Commercial vehicle friction plate life prediction method based on big data

Abstract

The present application relates to the technical field of life prediction detection, in particular to a commercial vehicle friction plate life prediction method based on big data, comprising: collecting target vehicle history and current stage multi-source operation data, including driving conditions, driver operation behavior, environmental state and friction plate monitoring data. The various types of heterogeneous data are cleaned, aligned and fused to construct a standardized feature time sequence that can represent the comprehensive wear state of the friction plate. A prediction model is built based on gradient boosting decision trees, and the algorithm structure is constrained by combining the physical mechanism of friction plate wear and data statistical characteristics. The model is optimized and improved. The optimized model is used to analyze the time series feature data, and the remaining service life prediction results and corresponding confidence intervals are calculated and output. The interval determination result is combined to output the adaptive maintenance decision information. This method broadens the data collection dimension, deeply excavates the wear time series change rule, optimizes the model fitting logic, and realizes the dynamic evaluation of the friction plate wear trend.

Description

Technical Field

[0001] This invention relates to the field of life prediction and testing technology, and in particular to a method for predicting the life of friction pads for commercial vehicles based on big data. Background Technology

[0002] The operational stability of a commercial vehicle's braking system directly affects the vehicle's driving condition. As a core wear component of the braking system, friction pads continuously experience wear and tear during long-term use, with the degree of wear constantly changing according to vehicle operating conditions. Existing friction pad life analysis methods generally only collect monitoring data from the friction pads themselves for analysis and calculation. This data collection dimension is limited, lacking the simultaneous use of related data such as vehicle driving conditions, human operation behavior, and external environmental conditions, resulting in a narrow data coverage. A single data system cannot fully reflect the multiple influencing factors of component wear, and various interference factors cannot be included in the analysis scope. The dimensions for characterizing wear state are lacking, making it difficult to adapt to the wear change characteristics under complex operating conditions.

[0003] Existing lifespan prediction technologies mostly employ general-purpose algorithm models for data computation. These algorithms maintain a generic overall structure, relying solely on basic data features for training and fitting, without considering the inherent physical laws governing wear of mechanical components or incorporating rules based on statistical patterns of actual data. The operational logic of these general-purpose algorithms cannot match the changing characteristics of continuous wear on friction plates, and their ability to analyze and process time-series operational data is insufficient, failing to systematically organize wear information that changes continuously over long periods. Traditional analysis methods result in lifespan assessments in a single format, lacking data interval references, and cannot meet the analytical standards for refined operation and maintenance management. The industry's demand for targeted optimization of multi-dimensional data fusion modeling algorithms is gradually increasing, necessitating new technical solutions to overcome the inherent shortcomings of existing technologies. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for predicting the lifespan of commercial vehicle friction pads based on big data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting the lifespan of commercial vehicle friction pads based on big data, comprising: Acquire multi-source operating data of the target vehicle during historical and current operation. The multi-source operating data includes vehicle driving condition data, driver operation behavior data, environmental condition data, and monitoring data of the friction pads themselves. The acquired multi-source operational data is cleaned, aligned, and fused to construct a standardized feature time series that reflects the comprehensive wear state of the friction plates; The standardized feature time series is input into the pre-trained improved gradient boosting decision tree lifetime prediction model. The improved gradient boosting decision tree lifetime prediction model is based on the gradient boosting decision tree algorithm, and the algorithm structure is constrained according to the physical mechanism and data statistical characteristics of friction plate wear. Based on the output of the improved gradient boosting decision tree lifetime prediction model, the remaining lifetime prediction value and its confidence interval of the target vehicle friction pad are calculated. Based on the predicted remaining service life and its confidence interval, maintenance decision recommendations are generated for the friction pads of the target vehicle.

[0006] As a further aspect of the present invention, the step of cleaning, aligning, and fusing the acquired multi-source operational data to construct a standardized feature time series that reflects the comprehensive wear state of the friction plates specifically includes: For driving condition data, continuous vehicle speed sequences, acceleration sequences, and vehicle mass sequences are extracted, and time periods containing braking events are identified. Based on driver operation behavior data, braking intensity, braking frequency, and braking duration characteristics are extracted from the brake pedal signal and time-aligned with the braking event time period; For environmental condition data, temperature data, humidity data, and road slope data are integrated and matched with vehicle location information and timestamps; For the monitoring data of the friction pad itself, the time-series measurement values ​​of the friction pad thickness sensor and the friction pad temperature sensor are collected, and their rate of change per unit time is calculated. All aligned data are divided into segments using a uniform fixed time window. Within each time window, statistics on driving condition characteristics, driver operation behavior characteristics, and environmental state characteristics are calculated and concatenated with statistics on friction pad monitoring data to form a fused feature vector for the corresponding time window. The fused feature vectors of all time windows are arranged in chronological order and normalized to form a standardized feature time sequence.

[0007] As a further aspect of the present invention, the improved gradient boosting decision tree lifetime prediction model has the following improved working principle: During the model building phase, the initial loss function is defined as the mean squared error between the predicted remaining lifetime and the actual remaining lifetime. In each round of improvement iteration, the residual between the current model prediction value and the true value is calculated, and this residual is used as the new training target; When constructing a single decision tree to fit the residual, a penalty term based on the physical mechanism of friction plate wear is introduced for each split point of the tree. The penalty term is associated with the theoretical wear rate of the feature combination corresponding to the split point calculated according to the physical model. When evaluating the quality of split points, instead of relying solely on information gain or the Gini coefficient, the penalty term is weighted and combined with a measure of data fit, and the splitting scheme with the highest weighted overall score is selected first. After constructing a single decision tree, the predictions of the newly constructed decision tree are added to the current model using an adaptively adjustable learning rate. The size of this learning rate is dynamically determined based on the performance of the current model on the validation set. The improvement process is repeated until a preset number of rounds is reached or the validation set error no longer decreases, ultimately resulting in a set of gradient boosting decision tree models that integrate data-driven and physical mechanism constraints, namely the improved gradient boosting decision tree lifetime prediction model.

[0008] As a further aspect of the present invention, based on the output of the improved gradient boosting decision tree lifetime prediction model, the predicted remaining lifetime value and its confidence interval of the target vehicle friction pad are calculated, specifically including: The standardized feature time series of the target vehicle is completely input into the improved gradient boosting decision tree lifetime prediction model; Each decision tree in the improved gradient boosting decision tree lifetime prediction model independently outputs a preliminary prediction value about the remaining lifetime based on the input feature time sequence. Summarize the preliminary predictions from all decision tree outputs, calculate the average of the preliminary predictions, and use this average as the final remaining useful life prediction. Simultaneously, the standard deviation of all preliminary predictions is calculated, and the error range at a confidence level is determined by combining the prediction error distribution obtained during the training phase of the improved gradient boosting decision tree lifetime prediction model. By combining the error range with the final predicted remaining useful life, a predicted range of remaining useful life, i.e., a confidence interval, is obtained.

