Unmanned heavy-load vehicle operation monitoring method and system based on big data

Abstract

The invention discloses an unmanned heavy-load vehicle operation monitoring method and system based on big data, and the method comprises the following steps: building an interpretable operation behavior model through a symbolic regression algorithm, carrying out the modeling of the healthy operation state of an unmanned heavy-load vehicle under various working conditions, and extracting the function relation between key variables. According to the method, a first operation baseline is generated through initial stable data, a second operation baseline is constructed under a model output stable condition, a double-baseline reference system is formed, real-time operation data is compared with the baseline after model prediction, a residual trajectory is generated, multi-dimensional residual features are extracted, and the features are input into a CUSUM judgment module. And statistical magnitude updating and sensitivity parameter self-adjustment are executed, dynamic recognition and multi-level early warning output of the degradation trend are achieved, and the method is suitable for deployment of the vehicle-mounted edge equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and vehicle operation monitoring technology, and in particular to a method and system for monitoring the operation of unmanned heavy-duty vehicles based on big data. Background Technology

[0002] With the continuous development of artificial intelligence and autonomous driving technologies, unmanned heavy-duty vehicles have been widely used in high-intensity, unattended operations such as ports, mines, and logistics. Due to their long operating cycles, complex working conditions, and unstable communication, real-time monitoring and fault warning of their operational status have become core issues in ensuring system reliability and operational safety. In environments lacking human intervention, performance degradation or potential malfunctions in heavy-duty unmanned vehicles can easily trigger high-risk events, necessitating a forward-looking, interpretable, and environmentally adaptable monitoring method.

[0003] In existing technologies, mainstream methods for monitoring the operation of autonomous vehicles mainly include black-box modeling methods based on neural networks or rule-based decision-making mechanisms based on fixed thresholds. These methods have the following significant shortcomings in practical engineering applications:

[0004] 1. Strong lack of interpretability of models: Prediction models based on deep neural networks are mostly black box structures, lacking physical interpretability of the operating mechanism, making it difficult to trace fault sources and conduct causal analysis under multiple operating conditions.

[0005] 2. Reliance on sudden anomaly identification: Most existing monitoring methods rely on the detection of sudden anomalies and lack the ability to identify gradual performance degradation processes.

[0006] 3. High dependence on computing resources: Deep learning-based monitoring systems typically rely on large-scale cloud computing resources, making them unsuitable for low-power edge computing deployment environments in unmanned vehicle-mounted devices.

[0007] 4. Poor model generalization ability: Most existing models are trained in specific scenarios and are difficult to transfer to other different road conditions, loads or environmental conditions, which limits their versatility in complex application scenarios.

[0008] 5. Rigid fixed threshold mechanism: Monitoring methods based on static rules cannot dynamically adapt to changes in vehicle status and cannot effectively identify behavioral patterns that deviate from the healthy state in the early stages but have not yet triggered a fault.

[0009] Therefore, how to provide a method and system for monitoring the operation of unmanned heavy-duty vehicles based on big data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0010] One objective of this invention is to propose a method and system for monitoring the operation of unmanned heavy-duty vehicles based on big data. This invention uses a symbolic regression algorithm to construct an interpretable operational expression model, combines a dual operational baseline mechanism to model the vehicle's health status, and utilizes the CUSUM statistical judgment module to dynamically judge residual characteristics, thereby achieving real-time monitoring and multi-level early warning output of operational performance degradation. It has the advantages of being deployable on edge devices, adapting to multiple working environments, and having early warning capabilities.

[0011] According to an embodiment of the present invention, a method and system for monitoring the operation of an unmanned heavy-duty vehicle based on big data includes the following steps:

[0012] Collect operational data of unmanned heavy-duty vehicles under various working conditions, including key operating parameters such as power system, braking system, steering control and environmental sensing, and construct a cross-scenario dataset;

[0013] The symbolic regression algorithm is applied to construct an operational expression model with variable selection and structural complexity control on a cross-scenario dataset, forming an interpretable set of functional relationships to reflect the dynamic behavior of vehicles in a healthy state.

[0014] A first baseline is built based on the stable running data in the early stage of the running expression model deployment. At the same time, a dynamic sliding mechanism for the expression is set. When the stability of the running expression model output meets the conditions, a second running baseline is built and activated to realize the dual baseline fusion mechanism.

[0015] Real-time running data is input into the running expression model to generate predicted values, and residual differences are calculated with two types of baseline values ​​respectively to form a residual evolution trajectory based on a sliding window;

[0016] Multidimensional residual features for health trend determination are extracted from the residual evolution trajectory, including indicators such as drift direction, perturbation frequency, fluctuation amplitude and slope change.

[0017] The multidimensional residual features are input into the CUSUM judgment module, and statistical update and sensitivity parameter self-adjustment operations are performed to dynamically generate degradation trend judgment results.

[0018] The low-power edge computing module integrates the running expression model, dual baseline fusion mechanism, multi-dimensional residual features and trend judgment logic, and outputs multi-level early warning signals and running anomaly markers when the trigger conditions are met.

