentioned in the background section.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for detecting the on-board mass of commercial vehicles, comprising the following steps.

[0007] Step S1, data preprocessing, involves data parsing, cleaning, and noise reduction of the original vehicle terminal data.

[0008] Step S2, Working condition identification and analysis: Further working condition analysis is performed on the preprocessed data.

[0009] Step S3: Data filtering. Determine whether the data for the working condition analysis belongs to test data or training data. If it belongs to test data, proceed to step S5; if it belongs to training data, proceed to step S4.

[0010] Step S4, Model Building and Fusion: Multiple different base models are built using the training data from step S2, and then fused after training and selection.

[0011] Step S5, Model Service Deployment: Deploy the trained model to the cloud server, build a Docker image using the trained model and model service code, and push it to the vehicle networking Docker image repository.

[0012] Preferably, step S1 includes the following steps:

[0013] Step S11, data parsing: Parse the original vehicle terminal data to extract the vehicle data required for modeling;

[0014] Step S12, data filtering, performs operations such as filling, alignment, interpolation and smoothing on the parsed data, and judges whether the data conforms to the dynamic principle based on conditions such as clutch and brake. If so, proceed to step S13; otherwise, send it back to the vehicle's original terminal.

[0015] Step S13, anomaly detection: determine whether the data is abnormal. If so, proceed to S14; otherwise, send it back to the vehicle's original terminal. The detection method is as follows:

[0016] First, rule-based methods mainly rely on the actual meaning of each attribute to make judgments, such as the speed cannot be negative and the torque cannot exceed a certain range.

[0017] 2. Anomaly samples are removed using the isolated forest model, which is commonly used in anomaly detection in model algorithms, and filtering is performed based on the center point peak.

[0018] S14, shift detection: based on the factors affecting vehicle load obtained from the dynamic formula, determine whether the data is abnormal. If so, proceed to S2; otherwise, send it back to the vehicle's original terminal.

[0019] Preferably, step S2 includes the following steps:

[0020] Step S21: Set the conditions for operating condition filtering: torque output percentage must be greater than 10%, engine speed must be greater than idle speed, and brake signal must be 0.

[0021] Step S22: Generate time periods for various defined working conditions, filter them based on acceleration, identify acceleration and deceleration time periods, and identify clutch engagement time periods based on clutch signals;

[0022] Step S23: Combine the time of various working conditions, select the effective working condition time, and select the basic working condition time period that conforms to the dynamic principle of not pressing the clutch and not pressing the brake under the working condition filtering conditions, and use it as the preliminary input data for the next step of model construction.

[0023] Preferably, step S4 includes the following steps:

[0024] Step S41, Model creation: Construct multiple different base models using the training data from step S2;

[0025] Step S42, Model Training: The selected basic regression model is trained based on the prepared model training sample set. The model parameters are optimized and selected to generate multiple basic models for vehicle weight prediction. The performance of the trained vehicle weight prediction model is then evaluated using the test sample set, for example, the load assessment model.

[0026] Step S43: Model screening. Determine whether the obtained model meets the fusion requirements. If yes, proceed to S44; otherwise, repeat step S42.

[0027] Step S44, Model fusion, which involves fusing the multiple better models obtained above.

[0028] Preferably, the construction of multiple different base models in step S41 can be carried out in the following ways:

[0029] 1. Filtering by gear position;

[0030] 2. Data filtering through triangular domain transformation.

[0031] Preferably, the gear filter design is as follows:

[0032] 1) Statistical distribution density of vehicle speed to engine speed ratio;

[0033] 2) Peak detection is performed using a distribution density map. The location of the peak is the center point of the gear speed-to-rev ratio.

[0034] 3) While performing peak detection, output the peak width to obtain the upper and lower boundaries of the peak width, which are the upper and lower boundaries of the vehicle speed-to-speed ratio for each gear.

[0035] 4) Filter data based on the upper and lower boundaries of the vehicle speed to engine speed ratio. Data within the upper and lower boundaries is retained, while data outside the upper and lower boundaries is filtered out.

[0036] Preferably, the triangular domain transformation data filtering design is as follows:

[0037] 1) Calculate the ratio between torque and acceleration, and then solve for its arctangent function;

[0038] 2) Statistical kernel density distribution;

[0039] 3) Theoretically, the calculated ratio coefficient corresponding to the load must have a high distribution density, so peak detection is used to detect the peak point and peak width.

[0040] 4) Use the upper limit boundary of the peak width for data filtering.