[0009] As a further aspect of the present invention, the step of segmenting all aligned data using a uniform fixed time window, and calculating statistics on driving condition characteristics, driver operation behavior characteristics, and environmental state characteristics within each time window, specifically includes: Set a fixed-length sliding time window, the length of which is determined based on the typical daily operating cycle of the vehicle or the braking behavior cluster cycle; Starting from the beginning of the time sequence, the sliding time window is slid across all aligned data sequentially, with a fixed step size each time. Within each sliding time window, calculate the average, maximum, standard deviation, and distribution ratio of the vehicle speed sequence within a specific speed range. Within the same sliding time window, calculate the cumulative values ​​of average deceleration, maximum deceleration, and negative acceleration for the acceleration sequence. Within the same sliding time window, for the braking intensity sequence, calculate its average braking intensity, the number of high-intensity braking events, and the variance of braking intensity fluctuations. Within the same sliding time window, calculate the average value of temperature and humidity data, and calculate the average value of absolute values ​​of road slope data. All the calculated statistics are arranged in a predetermined characteristic order to form a multidimensional driving condition feature vector, driver operation behavior feature vector, and environmental state feature vector.

[0010] As a further aspect of the present invention, the statistical data of the friction plate monitoring data are concatenated to form a fused feature vector corresponding to the time window, specifically including: Within the unified fixed time window, the time-series measurements of the friction pad thickness sensor are used to calculate the thickness reduction within the time window and the average wear thickness per unit time. Within the same time window, the peak temperature, average temperature, and cumulative time for the temperature to exceed the set threshold are calculated from the time-series measurements of the friction plate temperature sensor. Within the same time window, the correlation coefficient between the friction pad thickness change rate and the average braking intensity is calculated as an indicator reflecting the direct impact of driver operation on wear. Within the same time window, the difference between the average temperature of the friction plate and the average temperature of the environment is calculated as an indicator reflecting the heating state of the friction plate itself. The calculated thickness reduction, average wear thickness, peak temperature, average temperature, high temperature cumulative time, correlation coefficient, temperature difference statistics and indicators are combined to form a friction plate state characteristic vector. The driving condition feature vector, driver operation behavior feature vector, environmental state feature vector, and friction plate state feature vector are sequentially concatenated along the feature dimension to obtain a comprehensive fusion feature vector that represents the friction plate's operation and wear state within a time window.

[0011] As a further aspect of the present invention, maintenance decision-making recommendations for the friction pads of the target vehicle are generated based on the predicted remaining service life and its confidence interval, specifically including: Multiple remaining useful life threshold ranges are preset, and each threshold range corresponds to a maintenance decision level; The calculated predicted remaining useful life value is compared with a preset remaining useful life threshold range to determine its maintenance decision level. Based on the width of the confidence interval, a reliability assessment is performed on the determined maintenance decision level. If the confidence interval is too wide, an uncertainty marker is added to the corresponding decision level. Retrieve future planned operational task data for the target vehicle to determine whether there are specific operational tasks with high load or high safety requirements within the currently predicted remaining service life. If the specific operational task exists, the recommended maintenance time window for the specific operational task will be brought forward and integrated with the regular maintenance recommendations based on the remaining useful life prediction; The final output includes maintenance decision recommendation text containing suggested maintenance level, suggested maintenance time window, decision reliability description, and task adaptability adjustment information.

[0012] As a further aspect of the present invention, a penalty term based on the physical mechanism of friction plate wear is introduced for each split point of the tree. This penalty term is associated with the theoretical wear rate of the feature combination corresponding to the split point calculated according to the physical model, specifically including: For each feature split point to be evaluated during the decision tree construction process, the feature split point will divide the sample into two subsets according to the value of a certain feature; For the feature combination conditions defined by this feature split point, it is mapped to a simplified friction plate wear physical model, the input of which is the typical working condition represented by the feature combination conditions; The simplified physical model of friction plate wear, based on the principles of tribology, calculates a theoretical reference value for the wear rate of the friction plate according to the input typical working condition parameters. Obtain the statistical distribution of the actual wear rate corresponding to the sample subset that satisfies the aforementioned feature combination conditions in the actual data, and calculate the mean of the actual wear rate; Calculate the absolute difference between the reference value of the friction plate wear rate and the average value of the actual wear rate, and transform the absolute difference through a monotonically decreasing function to obtain the value of the penalty term; The larger the value of the penalty item, the greater the deviation between the data pattern and the expected physical mechanism at the corresponding split point, thus imposing a greater negative correction on its quality score during split point evaluation.

[0013] As a further aspect of the present invention, based on the prediction error distribution obtained during the training phase of the improved gradient boosting decision tree lifetime prediction model, an error range at a confidence level is determined, specifically including: After the model training is completed, the improved gradient boosting decision tree lifetime prediction model is tested using an independent validation dataset, and the prediction error of the model for each sample on the validation dataset is recorded. Collect the prediction errors of all validation samples to form a sample distribution of prediction errors, and calculate the standard deviation and skewness of the sample distribution; Based on the sample distribution of the prediction error, the quantile estimation method is used to find the upper and lower quantile points corresponding to the target confidence level. The range between the upper quantile point and the lower quantile point is used as the concentration interval estimate of the prediction error distribution; Half the width of the estimated central interval is weighted and fused with the standard deviation of the preliminary predictions output by all decision trees. The fusion weights are adjusted according to the size of the validation dataset to obtain the error range used to construct the confidence interval.

[0014] As a further aspect of the present invention, the calculation of the correlation coefficient between the friction pad thickness change rate and the average braking intensity, as an indicator reflecting the direct impact of driver operation on wear, specifically includes: Within the unified fixed time window, all time-series measurement points of the friction pad thickness sensor are acquired, the thickness difference between adjacent measurement points is calculated, and then all differences are divided by the corresponding time interval to obtain a series of instantaneous thickness change rate data points. Within the same time window, obtain the time-series data of braking intensity from the driver's operation behavior data, and calculate the average braking intensity within the current time window; The series of instantaneous thickness change rate data points are paired with the values ​​of braking intensity time series data at the same time to form a series of "braking intensity-instantaneous thickness change rate" data pairs; From the generated "braking intensity - instantaneous thickness change rate" data pairs, invalid data pairs caused by zero braking intensity or excessive thickness measurement noise are removed; For valid "braking intensity - instantaneous thickness change rate" data pairs, calculate their Pearson correlation coefficient. The absolute value of the correlation coefficient serves as an indicator reflecting the direct impact of the driver's braking operation on the friction pad thickness change within the corresponding time window.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This system collects four types of operational data: driving conditions, driver behavior, environmental conditions, and friction pad monitoring. A unified multi-source data processing workflow is implemented, employing standardized processes such as data cleaning, alignment, and fusion to generate a characteristic time-series sequence that comprehensively maps the wear process. After unifying and integrating these heterogeneous data, all influencing factors in the friction pad wear process are covered, eliminating the incomplete information coverage issues of single data acquisition modes. The time-series feature construction mode continuously records parameter changes throughout equipment operation, completely preserving continuous data information on wear evolution. This addresses the gaps in state recording caused by discrete data acquisition modes, ensuring a continuous and consistent expression of wear-related characteristic parameters, strengthening the completeness and correlation of wear state-related data, and improving the comprehensiveness of equipment wear state parameter representation.