[0019] Optionally, the application of the symbolic regression algorithm to construct a running expression model with variable selection and structural complexity control on a cross-scenario dataset specifically includes:

[0020] Correlation analysis was performed on the power system operating parameters, braking system operating parameters, steering control operating parameters and environmental sensing operating parameters collected centrally from cross-scenario datasets to eliminate highly redundant input dimensions and generate a candidate set of variables for model building.

[0021] A symbolic expression search space is constructed based on a set of candidate variables, the symbolic expression search space including a set of arithmetic operators, a set of mathematical functions, and a set of variable combination rules;

[0022] The symbolic regression algorithm is used to perform operations such as generating, mutating, and recombinating the symbolic expression search space to construct a set of candidate expression structures.

[0023] The prediction performance of the candidate expression structure set is calculated on a cross-scenario dataset, and the expression structure complexity index is constructed by combining the corresponding structure depth, number of symbols and combinatorial redundancy to form a comprehensive evaluation function.

[0024] Based on the comprehensive evaluation function, the candidate expression structure set is retained, mutated, and updated. Multiple rounds of expression evolution iterations are performed until the evaluation function value converges, at which point the running expression model is determined.

[0025] Optionally, the step of constructing a first operational baseline based on stable operational data from the initial deployment of the running expression model, and simultaneously setting a dynamic sliding mechanism for the running expression model, to construct and activate a second operational baseline when the output stability of the running expression model meets the conditions, specifically includes:

[0026] The running expression model is deployed to the on-board edge computing module of the unmanned heavy-duty vehicle. Real-time running data within a continuous running cycle is collected in the early stage of model deployment, and the real-time running data is input into the running expression model to obtain the prediction output results of the continuous time series.

[0027] The mean and fluctuation range within the sliding window are calculated for the predicted output results. The fluctuation amplitude and equilibrium stability within the specified time length are determined. If the stability determination condition is met, the first operating baseline is constructed based on the predicted output results of the corresponding time period, which serves as the model output reference value under the healthy operating state of the unmanned heavy-duty vehicle.

[0028] A dynamic sliding mechanism for the expression is set, which includes sliding detection window parameters, adaptive sampling interval parameters, and residual offset rate threshold parameters for identifying changes in residual trends, and is used to dynamically detect changes in the stability of the output of the running expression model.

[0029] When the residual sequence between the predicted output and the real-time running data satisfies the stability condition, the second running baseline construction logic is started, and the predicted output in the current sliding window is used to reconstruct the stability expression to generate the second running baseline.

[0030] The first and second operating baselines are retained in the edge computing module and used as reference benchmarks for subsequent residual difference calculation, multi-baseline fusion determination and degradation trend identification.

[0031] Optionally, the step of inputting real-time running data into the running expression model to generate predicted output values, and calculating the residual differences with the first running baseline and the second running baseline respectively to form a residual evolution trajectory based on a sliding time window specifically includes:

[0032] The real-time running data is input into the running expression model deployed in the vehicle edge computing module. The expression structure constructed in the running expression model is parsed and calculated item by item to generate the predicted output value at the current time point.

[0033] Extract the baseline reference output values ​​of the first and second operating baselines at corresponding time points, calculate the difference between the predicted output value of the running expression model and the first operating baseline reference output value as the first type of residual difference, and the difference between the predicted output value and the second operating baseline reference output value as the second type of residual difference;

[0034] Set the sliding time window parameters, and construct a sliding time window sequence with a fixed length and sliding step size during the operation. Sample the first type of residual difference and the second type of residual difference at continuous time points in each window to form a residual sequence.

[0035] The residual difference values ​​recorded in each sliding time window are arranged in chronological order to construct the first residual evolution trajectory and the second residual evolution trajectory, which serve as the input basis for the subsequent residual feature extraction and degradation trend determination modules.

[0036] Optionally, the step of inputting multidimensional residual features into the CUSUM judgment module, performing cumulative statistic updates and sensitivity parameter self-adjustment operations, and dynamically generating degradation trend judgment results specifically includes:

[0037] Extract the mean value of residual changes, fluctuation amplitude, gradient direction and change frequency within the sliding time window from the first residual evolution trajectory and the second residual evolution trajectory, and construct a multidimensional residual feature vector;

[0038] The multidimensional residual feature vector is fed into the CUSUM decision module. The CUSUM decision module updates the cumulative statistics of the residual feature vector at each time step and records the offset direction and increase in real time according to the cumulative change trend.

[0039] During the cumulative statistic update process, the sensitivity parameter set in the CUSUM judgment module is dynamically calculated based on the degree of change in the residual characteristic variance within the sliding time window, and the upper and lower control thresholds of the cumulative statistic are automatically adjusted.

[0040] When the cumulative statistical value reaches the judgment threshold set by the current sensitivity parameter, the CUSUM judgment module outputs the degradation trend judgment result for the current time period. The degradation trend judgment result includes degradation direction, degradation level and persistence index.

[0041] The cumulative statistical update data, sensitivity parameter adjustment records, and degradation trend judgment results are stored locally in the vehicle edge computing module and provided as input to the subsequent early warning signal triggering module and operation and maintenance decision module.