[0041] Preferably, the model selection method in step S43 is as follows:

[0042] 1) The coefficient C is obtained by calculating the slope of the characteristic torque and acceleration related to the load;

[0043] 2) Use this coefficient C to rotate the coordinate axes;

[0044] 3) Calculate the single standard deviation of the newly generated coordinates X and Y respectively, and divide the single standard deviation of Y by the single standard deviation of X to obtain Xyscore.

[0045] Preferably, the specific implementation algorithm of Xyscore

[0046] 1) Calculate the intermediate coefficients coef for the input data (DF0, DF20) of each constructed base model.

[0047] 2) Use this coefficient (coef) to calculate the new coordinate axis. The calculation formula is as follows:

[0048] X=cos(arctan(coef))*DF0+sin(arctan(coef))*DF20;

[0049] Y=cos(arctan(coef))*DF20-sin(arctan(coef))*DF0;

[0050] 3) Calculate the single standard deviation of the newly generated coordinates X and Y respectively, and divide the single standard deviation of Y by the single standard deviation of X to obtain the XY score.

[0051] Preferably, the underlying principle of this solution is based on dynamic formulas:

[0052] F t =F f +F w +F i +F j

[0053] Where F t F represents traction force. f F represents rolling resistance. w F represents air resistance. i F represents slope resistance. j Representing inertial force, the formula can be approximately expanded as follows:

[0054]

[0055] Where f is the rolling resistance coefficient, g is the acceleration due to gravity, and C is the acceleration due to gravity. DWhere A is the air resistance coefficient, ρ is the frontal area, m is the vehicle mass, and T is the air density. tq Where i is the engine torque, i is the gear ratio, n is the engine speed, and a is the transmission ratio. v For acceleration, I w For the moment of inertia, I f Let η be the rotational inertia of the flywheel, η be the efficiency, and t be a constant.

[0056] The formula can be approximately rearranged to obtain the reference principle equation for modeling as follows:

[0057]

[0058] Here, a, b, c, d, and e are all constants, and velocity, torque, and acceleration can be derived from the original data.

[0059] The beneficial effects of adopting the above technical solution are as follows: Relying on the advantages of the Internet of Vehicles big data platform, this invention designs and implements a multi-AI algorithm model screening and fusion method based on the principle of dynamics. In addition, this solution proposes a new model selection algorithm to select high-quality models for fusion, and builds models by analyzing massive historical data, realizing real-time monitoring of the on-board quality of commercial vehicles without adding extra hardware. This ensures the accuracy of the results and reduces the instability of the results caused by the randomness of statistical data. Attached Figure Description

[0060] Figure 1 This is a flowchart of the vehicle quality detection method provided by the present invention;

[0061] Figure 2 This is a flowchart of data preprocessing;

[0062] Figure 3 It is an assembly drawing for working condition identification and analysis;

[0063] Figure 4 This is a flowchart of model building and fusion; Detailed Implementation

[0064] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, in order to help those skilled in the art to have a more complete, accurate and in-depth understanding of the concept and technical solutions of the present invention, and to facilitate its implementation.

[0065] Example 1, as Figure 1 As shown:

[0066] Based on the principle of dynamics, this invention designs and implements a method for screening and fusing multiple AI algorithm models. In addition, this solution proposes a new model selection algorithm to select high-quality models for fusion, and constructs models by analyzing massive historical data. This enables real-time monitoring of the on-board quality of commercial vehicles without adding extra hardware, which not only ensures the accuracy of the results but also reduces the instability of the results caused by the randomness of statistical data.

[0067] Example 2, as Figure 2 As shown:

[0068] Based on Example 1, since the data is raw message data sent from the vehicle terminal, it needs to be parsed to extract the vehicle data required for modeling. Operations such as padding, alignment, interpolation, and data smoothing are performed on the parsed data. Data that does not conform to the principles of dynamics is initially filtered out based on conditions such as clutch and braking. For data anomalies, an identification and filtering module is designed. This module mainly uses two methods: one is rule-based, judging based on the actual meaning of each attribute, such as the engine speed not being negative and the torque not exceeding a certain range; the other is the isolation forest model commonly used in model algorithm anomaly detection to remove anomaly samples and filtering based on center point peaks. Both methods perform anomaly filtering from the single attribute dimension and the overall sample dimension, respectively. Then, for the factors affecting vehicle quality analyzed in the dynamic formula, feature value derivation operations are performed to generate the fields required for modeling. Since gear shifting operations have a significant impact on the data, gear shifting operations are identified and removed.