[0016] Based on the gradient boosting decision tree algorithm, and according to the physical mechanism of friction pad wear and the statistical patterns of massive operational data, constraints are added to the internal structure of the algorithm, and the model feature extraction logic and parameter iteration rules are adjusted. The algorithm's operating logic is optimized based on the inherent variation law of mechanical component wear, changing the general model's undirected fitting operation mode, enhancing the model's ability to analyze nonlinear wear data, and strengthening the extraction of hidden correlation features within time-series data. After the directional structural constraint adjustment, the model can simultaneously output two types of reference content: quantitative life values ​​and confidence intervals. This enriches the output format of wear assessment, supplements the limitations of single numerical results, improves the quantitative dimension of component wear assessment, adapts to the analysis standards for brake component wear control during normal vehicle operation, and is suitable for wear assessment work under diverse operating conditions. Attached Figure Description

[0017] Figure 1 This is a state diagram of the commercial vehicle friction pad life prediction method based on big data described in this invention. Figure 2 A flowchart for constructing standardized feature time-series sequences through cleaning, alignment, and fusion of multi-source operational data; Figure 3 This is a flowchart for predicting the remaining service life of friction plates and calculating the confidence interval. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 This invention provides a method for predicting the lifespan of friction pads in commercial vehicles based on big data. The core of this method lies in integrating multi-source operational data, constructing a standardized feature time-series sequence through data processing, and utilizing an improved gradient boosting decision tree lifespan prediction model to generate maintenance decision recommendations. The specific implementation scheme is as follows: Multi-source operational data of the target vehicle during its historical and current operations are acquired. This data covers vehicle driving condition data, driver operation behavior data, environmental state data, and monitoring data of the friction pads themselves. The acquired multi-source operational data is cleaned, aligned, and fused to construct a standardized feature time-series sequence that comprehensively reflects the overall wear state of the friction pads. The constructed standardized feature time-series sequence is input into a pre-trained improved gradient boosting decision tree lifespan prediction model. This model is built based on the gradient boosting decision tree algorithm, and the algorithm structure is specifically constrained according to the physical mechanism and statistical characteristics of friction pad wear. Based on the output of this improved gradient boosting decision tree lifespan prediction model, the predicted remaining lifespan of the target vehicle's friction pads is calculated, and the confidence interval of this predicted value is also given. Based on the obtained remaining service life prediction value and its confidence interval, targeted maintenance decision-making information for the friction pads of the target vehicle is generated.

[0021] In one embodiment of the present invention, see [reference] Figure 2This paper describes the specific process of constructing a standardized feature time series. For driving condition data, it extracts continuous vehicle speed, acceleration, and onboard mass sequences containing braking event periods. For driver operation behavior data, it extracts braking intensity, braking frequency, and braking duration features from brake pedal signals and aligns these features with the identified braking event periods on the timeline. For environmental condition data, it integrates temperature, humidity, and road slope data, matching and aligning these data with vehicle location information and timestamps. For friction pad monitoring data, it collects time-series measurements from friction pad thickness and temperature sensors and calculates their respective rates of change per unit time. A unified, fixed time window is set to segment all aligned data, sliding from the starting point of the time series data. Within each time window, statistics are calculated for driving condition features, driver operation behavior features, and environmental condition features. Statistics are also calculated for friction pad monitoring data within the same time window. The various feature statistics calculated within each time window are concatenated to form a fused feature vector representing the state of that time window. The fused feature vectors of all time windows are arranged in chronological order, and the entire sequence is normalized to form a standardized feature time sequence.

[0022] In practice, multi-source operational data of the target vehicle over a continuous 30-day operating cycle is acquired. Driving condition data includes time series of speed, acceleration, and vehicle weight collected during vehicle operation; driver behavior data includes raw signals from the brake pedal travel sensor; environmental condition data includes data from external temperature and humidity sensors and road slope information from a geographic information system; and friction pad monitoring data includes measurement sequences from friction pad thickness and temperature sensors mounted on the brake calipers. Specifically, for the driving condition data, the processing flow involves extracting speed messages, accelerometer signals, and suspension pressure sensor data from the vehicle controller's local area network (LAN). Speed ​​messages are parsed into a vehicle speed sequence occurring once per second. Accelerometer signals are filtered to generate an acceleration sequence. Suspension pressure data is calibrated and converted to generate a vehicle weight sequence. Based on the negative changes in the vehicle speed sequence and the braking signal status, time periods containing braking events are identified.

[0023] In practical implementation, for driver operation behavior data, the analog signal from the brake pedal travel sensor is converted from analog to digital. The original waveform of the brake pedal opening is acquired at a sampling frequency of 0.01 seconds. Braking intensity, braking frequency, and braking duration characteristics are analyzed from the original waveform. Braking intensity is defined as the percentage of pedal opening relative to the maximum opening; braking frequency is the number of times the brake pedal opening exceeds a 5% threshold per unit time; and braking duration is the length of time the pedal opening continues to exceed the threshold in a single braking event. The timestamps of the calculated braking intensity, braking frequency, and braking duration characteristics are then precisely matched and time-aligned with the braking event time periods identified in the aforementioned driving condition data. In practical implementation, for environmental condition data, the integration work includes receiving temperature and humidity data from the vehicle's external meteorological unit, as well as road slope data obtained from a high-precision map service interface that matches the vehicle's GPS coordinates. The received temperature, humidity, and road slope data are matched and aligned with the location information and timestamps recorded by the vehicle's GPS module. The alignment operation involves interpolating the environmental data to the same timestamps as the vehicle's driving data under a unified time reference.

[0024] In some embodiments, the monitoring data for the friction pad itself is acquired by periodically reading the resistance measurement sequence of the friction pad thickness sensor and the thermocouple voltage value sequence of the friction pad temperature sensor via the controller area network bus. The resistance measurement sequence of the friction pad thickness sensor is converted into thickness measurement values ​​according to the calibration curve, forming the time-series measurement values ​​of the friction pad thickness sensor. The thermocouple voltage value sequence of the friction pad temperature sensor is converted into temperature values ​​according to the thermocouple calibration table, forming the time-series measurement values ​​of the friction pad temperature sensor. In a specific implementation, the rate of change of the time-series measurement values ​​of the friction pad thickness sensor is calculated. The calculation formula is expressed as: the thickness change rate equals the current thickness value minus the previous thickness value, divided by the time interval between the two times. In a specific implementation, the rate of change of the time-series measurement values ​​of the friction pad temperature sensor is calculated. The calculation formula is expressed as: the temperature change rate equals the current temperature value minus the previous temperature value, divided by the time interval between the two times.