[0042] Optionally, the upper and lower control thresholds of the self-adjusting cumulative statistic specifically include:

[0043] Based on the residual sequence within the sliding time window constructed from the first residual evolution trajectory and the second residual evolution trajectory, the time difference score of the change in the mean residual in adjacent time periods is calculated, and the time difference score is defined as the change in the error steering rate.

[0044] The change in error steering rate is used as a sensitivity parameter to construct an adaptive decision threshold for controlling the trigger threshold of the CUSUM cumulative statistic. The threshold is calculated using the following formula:

[0045] h k =h0+γ·|Δθ k |;

[0046] Among them, h k The adaptive threshold value at the current moment, h0 is the initial threshold value, γ is the sensitivity adjustment coefficient, and Δθ is the sensitivity adjustment coefficient. k This represents the change in the error steering rate within the k-th sliding time window.

[0047] Optionally, the triggering conditions for outputting multi-level early warning signals and operation abnormality markers when the triggering conditions are met include:

[0048] When the current cumulative statistic in the CUSUM judgment module is greater than the adaptive judgment threshold value calculated from the change in error guidance rate, the first warning condition is triggered.

[0049] When the fluctuation amplitude and change frequency in the multidimensional residual feature vector simultaneously exceed the corresponding feature threshold within the sliding time window of the first residual evolution trajectory and the second residual evolution trajectory, and maintain the trend consistency for multiple consecutive time windows, the second warning condition is determined to be triggered.

[0050] Among them, the feature threshold is preset by the statistical maximum value of the multidimensional residual feature vector in the historical normal operation data during the initialization stage of the edge computing module, and is solidified as a built-in reference parameter of the system.

[0051] When either the first or second warning condition is met, the trend judgment logic control unit generates multi-level warning signals and operation abnormality markers.

[0052] Optionally, it includes an edge computing module deployed on the onboard platform of the autonomous heavy-duty vehicle, the edge computing module comprising:

[0053] The running expression model processing unit is used to receive the power system operating parameters, braking system operating parameters, steering control operating parameters and environmental sensing operating parameters collected during the operation of the unmanned heavy-duty vehicle, and input the above operating parameters into the running expression model to generate predicted output values.

[0054] The dual-baseline fusion processing unit is used to call the first operating baseline and the second operating baseline, calculate the residual difference for the predicted output value, and generate the first residual evolution trajectory and the second residual evolution trajectory.

[0055] The multidimensional residual feature extraction unit is used to extract the mean value of residual change, fluctuation amplitude, change frequency and gradient direction from the first residual evolution trajectory and the second residual evolution trajectory based on a sliding time window, and construct a multidimensional residual feature vector;

[0056] The CUSUM determination module is used to receive the multidimensional residual feature vector, update the cumulative statistics based on the error guidance rate, and perform degradation trend determination according to the adaptive determination threshold, wherein the adaptive determination threshold is constructed based on the change of the error guidance rate.

[0057] The trend judgment logic control unit is used to determine whether the current cumulative statistics exceed the adaptive judgment threshold, or to determine whether the fluctuation amplitude and change frequency in the multidimensional residual feature vector exceed the initial feature threshold for multiple consecutive sliding time windows, and outputs multi-level early warning signals and operation abnormality markers when either condition is met.

[0058] The beneficial effects of this invention are:

[0059] (1) This invention introduces a symbolic regression algorithm to construct an operational expression model of an unmanned heavy-duty vehicle in a healthy state, thereby realizing interpretable modeling of operational behavior, improving the model's adaptability and physical transferability in multiple working conditions, and effectively solving the shortcomings of traditional black-box models in terms of interpretability and deployment universality. At the same time, by constructing an expression modeling process that includes variable selection and structural complexity control mechanisms, the stability and generalization performance of the model structure are enhanced.

[0060] (2) The dual-baseline fusion mechanism designed in this invention combines the stable state output in the early stage of model deployment with the dynamic slip expression in the running process. It can simultaneously refer to the long-term and short-term running deviation behavior, provide a more robust benchmark for residual evolution trend identification, and improve the system's sensitivity and accuracy in identifying gradual changes in running behavior.

[0061] (3) The CUSUM judgment module proposed in this invention integrates a statistical update mechanism based on multidimensional residual features and a sensitivity adaptive adjustment strategy. It can dynamically adapt to the feature fluctuations under different operating conditions and output the degradation trend judgment result including trend direction and offset degree, effectively enhancing the system's early warning capability for gradual performance degradation.