[0069] Example 3, as Figure 3 As shown:

[0070] Based on Example 1, the choice of the base model is linear regression, which is the most common learning algorithm. The many advantages of linear models are well known. The computational cost involved in fitting these models is usually moderate and always predictable. They usually require minimal tuning, and their expression naturally facilitates interpretation. Therefore, this scheme uses regression algorithms from the sklearn toolbox to implement the base algorithm model, including LinearRegression, RANSACRegressor, Ridge, Huber, and TheilSen.

[0071] After the basic models are selected, modeling is carried out for each basic model. The training data is processed and divided in different ways. Specifically, different basic models are built and trained for driving data filtered by gear, data filtered by gear superposition working conditions, data filtered by gear superposition triangular domain transformation, data filtered by gear superposition working conditions superposition triangular domain transformation, and data divided for different working conditions.

[0072] Example 4, as Figure 4 As shown:

[0073] Based on Example 1, the data in this scheme is divided into 5 types, and 5 selected basic models are constructed for each type. Then, the model training process begins. The specific method for constructing the basic models is as follows:

[0074] Model 1: Input data: Data after gear filtering (DF0, DF20), basic model: LinearRegression;

[0075] Model 2: Input data: Data after gear filtering and operating condition filtering (DF0, DF20), basic model: RANSACRegressor;

[0076] Model 3: Input data: Data after gear filtering plus triangular threshold transformation filtering (DF0, DF20, basic model: Ridge);

[0077] Model 4: Input data: Data after gear filtering, condition filtering and triangular domain transformation filtering (DF0, DF20), basic model: TheilSen;

[0078] Model 5: Input data: Data filtered by trigonometric threshold transformation (DF0, DF20), base model: Huber.

[0079] Example 5, as Figure 4 As shown:

[0080] Based on Example 4, the five basic models that have been constructed are trained and their parameters are tuned. The training phase requires real load data. The training steps for each basic model are as follows:

[0081] 1. Input the data (DF0, DF20) filtered according to the different filtering conditions in the previous step into the regression model RANSACRegressor to fit the relationship between DF0 and DF20 and obtain the intermediate coefficient coef;

[0082] 2. Then, the intermediate coefficient coef is integrated with the actual load data (coef, load) to generate the data to be input into the base model, and each base model selected in the previous step is trained.

[0083] 3. During the model training process, the model parameters are optimized and selected to generate a basic prediction model. Then, the performance of the trained vehicle weight prediction model is evaluated using a test sample set. For the load assessment model, since the model output is a numerical value, which is a common regression problem, some commonly used indicators for machine learning regression models, such as MAPE, MAE, and R2, can be used. Here, this solution adopts MAPE as the main evaluation standard and continuously guides the iterative training of the model through evaluation indicators until the model training is completed.

[0084] 4. After training is complete, the load prediction model is output, which can predict unknown data and output the predicted load value.

[0085] Finally, the XYscore algorithm can be used to select better models from multiple models for final model fusion. The higher the XYscore score in the rating mechanism, the better the model performance.

[0086] For the multiple basic models generated during the calculation process, XYscore is used for screening. The higher the score, the better the model performance. The top-n models are selected for the final model fusion. In this scheme, n=2.

[0087] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other situations without modification, are all within the protection scope of the present invention.

Claims

1. A method for detecting the on-board mass of a commercial vehicle, characterized in that: The steps are as follows: Step S1, data preprocessing, involves data parsing, cleaning, and noise processing of the original vehicle terminal data, including but not limited to filtering and outlier removal. Step S2, Working condition identification and analysis: Further working condition analysis is performed on the preprocessed data. Step S3: Data filtering. Determine whether the data for the working condition analysis belongs to test data or training data. If it belongs to test data, proceed to step S5; if it belongs to training data, proceed to step S4. Step S4, Model Building and Fusion: Multiple different base models are built using the training data from step S3, and after training and filtering, the models are finally fused. Step S41, Model creation: Construct multiple different base models using the training data from step S3; Step S42, Model Training: The selected basic regression model is trained based on the prepared model training sample set. The model parameters are optimized and selected to generate multiple basic models for vehicle weight prediction. The performance of the trained vehicle weight prediction model is then evaluated using the test sample set. The performance indicators of the load assessment model are analyzed and judged. Step S43: Model screening. Determine whether the obtained model meets the fusion requirements. If yes, proceed to S44; otherwise, repeat step S42. The model selection method in step S43 is as follows: 1) The coefficient C is obtained by calculating the slope of the characteristic torque and acceleration related to the load; 2) Use this coefficient C to rotate the coordinate axes; 3) Calculate the single standard deviation of the newly generated coordinates X and Y respectively, and divide the single standard deviation of Y by the single standard deviation of X to obtain the Xyscore; The specific implementation algorithm of Xyscore: 1) Calculate the intermediate coefficients coef for the input data (DF0, DF20) of each constructed base model; 2) Use this coefficient (coef) to calculate the new coordinate axis. The calculation formula is as follows: ； ； 3) Calculate the single standard deviation of the newly generated coordinates X and Y respectively, and divide the single standard deviation of Y by the single standard deviation of X to obtain the XY score; Step S44, Model fusion: The multiple models obtained are fused together. Step S5, Model Service Deployment: Deploy the trained model to the cloud server, build a Docker image using the trained model and model service code, and push it to the vehicle networking Docker image repository.