[0025] In the implementation, all aligned data are segmented using a uniform fixed time window, with a length of 24 hours and a step size of 1 hour. Within each sliding time window, the average, maximum, standard deviation, and distribution proportions of vehicle speed sequences across three specific speed ranges (0-30 km / h, 30-60 km / h, and above 60 km / h) are calculated. Within the same sliding time window, the average deceleration, maximum deceleration, and cumulative negative acceleration are calculated for the acceleration sequence; the cumulative negative acceleration is defined as the sum of the absolute values ​​of all negative acceleration samples within the time window. Within the same sliding time window, the average braking intensity, the number of high-intensity braking events, and the variance of braking intensity fluctuations are calculated for the braking intensity sequence; the number of high-intensity braking events is defined as the number of braking events with braking intensity values ​​exceeding 70%. Within the same sliding time window, the average values ​​of temperature and humidity data are calculated, and the average absolute value of road slope data is calculated. In practical implementation, within the same sliding time window, the difference between the timing measurements of the friction pad thickness sensor at the start and end of the time window is calculated as the thickness reduction. This thickness reduction is then divided by the time window length to calculate the average wear thickness per unit time. In practical implementation, within the same sliding time window, the maximum value of the timing measurements of the friction pad temperature sensor is identified as the peak temperature. The arithmetic mean of these values ​​is calculated as the average temperature. The cumulative time during which the temperature exceeds the 300-degree threshold is recorded as the high-temperature cumulative time.

[0026] In the specific implementation, the calculated statistical quantities, including average vehicle speed, maximum vehicle speed, standard deviation of vehicle speed, distribution ratio of the 0-30 km / h interval, distribution ratio of the 30-60 km / h interval, distribution ratio of the interval above 60 km / h, average deceleration, maximum deceleration, cumulative negative acceleration, average braking intensity, number of high-intensity braking events, variance of braking intensity fluctuation, average temperature, average humidity, average absolute value of road slope, thickness reduction, average wear thickness, peak temperature, average temperature, and cumulative high-temperature time, are arranged in a predetermined characteristic order to form a multi-dimensional fused feature vector. In the specific implementation, the fused feature vectors formed for each time window are arranged in chronological order according to the start time of the time window to form the original feature time series sequence. In the specific implementation, all original feature time series sequences are normalized. The minimum and maximum values ​​of each feature dimension in the sequence are calculated over the entire time span, and each feature value is subjected to minimum-maximum scaling to finally form a standardized feature time series sequence.

[0027] In one embodiment of the present invention, the construction and working principle of the improved gradient boosting decision tree lifetime prediction model are described in detail. During the model construction phase, the initial loss function is defined as the mean square error between the predicted remaining lifetime and the actual remaining lifetime. In each boosting iteration, the residual between the current model prediction and the actual value is calculated, and this residual is used as the new training target to be fitted in this iteration. When constructing a single decision tree for fitting the residual, a penalty term based on the physical mechanism of friction plate wear is introduced for each split point to be evaluated in the tree. For each feature split point to be evaluated during the decision tree construction process, the split point divides the samples into two subsets based on the value of a certain feature. The feature combination conditions defined by the split point are mapped to a simplified friction plate wear physical model, the input of which is the typical working condition represented by the feature combination conditions. This simplified model calculates a theoretical wear rate reference value based on tribological principles. The mean actual wear rate of the sample subset satisfying the conditions in the actual data is obtained, and the absolute difference between the theoretical reference value and the actual mean is calculated. This difference is transformed by a monotonically decreasing function and used as the penalty term value. When evaluating the quality of split points, instead of solely relying on information gain, a weighted combination of this penalty term and a data fit metric is used, prioritizing splitting schemes with higher weighted composite scores. After constructing a single decision tree, the predictions from the new tree are accumulated into the current model using an adaptively adjustable learning rate, dynamically determined based on the model's performance on the validation set. This boosting iteration process is repeated until a preset number of iterations is reached or the validation set error no longer decreases, ultimately resulting in a gradient boosting decision tree model set that integrates data-driven and physical mechanism constraints.

[0028] In its implementation, the improved gradient boosting decision tree lifetime prediction model uses a standardized set of historical feature time series as its training set. Each sample in the training set contains a feature time series and its corresponding true value of the remaining lifetime of the friction pad. During the model building phase, the initial loss function is defined as the mean squared error between the predicted and actual remaining lifetimes. In each boosting iteration, the model obtained from the previous iteration is used to predict the training set samples. The residual between the model's predicted value and the actual value is calculated, and this residual sequence is used as the new training target for the current round. A new decision tree model is then constructed to fit this residual.

[0029] In some embodiments, when constructing a single decision tree for the fitted residuals, the tree growth process starts from the root node and divides the samples into child nodes by continuously selecting features and split points. When evaluating the quality of each candidate feature split point, a penalty term based on the physical mechanism of friction pad wear is introduced for each feature split point. In a specific implementation, for each feature split point to be evaluated during the decision tree construction process, the feature split point will divide the sample set of the current node into two subsets according to whether the value of a certain feature meets a specific condition. For example, the condition for a split point is that the average deceleration of the vehicle is greater than 0.3 m / s² and the braking intensity is greater than 50%. In a specific implementation, the feature combination conditions defined by the feature split point are mapped to a simplified friction pad wear physical model. This physical model is a simplified version based on the Arcard wear formula. The input parameters include the average braking pressure, average sliding speed, and material constant under the typical working conditions represented by the feature combination conditions. The simplified friction pad wear physical model outputs a theoretical reference value for the friction pad wear rate. In practice, from the actual training data, a subset of samples that satisfy the feature combination conditions defined by the current feature split point is found, and the mean of the actual wear rate corresponding to this subset is calculated. The actual wear rate is calculated from the change in the measured value of the friction pad thickness sensor within the corresponding time period of the sample. In practice, the absolute difference between the theoretical wear rate reference value output by the simplified friction pad wear physical model and the mean of the actual wear rate is calculated. This absolute difference is then transformed by a monotonically decreasing function and used as the value of the penalty term. This monotonically decreasing function can be a negative exponential function.

[0030] in: It is the numerical value of the penalty term (dimensionless). This is the theoretical wear rate reference value output by the physical model. It is the mean of the actual wear rate of a subset of samples that meet the conditions in the actual data. It is a normalized reference value for the wear rate (and) , (Same dimensions). In practice, the larger the value of the penalty term, the greater the deviation between the data pattern at the corresponding split point and the expected physical mechanism.