[0062] (4) This invention integrates the running expression model, dual running baseline fusion mechanism, multi-dimensional residual feature extraction and trend judgment logic into a low-power edge computing module, realizing local real-time monitoring and anomaly identification of vehicle operation status in environments such as unattended operation and unstable communication. It has the advantages of high reliability, low latency and low resource consumption, and meets the real-time security requirements in complex application scenarios. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a flowchart of a method and system for monitoring the operation of unmanned heavy-duty vehicles based on big data, as proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1 A method and system for monitoring the operation of unmanned heavy-duty vehicles based on big data, comprising the following steps:

[0067] Collect operational data of unmanned heavy-duty vehicles under various working conditions, including key operating parameters such as power system, braking system, steering control and environmental sensing, and construct a cross-scenario dataset;

[0068] The symbolic regression algorithm is applied to construct an operational expression model with variable selection and structural complexity control on a cross-scenario dataset, forming an interpretable set of functional relationships to reflect the dynamic behavior of vehicles in a healthy state.

[0069] A first baseline is built based on the stable running data in the early stage of the running expression model deployment. At the same time, a dynamic sliding mechanism for the expression is set. When the stability of the running expression model output meets the conditions, a second running baseline is built and activated to realize the dual baseline fusion mechanism.

[0070] Real-time running data is input into the running expression model to generate predicted values, and residual differences are calculated with two types of baseline values ​​respectively to form a residual evolution trajectory based on a sliding window;

[0071] Multidimensional residual features for health trend determination are extracted from the residual evolution trajectory, including indicators such as drift direction, perturbation frequency, fluctuation amplitude and slope change.

[0072] The multidimensional residual features are input into the CUSUM judgment module, and statistical update and sensitivity parameter self-adjustment operations are performed to dynamically generate degradation trend judgment results.

[0073] The low-power edge computing module integrates the running expression model, dual baseline fusion mechanism, multi-dimensional residual features and trend judgment logic, and outputs multi-level early warning signals and running anomaly markers when the trigger conditions are met.

[0074] In this embodiment, the application of the symbolic regression algorithm to construct a running expression model with variable selection and structural complexity control on a cross-scenario dataset specifically includes:

[0075] Correlation analysis was performed on the power system operating parameters, braking system operating parameters, steering control operating parameters and environmental sensing operating parameters collected centrally from cross-scenario datasets to eliminate highly redundant input dimensions and generate a candidate set of variables for model building.

[0076] A symbolic expression search space is constructed based on a set of candidate variables, the symbolic expression search space including a set of arithmetic operators, a set of mathematical functions, and a set of variable combination rules;

[0077] The symbolic regression algorithm is used to perform operations such as generating, mutating, and recombinating the symbolic expression search space to construct a set of candidate expression structures.

[0078] The prediction performance of the candidate expression structure set is calculated on a cross-scenario dataset, and the expression structure complexity index is constructed by combining the corresponding structure depth, number of symbols and combinatorial redundancy to form a comprehensive evaluation function.

[0079] Based on the comprehensive evaluation function, the candidate expression structure set is retained, mutated, and updated. Multiple rounds of expression evolution iterations are performed until the evaluation function value converges, at which point the running expression model is determined.

[0080] This implementation method performs correlation analysis on the operating parameters of the power system, braking system, steering control, and environmental sensing collected from cross-scenario datasets. After eliminating redundant dimensions, a candidate set of variables is constructed. Within the symbolic expression search space containing arithmetic operators, mathematical functions, and combination rules, a symbolic regression algorithm is used to generate, mutate, and recombine the expression structure. In multiple iterations, a comprehensive evaluation function is constructed based on structural complexity and prediction performance. The system optimizes and finally determines an interpretable operating expression model, thereby achieving accurate modeling of the operating behavior of unmanned heavy-duty vehicles and improving the model's structural stability and cross-scenario adaptability.

[0081] In this embodiment, the construction of a first operational baseline based on stable operational data from the initial deployment of the operational expression model, and the setting of a dynamic sliding mechanism for the operational expression model, to construct and activate a second operational baseline when the output stability of the operational expression model meets the conditions, specifically includes:

[0082] The running expression model is deployed to the on-board edge computing module of the unmanned heavy-duty vehicle. Real-time running data within a continuous running cycle is collected in the early stage of model deployment, and the real-time running data is input into the running expression model to obtain the prediction output results of the continuous time series.

[0083] The mean and fluctuation range within the sliding window are calculated for the predicted output results. The fluctuation amplitude and equilibrium stability within the specified time length are determined. If the stability determination condition is met, the first operating baseline is constructed based on the predicted output results of the corresponding time period, which serves as the model output reference value under the healthy operating state of the unmanned heavy-duty vehicle.

[0084] A dynamic sliding mechanism for the expression is set, which includes sliding detection window parameters, adaptive sampling interval parameters, and residual offset rate threshold parameters for identifying changes in residual trends, and is used to dynamically detect changes in the stability of the output of the running expression model.

[0085] When the residual sequence between the predicted output and the real-time running data satisfies the stability condition, the second running baseline construction logic is started, and the predicted output in the current sliding window is used to reconstruct the stability expression to generate the second running baseline.

[0086] The first and second operating baselines are retained in the edge computing module and used as reference benchmarks for subsequent residual difference calculation, multi-baseline fusion determination and degradation trend identification.