2. The method for detecting the on-board mass of a commercial vehicle according to claim 1, characterized in that: Step S1 includes the following steps: Step S11, data parsing: Parse the original vehicle terminal data to extract the vehicle data required for modeling; Step S12: Data filtering. The parsed data is filled, aligned, interpolated and smoothed. The system determines whether the data conforms to the dynamics principle based on the clutch and brake operation conditions. If it does, proceed to step S13. Otherwise, it is sent back to the vehicle's original terminal. Step S13, anomaly detection: determine whether the data is abnormal. If so, proceed to S14; otherwise, send it back to the vehicle's original terminal. The detection method is as follows:

1. Make rule-based judgments based on the actual consequences of each attribute, where the rules include: speed cannot be negative and torque cannot exceed a certain range; 2. Anomaly samples are removed using the isolated forest model, which is commonly used in anomaly detection in model algorithms, and filtering is performed based on the center point peak. S14, shift detection: based on the factors affecting load capacity analyzed in the dynamic formula, determine whether the data is abnormal. If so, proceed to S2; otherwise, send it back to the vehicle's original terminal.

3. The method for detecting the on-board mass of a commercial vehicle according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: Set the conditions for working condition filtering: torque output percentage must be greater than 10%, engine speed must be greater than idle speed, and brake signal must be 0. Step S22: Generate time periods for various defined working conditions, filter them based on acceleration, identify acceleration and deceleration time periods, and identify clutch engagement time periods based on clutch signals; Step S23: Combine the times of various working conditions, select the effective working condition time, and select the data within the basic working condition time period that is not engaged in clutch or brake and conforms to the dynamic principle under the working condition filtering conditions, as the initial input data for the next step of model construction. The underlying principles of this scheme are based on dynamic formulas: ； in Represents traction force. Represents rolling resistance. Represents air resistance, Represents slope resistance. Representing inertial force, the formula can be approximately expanded as follows: Where f is the rolling resistance coefficient, and g is the acceleration due to gravity. Where A is the air resistance coefficient and A is the frontal area. air density, m is the load mass, This refers to engine torque. Where n is the transmission ratio and n is the engine speed. For acceleration, For rotational inertia, Due to the rotational inertia of the flywheel, efficiency, It is a constant; Based on relevant mathematical principles, the formula can be approximately expanded and rearranged as follows to obtain the reference principle equation for modeling: ； Here, a, b, c, d, and e are all constants, and velocity, torque, and acceleration can be derived from the original data.

4. The method for detecting the on-board mass of a commercial vehicle according to claim 1, characterized in that: The construction of multiple different base models in step S41 can be achieved in the following ways:

1. Filtering by gear position; 2. Data filtering through triangular domain transformation.

5. The method for detecting the on-board mass of a commercial vehicle according to claim 4, characterized in that: The gear filter design is as follows. 1) Statistical distribution density of vehicle speed to engine speed ratio; 2) Peak detection is performed using a distribution density map. The location of the peak is the center point of the gear speed-to-rev ratio. 3) While performing peak detection, output the peak width to obtain the upper and lower boundaries of the peak width, which are the upper and lower boundaries of the vehicle speed-to-speed ratio for each gear. 4) Filter data based on the upper and lower boundaries of the vehicle speed to engine speed ratio. Data within the upper and lower boundaries is retained, while data outside the upper and lower boundaries is filtered out.

6. The method for detecting the on-board mass of a commercial vehicle according to claim 5, characterized in that: The triangular domain transformation data filtering design is as follows. 1) Calculate the ratio between torque and acceleration, and then solve for its arctangent function; 2) Statistical kernel density distribution; 3) Theoretically, the calculated ratio coefficient corresponding to the load must have a high distribution density, so peak detection is used to detect the peak point and peak width; 4) Use the upper limit boundary of the peak width for data filtering.

Images