[0031] In practical implementation, the conventional metric for evaluating the quality of feature split points is information gain. Information gain reflects the degree to which the purity of the sample set is improved after splitting the samples according to this feature split point. In some embodiments, split points are not selected solely based on information gain or the Gini coefficient, but rather by weighting the penalty term and information gain. The formula for calculating the comprehensive score of the weighted combination is expressed as: comprehensive score equals information gain minus the penalty term multiplied by an adjustment coefficient. Splitting schemes with higher comprehensive scores are preferred. In practical implementation, after the construction of a single decision tree is completed, the predicted values ​​of the newly constructed decision tree are accumulated into the current model through an adaptively adjustable learning rate. The magnitude of this learning rate is dynamically determined based on the performance of the current model on an independent validation set. The method for determining this rate is to monitor the mean squared error on the validation set. If the error decreases for five consecutive rounds, the learning rate is maintained or slightly increased; if the error increases, the learning rate is decreased. In practice, the above-mentioned improvement iteration process is repeated. In each round, the residual is calculated, a decision tree with physical penalty term is constructed, and the model is updated with an adaptive learning rate until the preset number of 1000 iterations is reached or the validation set error no longer decreases in 50 consecutive iterations. This results in a gradient boosting decision tree model set that integrates data-driven and physical mechanism constraints, namely the improved gradient boosting decision tree lifetime prediction model.

[0032] In one embodiment of the present invention, specific steps for calculating the predicted remaining useful life and its confidence interval are described. See also... Figure 3 The complete, standardized time-series sequence of the target vehicle's features is input into an improved gradient boosting decision tree lifetime prediction model. Each decision tree in the model independently outputs a preliminary prediction of the remaining lifetime based on the input feature sequence. The preliminary predictions from all decision trees are aggregated, and their average is calculated as the final remaining lifetime prediction. Simultaneously, the standard deviation of all preliminary predictions is calculated. After model training, the model is tested using an independent validation dataset, and its prediction error on each validation sample is recorded. The prediction errors from all validation samples are collected, forming a sample distribution of prediction errors. Based on this sample distribution, a quantile estimation method is used to find the upper and lower quantile points corresponding to the target confidence level, and the range between them is used as the estimated central interval of the prediction error distribution. Half the width of this estimated central interval is weighted and fused with the standard deviation of the preliminary predictions from all decision trees, with the fusion weight adjusted according to the size of the validation dataset, ultimately yielding an error range. This error range is combined with the final remaining lifetime prediction to obtain a prediction interval for the remaining lifetime, i.e., a confidence interval.

[0033] In practical implementation, the process of calculating the predicted remaining service life and confidence interval of the friction pads of the target vehicle involves inputting the complete standardized feature time series sequence of the target vehicle into a trained improved gradient boosting decision tree life prediction model. The improved gradient boosting decision tree life prediction model receives this standardized feature time series sequence as input during the prediction phase. In practice, each decision tree in the improved gradient boosting decision tree life prediction model independently runs its decision rules based on the input standardized feature time series sequence and outputs a preliminary prediction value for the remaining service life from its respective leaf node. Assuming the model contains 500 decision trees, 500 preliminary prediction values ​​for the same input sequence will be obtained. In practice, the preliminary prediction values ​​output by all decision trees are aggregated, and the arithmetic mean of these preliminary prediction values ​​is calculated. This arithmetic mean is used as the final predicted remaining service life value. The calculation formula is expressed as:

[0034] in: This is the final predicted remaining useful life. It is the total number of decision trees in the model. It is the first The initial predictions output by each decision tree. In practice, the standard deviation of the initial predictions output by all decision trees is calculated simultaneously. The formula for calculating the standard deviation is: ,in This represents the sample standard deviation of the preliminary predicted values.

[0035] In some embodiments, after model training is complete, the improved gradient boosting decision tree lifetime prediction model is tested using an independent validation dataset. The validation dataset contains time-series sequences of vehicle friction pad features not seen by the model during training and their corresponding true remaining lifetimes. The prediction error of the model for each sample on the validation dataset is defined as the true remaining lifetime value of that sample minus the model's predicted remaining lifetime value for that sample. In a specific implementation, the prediction errors of all validation samples are collected to form a sample distribution of prediction errors, and the standard deviation and skewness of the sample distribution of prediction errors are calculated. In a specific implementation, based on the sample distribution of prediction errors, a quantile estimation method is used to find the upper and lower quantile points corresponding to the target confidence level, which is set to 95%. In a specific implementation, the range between the upper and lower quantile points is used as the concentration interval estimate of the prediction error distribution. In practice, half the width of the set interval estimate is weighted and fused with the standard deviations of the preliminary predictions from all decision tree outputs obtained during the target vehicle prediction process. The weights for this weighting are adjusted based on the size of the validation dataset; the larger the validation dataset, the higher the weight allocated to half the width of the set interval estimate. The final result of the weighted fusion is the error range used to construct the confidence interval, denoted as . In practical implementation, the error range will be... Compared with the final predicted remaining useful life Combining these, we obtain a prediction interval for the remaining useful life, i.e., a confidence interval, denoted as: Refer to Table 1, which shows the calculation of prediction error quantiles for a validation set containing 5 samples: Table 1: Calculation of Quantiles for Prediction Error on the Validation Set

[0036] In one embodiment of the present invention, a detailed method for calculating various characteristic statistics and constructing fused feature vectors within a fixed time window is described. A fixed-length sliding time window is set, the length of which is determined according to the typical daily operating cycle of the vehicle. Starting from the time series start point, the time window slides with a fixed step size. Within each sliding time window, the average, maximum, standard deviation, and distribution ratio within a specific speed range of the vehicle speed series are calculated. Within the same sliding time window, the average deceleration, maximum deceleration, and cumulative negative acceleration of the acceleration series are calculated. Within the same sliding time window, the average braking intensity, the number of high-intensity braking events, and the variance of braking intensity fluctuations are calculated for the braking intensity series. Within the same sliding time window, the average values ​​of temperature and humidity data are calculated, and the average absolute value of road slope data is calculated. Within the same time window, the thickness reduction within the window and the average wear thickness per unit time are calculated for the time series measurements of the friction pad thickness sensor. Within the same time window, the peak temperature, average temperature, and cumulative time for the temperature to exceed a set threshold are calculated for the time series measurements of the friction pad temperature sensor. Within the same time window, the correlation coefficient between the friction pad thickness change rate and the average braking intensity is calculated. A series of instantaneous thickness change rate data points of the friction pad thickness sensor are obtained within the time window, and the average braking intensity within the window is calculated.