[0087] This implementation deploys the running expression model to the onboard edge computing module of an unmanned heavy-duty vehicle. It constructs a first operating baseline using continuous operating data collected during the initial deployment phase and, combined with a sliding window mechanism, sets a dynamic sliding mechanism for the expression to generate a second operating baseline when stability conditions are met. This achieves dynamic construction and parallel retention of dual baselines. This method effectively improves the model's responsiveness to changes in operating status, enabling the system to simultaneously capture short-term fluctuations and long-term offset characteristics. It enhances the continuity and sensitivity of operational monitoring, providing a highly timely and adaptable benchmark mechanism for safety monitoring of heavy-duty vehicles under unattended conditions.

[0088] In this embodiment, the step of inputting real-time running data into the running expression model to generate predicted output values, and calculating the residual differences with the first running baseline and the second running baseline respectively to form a residual evolution trajectory based on a sliding time window specifically includes:

[0089] The real-time running data is input into the running expression model deployed in the vehicle edge computing module. The expression structure constructed in the running expression model is parsed and calculated item by item to generate the predicted output value at the current time point.

[0090] Extract the baseline reference output values ​​of the first and second operating baselines at corresponding time points, calculate the difference between the predicted output value of the running expression model and the first operating baseline reference output value as the first type of residual difference, and the difference between the predicted output value and the second operating baseline reference output value as the second type of residual difference;

[0091] Set the sliding time window parameters, and construct a sliding time window sequence with a fixed length and sliding step size during the operation. Sample the first type of residual difference and the second type of residual difference at continuous time points in each window to form a residual sequence.

[0092] The residual difference values ​​recorded in each sliding time window are arranged in chronological order to construct the first residual evolution trajectory and the second residual evolution trajectory, which serve as the input basis for the subsequent residual feature extraction and degradation trend determination modules.

[0093] This implementation inputs real-time operating data into an operating expression model deployed in an onboard edge computing module. Based on the expression structure, it performs item-by-item parsing and calculation to generate predicted output values. It then calculates two types of residual differences by combining the reference output values ​​of the first and second operating baselines at corresponding time points. Using a set sliding time window, it samples and sorts the continuous residual data, and finally constructs the first and second residual evolution trajectories. This enables a dynamic expression of the operating state deviation of unmanned heavy-duty vehicles, improving the sensitivity of the operation monitoring system to gradual performance changes and its forward-looking early warning capability.

[0094] In this embodiment, the step of inputting multidimensional residual features into the CUSUM judgment module, performing cumulative statistic updates and sensitivity parameter self-adjustment operations, and dynamically generating degradation trend judgment results specifically includes:

[0095] Extract the mean value of residual changes, fluctuation amplitude, gradient direction and change frequency within the sliding time window from the first residual evolution trajectory and the second residual evolution trajectory, and construct a multidimensional residual feature vector;

[0096] The multidimensional residual feature vector is fed into the CUSUM decision module. The CUSUM decision module updates the cumulative statistics of the residual feature vector at each time step and records the offset direction and increase in real time according to the cumulative change trend.

[0097] During the cumulative statistic update process, the sensitivity parameter set in the CUSUM judgment module is dynamically calculated based on the degree of change in the residual characteristic variance within the sliding time window, and the upper and lower control thresholds of the cumulative statistic are automatically adjusted.

[0098] When the cumulative statistical value reaches the judgment threshold set by the current sensitivity parameter, the CUSUM judgment module outputs the degradation trend judgment result for the current time period. The degradation trend judgment result includes degradation direction, degradation level and persistence index.

[0099] The cumulative statistical update data, sensitivity parameter adjustment records, and degradation trend judgment results are stored locally in the vehicle edge computing module and provided as input to the subsequent early warning signal triggering module and operation and maintenance decision module.

[0100] This implementation extracts information such as the mean value, fluctuation amplitude, gradient direction, and frequency of residual changes within the sliding time window from the first and second residual evolution trajectories to construct a multidimensional residual feature vector. This vector is then input into the CUSUM decision module to achieve dynamic monitoring of residual change trends. The CUSUM decision module dynamically adjusts the sensitivity parameters based on changes in feature variance, adaptively updating the cumulative statistics and control thresholds. When the statistical value exceeds a set threshold, it outputs the degradation trend judgment result, including degradation direction, degradation level, and persistence index. All intermediate judgment results are stored locally in the vehicle edge computing module, thereby achieving early identification and closed-loop response to the gradual performance degradation of unmanned heavy-duty vehicles, effectively improving the system's stability and safe operation assurance capabilities.

[0101] In this embodiment, the control upper and lower threshold values ​​of the self-adjusting cumulative statistic specifically include:

[0102] Based on the residual sequence within the sliding time window constructed from the first residual evolution trajectory and the second residual evolution trajectory, the time difference score of the change in the mean residual in adjacent time periods is calculated, and the time difference score is defined as the change in the error steering rate.

[0103] The change in error steering rate is used as a sensitivity parameter to construct an adaptive decision threshold for controlling the trigger threshold of the CUSUM cumulative statistic. The threshold is calculated using the following formula:

[0104] h k =h0+γ·|Δθ k |;

[0105] Among them, h k The adaptive threshold value at the current moment, h0 is the initial threshold value, γ is the sensitivity adjustment coefficient, and Δθ is the sensitivity adjustment coefficient. k This represents the change in the error steering rate within the k-th sliding time window.

[0106] In this embodiment, the time difference value of the residual mean change is extracted within the sliding time window constructed based on the first residual evolution trajectory and the second residual evolution trajectory. This difference value is defined as the change in error guidance rate and is introduced into the CUSUM determination module as a sensitivity parameter to dynamically calculate the adaptive trigger threshold of the CUSUM cumulative statistics.

[0107] Specifically, the threshold is based on formula T. t =T0+α·ΔR t Construct, where T t This represents the adaptive threshold value at the current moment, T0 is the initial threshold value, α is the sensitivity adjustment coefficient, and ΔR is the sensitivity adjustment coefficient. t Let be the change in error guidance rate within the t-th sliding window. This method, by combining the linkage adjustment mechanism of real-time residual dynamic characteristics and judgment sensitivity, effectively improves the adaptability of the CUSUM judgment module to changes in operating trends and the accuracy of early degradation identification, and realizes flexible response and sensitive detection of progressive performance degradation of unmanned heavy-duty vehicles.

[0108] This formula incorporates the change in error steering rate as a sensitivity adjustment factor, dynamically participating in the construction of the CUSUM cumulative statistic threshold, thus achieving adaptive adjustment of the threshold value according to the error evolution trend. Its principle lies in combining the absolute magnitude of the error steering rate change with the sensitivity adjustment parameter to form a judgment benchmark that can be dynamically corrected according to the residual trend. This method innovatively breaks through the traditional fixed threshold setting method of CUSUM, introducing real-time error steering trend information into sensitivity determination for the first time. This enables the system to respond to gradual, subtle degradation, significantly improving the adaptability and forward-looking judgment capability of the monitoring system.

[0109] In this embodiment, the triggering conditions for outputting multi-level early warning signals and operation abnormality markers when the triggering conditions are met include:

[0110] When the current cumulative statistic in the CUSUM judgment module is greater than the adaptive judgment threshold value calculated from the change in error guidance rate, the first warning condition is triggered.

[0111] When the fluctuation amplitude and change frequency in the multidimensional residual feature vector simultaneously exceed the corresponding feature threshold within the sliding time window of the first residual evolution trajectory and the second residual evolution trajectory, and maintain the trend consistency for multiple consecutive time windows, the second warning condition is determined to be triggered.

[0112] Among them, the feature threshold is preset by the statistical maximum value of the multidimensional residual feature vector in the historical normal operation data during the initialization stage of the edge computing module, and is solidified as a built-in reference parameter of the system.

[0113] When either the first or second warning condition is met, the trend judgment logic control unit generates multi-level warning signals and operation abnormality markers.

[0114] In this implementation, the low-power edge computing module is equipped with an adaptive judgment threshold calculated based on the change in error guidance rate in the CUSUM judgment module and a cumulative statistic in real time to trigger a first warning condition. Combined with the continuous over-threshold behavior of the fluctuation amplitude and change frequency of the multidimensional residual feature vector in the first and second residual evolution trajectories within the sliding time window, a second warning condition is triggered. When either condition is met, the trend judgment logic control unit issues multi-level warning signals and operation anomaly markers, which significantly enhances the adaptive response capability and real-time risk intervention effect for abnormal operation status of the unmanned heavy-duty vehicle system.

[0115] This embodiment includes an edge computing module deployed on the onboard platform of an unmanned heavy-duty vehicle, the edge computing module comprising:

[0116] The running expression model processing unit is used to receive the power system operating parameters, braking system operating parameters, steering control operating parameters and environmental sensing operating parameters collected during the operation of the unmanned heavy-duty vehicle, and input the above operating parameters into the running expression model to generate predicted output values.

[0117] The dual-baseline fusion processing unit is used to call the first operating baseline and the second operating baseline, calculate the residual difference for the predicted output value, and generate the first residual evolution trajectory and the second residual evolution trajectory.

[0118] The multidimensional residual feature extraction unit is used to extract the mean value of residual change, fluctuation amplitude, change frequency and gradient direction from the first residual evolution trajectory and the second residual evolution trajectory based on a sliding time window, and construct a multidimensional residual feature vector;

[0119] The CUSUM determination module is used to receive the multidimensional residual feature vector, update the cumulative statistics based on the error guidance rate, and perform degradation trend determination according to the adaptive determination threshold, wherein the adaptive determination threshold is constructed based on the change of the error guidance rate.

[0120] The trend judgment logic control unit is used to determine whether the current cumulative statistics exceed the adaptive judgment threshold, or to determine whether the fluctuation amplitude and change frequency in the multidimensional residual feature vector exceed the initial feature threshold for multiple consecutive sliding time windows, and outputs multi-level early warning signals and operation abnormality markers when either condition is met.