[0037] The instantaneous thickness change rate data points are paired with the braking intensity values ​​at the same time. Invalid data pairs caused by zero braking intensity or excessive thickness measurement noise are eliminated. For valid data pairs, the Pearson correlation coefficient is calculated, and the absolute value of this correlation coefficient serves as an indicator reflecting the direct impact of the driver's braking operation on the thickness change. Within the same time window, the difference between the average temperature of the friction pad and the average ambient temperature is calculated as an indicator reflecting the friction pad's own heating state. All calculated driving condition characteristic statistics, driver operation behavior characteristic statistics, and environmental state characteristic statistics are combined with friction pad state statistics and indicators such as thickness reduction, average wear thickness, peak temperature, average temperature, high temperature cumulative time, correlation coefficient, and temperature difference. These are then sequentially concatenated along the feature dimension to form a comprehensive fusion feature vector representing the friction pad's operation and wear state within the time window. In specific implementation, a fixed-length sliding time window is set. The length of the sliding time window is determined based on the typical daily operating cycle of the vehicle, and is set to 24 hours. The step size of each sliding time window is set to 1 hour. In practice, starting from the beginning of the time series data, a fixed-length sliding time window is sequentially slid over all the data that has been timestamped and aligned. Each slide generates a new time window for subsequent feature statistical calculations.

[0038] In practical implementation, within each sliding time window, statistics are calculated for the vehicle speed sequence. The vehicle speed sequence originates from data recorded once per second on the controller area network bus. Within a 24-hour time window, the average, maximum, and standard deviation of the vehicle speed sequence are calculated, along with the distribution proportions of vehicle speeds within the 0-30 km / h, 30-60 km / h, and above 60 km / h ranges. Within the same sliding time window, statistics are calculated for the acceleration sequence, which originates from data collected by the inertial measurement unit. Within a 24-hour time window, the average deceleration and maximum deceleration of the acceleration sequence are calculated, along with the cumulative value of negative acceleration, defined as the sum of the absolute values ​​of all acceleration samples less than 0 within the time window. Within the same sliding time window, statistics are calculated for the braking intensity sequence, which originates from data from the brake pedal position sensor. Within a 24-hour time window, the average braking intensity, the number of high-intensity braking events, and the variance of braking intensity fluctuations are calculated. The number of high-intensity braking events is defined as the number of braking events with a braking intensity value exceeding 70%. Within the same sliding time window, the average values ​​of temperature and humidity data in the environmental status data are calculated, and the average value of the absolute value of road slope data in the environmental status data is calculated.

[0039] In practical implementation, within the same fixed sliding time window, the time-series measurements of the friction pad thickness sensor are calculated. These measurements are recorded once per minute. The difference between the initial and final measurements within the time window is calculated and used as the thickness reduction. This thickness reduction is then divided by 24 hours to calculate the average wear thickness per unit time. Within the same sliding time window, the time-series measurements of the friction pad temperature sensor are also calculated, recorded once per minute. The maximum value of the friction pad temperature sensor measurement within the time window is taken as the peak temperature. The arithmetic mean of the friction pad temperature sensor measurements within the time window is calculated as the average temperature. The cumulative time during which the friction pad temperature sensor measurement exceeds the 300-degree Celsius threshold is recorded as the high-temperature cumulative time.

[0040] In some embodiments, within the same sliding time window, the correlation coefficient between the friction pad thickness change rate and the average braking intensity is calculated. Within a specific sliding time window, all time-series measurement points of the friction pad thickness sensor are acquired, with one measurement point per minute. The thickness difference between two adjacent measurement points is calculated, and each thickness difference is divided by the corresponding time interval of one minute to obtain a series of instantaneous thickness change rate data points. Within the same sliding time window, the braking intensity time-series data from the driver's operation behavior data is acquired, also recorded once per minute. The braking intensity time-series data within the current time window is calculated. The average value of the braking intensity time series data is used as the average braking intensity. A series of instantaneous thickness change rate data points are paired one-to-one with the values ​​of the braking intensity time series data at the same time, forming a series of "braking intensity - instantaneous thickness change rate" data pairs. From the formed "braking intensity - instantaneous thickness change rate" data pairs, invalid data pairs caused by zero braking intensity or excessive thickness measurement noise are removed. Excessive thickness measurement noise is defined as the absolute value of the instantaneous thickness change rate exceeding a preset threshold. After removing invalid data pairs, the Pearson correlation coefficient is calculated for the remaining valid "braking intensity - instantaneous thickness change rate" data pairs. The formula for calculating the correlation coefficient is:

[0041] in: It is the Pearson correlation coefficient. It is the first An effective braking intensity value, It is the average value of the effective braking intensity. It is the first An instantaneous thickness change rate value paired with the braking intensity value. It is the average value of the instantaneous thickness change rate, and the correlation coefficient. The absolute value of the friction pad thickness is used as an indicator of the direct impact of the driver's braking operation on the friction pad thickness change within the corresponding time window. Within the same sliding time window, the difference between the average temperature of the friction pad and the average ambient temperature is calculated as an indicator of the friction pad's own heating state.

[0042] In practical implementation, all calculated statistics are arranged in a predetermined characteristic order to form a multi-dimensional driving condition feature vector, driver operation behavior feature vector, and environmental state feature vector. In practical implementation, the calculated friction plate state statistics and indicators, such as thickness reduction, average wear thickness, peak temperature, average temperature, high-temperature cumulative time, correlation coefficient, and temperature difference, are combined to form a friction plate state feature vector. In practical implementation, the driving condition feature vector, driver operation behavior feature vector, environmental state feature vector, and friction plate state feature vector are sequentially concatenated along their feature dimensions to obtain a comprehensive fused feature vector representing the friction plate's operation and wear state within a time window. Refer to Table 2, which shows the composition of a fused feature vector for a time window. Table 2: Fusion Feature Vector Table for a Single Time Window

[0043] In one embodiment of the present invention, the specific process for generating maintenance decision recommendation information is described. Multiple different remaining service life threshold ranges are preset, each threshold range being associated with a maintenance decision level. The calculated remaining service life prediction value is compared with these preset threshold ranges to determine its corresponding maintenance decision level. The determined maintenance decision level is reliably assessed based on the width of the calculated confidence interval. If the confidence interval width exceeds a certain range, an uncertainty marker is added to the corresponding decision level. Future planned operation task data of the target vehicle is retrieved to determine whether there are specific operation tasks with high load or high safety requirements within the currently predicted remaining service life. If such specific operation tasks exist, a recommended maintenance time window for that specific operation task is pre-set, and this adjusted recommendation is integrated with the regular maintenance recommendations derived from the remaining service life prediction. Finally, a maintenance decision recommendation text containing the recommended maintenance level, recommended maintenance time window, decision reliability description, and task adaptability adjustment information is generated.