[0121] Example 1:

[0122] To verify the feasibility of this invention in practice, it was experimentally deployed in an automated transportation operation scenario within a mining area. This mining area is located in an open-pit coal mine, where multiple unmanned heavy-duty transport vehicles are required to complete short-distance, high-frequency transportation of coal and stone in daily operations. The operating environment in this area is complex, with high vehicle operating frequency and large load variations, making it prone to hidden degradation problems such as system thermal decay, sensor drift, and slow control response caused by road conditions. Traditional operation monitoring methods rely on single-point alarm logic, which has significant drawbacks such as high false alarm rates and inability to detect gradual degradation. Therefore, a more intelligent, dynamic, and interpretable monitoring method is urgently needed.

[0123] The method of this invention was deployed on five unmanned heavy-duty vehicles in the aforementioned mining area, using customized vehicle-mounted edge computing terminals for model calculation and data judgment. The deployment scheme covers the entire process from operational data acquisition, expression modeling, residual construction, multi-dimensional feature extraction, CUSUM judgment to early warning output.

[0124] During operation, the on-board system collects real-time operating parameters of the power system (such as engine speed and throttle opening), braking system (such as braking force distribution and hydraulic status), steering control (such as steering angle response and deviation between target and actual values), and environmental sensing parameters (such as lidar ranging and temperature and humidity sensor values). It constructs a cross-scene dataset and inputs it into a symbolic regression algorithm for expression modeling training. In the early stage of model deployment, the first operating baseline is constructed using data from 10 consecutive hours of operation. Through the residual stability judgment mechanism, a second operating baseline is constructed at the 36th hour, thus establishing a dual-baseline reference system.

[0125] The expression model output and the two baseline values ​​are used to calculate residuals within a sliding time window, forming the first and second residual evolution trajectories. Their multidimensional features are then extracted and input into the CUSUM decision module. The CUSUM module dynamically constructs cumulative statistics based on the error guidance rate and combines this with a sensitivity parameter self-adjustment mechanism to determine trends and trigger multi-level early warning logic.

[0126] After 5 consecutive days of operation in the mining area (16 hours per day), the cumulative monitoring data of the system is shown in the table below:

[0127] Table 1 Comparison of anomaly detection results for different systems under typical operating days.

[0128]

[0129]

[0130] Table 2 compares the anomaly detection performance of traditional methods and the system of this invention.

[0131]

[0132] As can be seen from the data in the two tables above, the big data-based unmanned heavy-duty vehicle operation monitoring system described in this invention exhibits significant advantages in actual mining area applications. Compared to traditional baseline monitoring methods, this system shows substantial improvements in false alarm control, degradation trend identification, and fault early warning timeliness. The average daily number of false alarms has decreased from 3.6 to 0.8, the average anomaly response delay has been shortened from 28 minutes to 9 minutes, the accuracy rate of gradual degradation identification has increased to 91.7%, and the early fault detection rate has reached 88.2%. Monitoring results of different vehicle systems over specific operating days show that the method of this invention can not only effectively identify multiple types of anomalies, such as steering response delay, engine fuel injection anomalies, and sensor drift, but also distinguish between short-term disturbances and long-term trend deviations, achieving more intelligent and accurate operation monitoring.

[0133] The symbolic regression model, combined with variable selection and expression structure control, effectively avoids the problems of black-box modeling, achieving accurate modeling of vehicle health status under multiple operating conditions. Secondly, the dual-baseline mechanism can adaptively adapt to the data characteristics of different time periods during the initial modeling phase and after stabilization, greatly enhancing the robustness of residual calculation. Through the extraction of multi-dimensional features and trend modeling in the residual evolution trajectory, this system can not only capture sudden anomalies but also possess the ability to proactively identify gradual degradation trends. Finally, by constructing a CUSUM judgment mechanism based on error guidance rate, the system effectively suppresses the false alarm rate while maintaining sensitivity, greatly improving the practical value of early warning.

[0134] For example, vehicle CX-02 detected a sustained deviation between engine speed and expected response. The CUSUM determination module's cumulative statistics exceeded the sensitivity-adjusted threshold within 15 minutes, triggering a Level 1 warning and recording the abnormal event. Subsequent inspection by maintenance personnel confirmed it as an early sign of injector blockage, successfully preventing a larger-scale control failure.

[0135] In summary, the big data-based unmanned heavy-duty vehicle operation monitoring method of the present invention, by integrating symbolic expression modeling, dual baseline determination, multidimensional residual evolution analysis and CUSUM trend determination mechanism, not only improves the accuracy, real-time performance and interpretability of monitoring, but also significantly reduces the probability of false alarms and missed alarms. It demonstrates excellent engineering practicality and technological leadership in complex and ever-changing actual operation scenarios.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring the operation of unmanned heavy-duty vehicles based on big data, characterized in that, Includes the following steps: Collect operational data of unmanned heavy-duty vehicles under various working conditions to construct a cross-scenario dataset; The symbolic regression algorithm is applied to construct a running expression model with variable selection and structural complexity control on a cross-scenario dataset; A first baseline is built based on the stable running data in the early stage of the running expression model deployment. At the same time, a dynamic sliding mechanism for the expression is set up to build and activate a second running baseline when the stability of the running expression model output meets the conditions. Real-time running data is input into the running expression model to generate predicted values, and residual differences are calculated with two types of baseline values ​​respectively to form a residual evolution trajectory based on a sliding window; Extract multidimensional residual features for health trend determination from residual evolution trajectories; The multidimensional residual features are input into the CUSUM judgment module, and statistical update and sensitivity parameter self-adjustment operations are performed to dynamically generate degradation trend judgment results. The low-power edge computing module integrates the running expression model, dual baseline fusion mechanism, multi-dimensional residual features and trend judgment logic, and outputs multi-level early warning signals and running anomaly markers when the trigger conditions are met.