[0044] In practice, maintenance decision recommendations for the friction pads of the target vehicle are generated based on the predicted remaining service life and its confidence interval. Multiple remaining service life threshold intervals are preset, each corresponding to a maintenance decision level. For example, three threshold intervals can be preset: remaining service life less than or equal to 7 days corresponds to the "immediate maintenance" level; remaining service life greater than 7 days and less than or equal to 30 days corresponds to the "planned maintenance" level; and remaining service life greater than 30 days corresponds to the "continuous monitoring" level. In practice, the calculated predicted remaining service life is compared with the preset remaining service life threshold intervals to determine its corresponding maintenance decision level. For example, if the calculated predicted remaining service life is 15 days, falling within the "greater than 7 days and less than or equal to 30 days" interval, the maintenance decision level is determined to be "planned maintenance." In practice, the reliability of the determined maintenance decision level is assessed based on the width of the calculated confidence interval. The calculation formula is:

[0045] in: This represents the upper limit of the confidence interval. This indicates the lower limit of the confidence interval. If the confidence interval is too wide, for example, exceeding 20 days, then a "low reliability" or "uncertainty" label is added to the corresponding decision level.

[0046] In some embodiments, future planned operational task data of the target vehicle is retrieved. This data originates from the scheduling plan of the vehicle dispatch management system. It is determined whether there are specific operational tasks with high load or high safety requirements within the currently predicted remaining service life. High-load operational tasks are defined as tasks with a planned cargo load exceeding 90% of the vehicle's rated load capacity, and high-safety-requirement operational tasks are defined as tasks transporting dangerous goods, valuables, or long-distance passenger transport. In a specific implementation, if such specific operational tasks exist, the recommended maintenance time window for those tasks is advanced. The logic for this advancement is to set the maintenance recommendation time point a preset safety margin time before the start of the specific operational task. This preset safety margin time can be set to 3 days, and it is integrated with the regular maintenance recommendations based on the remaining service life prediction. In a specific implementation, a maintenance decision recommendation information text is generated, containing the recommended maintenance level, recommended maintenance time window, decision reliability description, and task adaptability adjustment information. In practice, the generation of maintenance decision recommendation information text follows a fixed template. Specific information is filled into the corresponding positions in the template. For example, a generated text message might read: "Recommended maintenance level: Planned maintenance. Predicted remaining lifespan is 15 days, with a 95% confidence interval of [10, 20] days. Confidence interval width is 10 days, and decision reliability is assessed as moderate. Based on the vehicle operation plan, there are high-load transportation tasks within the next 14 days. It is recommended to adjust the maintenance time window to before the start of this task, i.e., it is recommended to arrange inspection and maintenance within 3 days."

[0047] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting the lifespan of friction pads for commercial vehicles based on big data, characterized in that, Includes the following steps: Acquire multi-source operating data of the target vehicle during historical and current operation. The multi-source operating data includes vehicle driving condition data, driver operation behavior data, environmental condition data, and monitoring data of the friction pads themselves. The acquired multi-source operational data is cleaned, aligned, and fused to construct a standardized feature time series that reflects the comprehensive wear state of the friction plates; The standardized feature time series is input into the pre-trained improved gradient boosting decision tree lifetime prediction model. The improved gradient boosting decision tree lifetime prediction model is based on the gradient boosting decision tree algorithm, and the algorithm structure is constrained according to the physical mechanism and data statistical characteristics of friction plate wear. Based on the output of the improved gradient boosting decision tree lifetime prediction model, the remaining lifetime prediction value and its confidence interval of the target vehicle friction pad are calculated. Based on the predicted remaining service life and its confidence interval, maintenance decision recommendations are generated for the friction pads of the target vehicle.

2. The method for predicting the lifespan of commercial vehicle friction pads based on big data according to claim 1, characterized in that, The process of cleaning, aligning, and fusing the acquired multi-source operational data to construct a standardized feature time series that reflects the comprehensive wear state of the friction plates specifically includes: For driving condition data, continuous vehicle speed sequences, acceleration sequences, and vehicle mass sequences are extracted, and time periods containing braking events are identified. Based on driver operation behavior data, braking intensity, braking frequency, and braking duration characteristics are extracted from the brake pedal signal and time-aligned with the braking event time period; For environmental condition data, temperature data, humidity data, and road slope data are integrated and matched with vehicle location information and timestamps; For the monitoring data of the friction pad itself, the time-series measurement values ​​of the friction pad thickness sensor and the friction pad temperature sensor are collected, and their rate of change per unit time is calculated. All aligned data are divided into segments using a uniform fixed time window. Within each time window, statistics on driving condition characteristics, driver operation behavior characteristics, and environmental state characteristics are calculated and concatenated with statistics on friction pad monitoring data to form a fused feature vector for the corresponding time window. The fused feature vectors of all time windows are arranged in chronological order and normalized to form a standardized feature time sequence.

3. The method for predicting the lifespan of commercial vehicle friction pads based on big data according to claim 1, characterized in that, The improved gradient boosting decision tree lifetime prediction model has the following improved working principle: During the model building phase, the initial loss function is defined as the mean squared error between the predicted remaining lifetime and the actual remaining lifetime. In each round of improvement iteration, the residual between the current model prediction value and the true value is calculated, and this residual is used as the new training target; When constructing a single decision tree to fit the residual, a penalty term based on the physical mechanism of friction plate wear is introduced for each split point of the tree. The penalty term is associated with the theoretical wear rate of the feature combination corresponding to the split point calculated according to the physical model. When evaluating the quality of split points, instead of relying solely on information gain or the Gini coefficient, the penalty term is weighted and combined with a measure of data fit, and the splitting scheme with the highest weighted overall score is selected first. After constructing a single decision tree, the predictions of the newly constructed decision tree are added to the current model using an adaptively adjustable learning rate. The size of this learning rate is dynamically determined based on the performance of the current model on the validation set. The improvement process is repeated until a preset number of rounds is reached or the validation set error no longer decreases, ultimately resulting in a set of gradient boosting decision tree models that integrate data-driven and physical mechanism constraints, namely the improved gradient boosting decision tree lifetime prediction model.

4. The method for predicting the lifespan of commercial vehicle friction pads based on big data according to claim 1, characterized in that, Based on the output of the improved gradient boosting decision tree lifetime prediction model, the predicted remaining lifetime of the target vehicle friction pads and its confidence interval are calculated, specifically including: The standardized feature time series of the target vehicle is completely input into the improved gradient boosting decision tree lifetime prediction model; Each decision tree in the improved gradient boosting decision tree lifetime prediction model independently outputs a preliminary prediction value about the remaining lifetime based on the input feature time sequence. Summarize the preliminary predictions from all decision tree outputs, calculate the average of the preliminary predictions, and use this average as the final remaining useful life prediction. Simultaneously, the standard deviation of all preliminary predictions is calculated, and the error range at a confidence level is determined by combining the prediction error distribution obtained during the training phase of the improved gradient boosting decision tree lifetime prediction model. By combining the error range with the final predicted remaining useful life, a predicted range of remaining useful life, i.e., a confidence interval, is obtained.

5. The method for predicting the lifespan of commercial vehicle friction pads based on big data according to claim 2, characterized in that, The process involves segmenting all aligned data using a uniform, fixed time window. Within each time window, statistics on driving condition characteristics, driver operational behavior characteristics, and environmental state characteristics are calculated. Specifically, this includes: Set a fixed-length sliding time window, the length of which is determined based on the typical daily operating cycle of the vehicle or the braking behavior cluster cycle; Starting from the beginning of the time sequence, the sliding time window is slid across all aligned data sequentially, with a fixed step size each time. Within each sliding time window, calculate the average, maximum, standard deviation, and distribution ratio of the vehicle speed sequence within a specific speed range. Within the same sliding time window, calculate the cumulative values ​​of average deceleration, maximum deceleration, and negative acceleration for the acceleration sequence. Within the same sliding time window, for the braking intensity sequence, calculate its average braking intensity, the number of high-intensity braking events, and the variance of braking intensity fluctuations. Within the same sliding time window, calculate the average value of temperature and humidity data, and calculate the average value of absolute values ​​of road slope data. All the calculated statistics are arranged in a predetermined characteristic order to form a multidimensional driving condition feature vector, driver operation behavior feature vector, and environmental state feature vector.

6. The method for predicting the lifespan of commercial vehicle friction pads based on big data according to claim 5, characterized in that, The statistics of the friction plate monitoring data are concatenated to form a fusion feature vector corresponding to the time window, specifically including: Within the unified fixed time window, the time-series measurements of the friction pad thickness sensor are used to calculate the thickness reduction within the time window and the average wear thickness per unit time. Within the same time window, the peak temperature, average temperature, and cumulative time for the temperature to exceed the set threshold are calculated from the time-series measurements of the friction plate temperature sensor. Within the same time window, the correlation coefficient between the friction pad thickness change rate and the average braking intensity is calculated as an indicator reflecting the direct impact of driver operation on wear. Within the same time window, the difference between the average temperature of the friction plate and the average temperature of the environment is calculated as an indicator reflecting the heating state of the friction plate itself. The calculated thickness reduction, average wear thickness, peak temperature, average temperature, high temperature cumulative time, correlation coefficient, temperature difference statistics and indicators are combined to form a friction plate state characteristic vector. The driving condition feature vector, driver operation behavior feature vector, environmental state feature vector, and friction plate state feature vector are sequentially concatenated along the feature dimension to obtain a comprehensive fusion feature vector that represents the friction plate's operation and wear state within a time window.

7. The method for predicting the lifespan of commercial vehicle friction pads based on big data according to claim 1, characterized in that, Based on the predicted remaining service life and its confidence interval, maintenance decision recommendations for the friction pads of the target vehicle are generated, specifically including: Multiple remaining useful life threshold ranges are preset, and each threshold range corresponds to a maintenance decision level; The calculated predicted remaining useful life value is compared with a preset remaining useful life threshold range to determine its maintenance decision level. Based on the width of the confidence interval, a reliability assessment is performed on the determined maintenance decision level. If the confidence interval is too wide, an uncertainty marker is added to the corresponding decision level. Retrieve future planned operational task data for the target vehicle to determine whether there are specific operational tasks with high load or high safety requirements within the currently predicted remaining service life. If the specific operational task exists, the recommended maintenance time window for the specific operational task will be brought forward and integrated with the regular maintenance recommendations based on the remaining useful life prediction; The final output includes maintenance decision recommendation text containing suggested maintenance level, suggested maintenance time window, decision reliability description, and task adaptability adjustment information.

8. The method for predicting the lifespan of commercial vehicle friction pads based on big data according to claim 3, characterized in that, The method introduces a penalty term based on the physical mechanism of friction plate wear at each split point of the tree. This penalty term is associated with the theoretical wear rate of the feature combination corresponding to the split point calculated according to the physical model, specifically including: For each feature split point to be evaluated during the decision tree construction process, the feature split point will divide the sample into two subsets according to the value of a certain feature; For the feature combination conditions defined by this feature split point, it is mapped to a simplified friction plate wear physical model, the input of which is the typical working condition represented by the feature combination conditions; The simplified physical model of friction plate wear, based on the principles of tribology, calculates a theoretical reference value for the wear rate of the friction plate according to the input typical working condition parameters. Obtain the statistical distribution of the actual wear rate corresponding to the sample subset that satisfies the aforementioned feature combination conditions in the actual data, and calculate the mean of the actual wear rate; Calculate the absolute difference between the reference value of the friction plate wear rate and the average value of the actual wear rate, and transform the absolute difference through a monotonically decreasing function to obtain the value of the penalty term; The larger the value of the penalty item, the greater the deviation between the data pattern and the expected physical mechanism at the corresponding split point, thus imposing a greater negative correction on its quality score during split point evaluation.

9. The method for predicting the lifespan of commercial vehicle friction pads based on big data according to claim 4, characterized in that, Based on the prediction error distribution obtained during the training phase of the improved gradient boosting decision tree lifetime prediction model, an error range at a confidence level is determined, specifically including: After the model training is completed, the improved gradient boosting decision tree lifetime prediction model is tested using an independent validation dataset, and the prediction error of the model for each sample on the validation dataset is recorded. Collect the prediction errors of all validation samples to form a sample distribution of prediction errors, and calculate the standard deviation and skewness of the sample distribution; Based on the sample distribution of the prediction error, the quantile estimation method is used to find the upper and lower quantile points corresponding to the target confidence level. The range between the upper quantile point and the lower quantile point is used as the concentration interval estimate of the prediction error distribution; Half the width of the estimated central interval is weighted and fused with the standard deviation of the preliminary predictions output by all decision trees. The fusion weights are adjusted according to the size of the validation dataset to obtain the error range used to construct the confidence interval.

10. The method for predicting the lifespan of commercial vehicle friction pads based on big data according to claim 6, characterized in that, The correlation coefficient between the friction pad thickness change rate and the average braking intensity is calculated as an indicator reflecting the direct impact of driver operation on wear, specifically including: Within the unified fixed time window, all time-series measurement points of the friction pad thickness sensor are acquired, the thickness difference between adjacent measurement points is calculated, and then all differences are divided by the corresponding time interval to obtain a series of instantaneous thickness change rate data points. Within the same time window, obtain the time-series data of braking intensity from the driver's operation behavior data, and calculate the average braking intensity within the current time window; The series of instantaneous thickness change rate data points are paired with the values ​​of braking intensity time series data at the same time to form a series of "braking intensity-instantaneous thickness change rate" data pairs; From the generated "braking intensity - instantaneous thickness change rate" data pairs, invalid data pairs caused by zero braking intensity or excessive thickness measurement noise are removed; For valid "braking intensity - instantaneous thickness change rate" data pairs, calculate their Pearson correlation coefficient. The absolute value of the correlation coefficient serves as an indicator reflecting the direct impact of the driver's braking operation on the friction pad thickness change within the corresponding time window.

Images