2. The method for monitoring the operation of unmanned heavy-duty vehicles based on big data according to claim 1, characterized in that, The application of the symbolic regression algorithm to construct a running expression model with variable selection and structural complexity control on cross-scenario datasets specifically includes: Correlation analysis of variables across cross-scenario datasets is performed to remove highly redundant input dimensions and generate a candidate set of variables for model building. A symbolic expression search space is constructed based on a set of candidate variables, the symbolic expression search space including a set of arithmetic operators, a set of mathematical functions, and a set of variable combination rules; The symbolic regression algorithm is used to perform operations such as generating, mutating, and recombinating the symbolic expression search space to construct a set of candidate expression structures. The prediction performance of the candidate expression structure set is calculated on a cross-scenario dataset, and the expression structure complexity index is constructed by combining the corresponding structure depth, number of symbols and combinatorial redundancy to form a comprehensive evaluation function. Based on the comprehensive evaluation function, the candidate expression structure set is retained, mutated, and updated. Multiple rounds of expression evolution iterations are performed until the evaluation function value converges, at which point the running expression model is determined.

3. The method for monitoring the operation of unmanned heavy-duty vehicles based on big data according to claim 1, characterized in that, The process of constructing a first operational baseline based on stable operational data from the initial deployment of the running expression model, and simultaneously setting a dynamic sliding mechanism for the running expression model to construct and activate a second operational baseline when the output stability of the running expression model meets the conditions, specifically includes: The running expression model is deployed to the on-board edge computing module of the unmanned heavy-duty vehicle. Real-time running data within a continuous running cycle is collected in the early stage of model deployment, and the real-time running data is input into the running expression model to obtain the prediction output results of the continuous time series. The mean and fluctuation range within the sliding window are calculated for the predicted output results. The fluctuation amplitude and equilibrium stability within a specified time length are determined. If the stability determination condition is met, the first operating baseline is constructed based on the predicted output results for the corresponding time period. A dynamic sliding mechanism for the expression is set, which includes sliding detection window parameters, adaptive sampling interval parameters, and residual offset rate threshold parameters for identifying changes in residual trends. When the residual sequence between the predicted output and the real-time running data satisfies the stability condition, the second running baseline construction logic is initiated, and the predicted output within the current sliding window is used to reconstruct the stability expression to generate the second running baseline.

4. The method for monitoring the operation of unmanned heavy-duty vehicles based on big data according to claim 1, characterized in that, The step of inputting real-time running data into the running expression model to generate predicted output values, and calculating the residual differences with the first running baseline and the second running baseline respectively to form a residual evolution trajectory based on a sliding time window specifically includes: The real-time running data is input into the running expression model deployed in the vehicle edge computing module. The expression structure constructed in the running expression model is parsed and calculated item by item to generate the predicted output value at the current time point. Extract the baseline reference output values ​​of the first and second operating baselines at corresponding time points, calculate the difference between the predicted output value of the running expression model and the first operating baseline reference output value as the first type of residual difference, and the difference between the predicted output value and the second operating baseline reference output value as the second type of residual difference; Set the sliding time window parameters, and construct a sliding time window sequence with a fixed length and sliding step size during the operation. Sample the first type of residual difference and the second type of residual difference at continuous time points in each window to form a residual sequence. The residual difference values ​​recorded in each sliding time window are arranged in chronological order to construct the first residual evolution trajectory and the second residual evolution trajectory, respectively.

5. The method for monitoring the operation of unmanned heavy-duty vehicles based on big data according to claim 1, characterized in that, The step of inputting multidimensional residual features into the CUSUM judgment module, performing cumulative statistic updates and sensitivity parameter self-adjustment operations, and dynamically generating degradation trend judgment results specifically includes: Extract the mean value of residual changes, fluctuation amplitude, gradient direction and change frequency within the sliding time window from the first residual evolution trajectory and the second residual evolution trajectory, and construct a multidimensional residual feature vector; The multidimensional residual feature vector is fed into the CUSUM decision module. The CUSUM decision module updates the cumulative statistics of the residual feature vector at each time step and records the offset direction and increase in real time according to the cumulative change trend. During the cumulative statistic update process, the sensitivity parameter set in the CUSUM judgment module is dynamically calculated based on the degree of change in the residual characteristic variance within the sliding time window, and the upper and lower control thresholds of the cumulative statistic are automatically adjusted. When the cumulative statistical value reaches the judgment threshold set by the current sensitivity parameter, the CUSUM judgment module outputs the degradation trend judgment result for the current time period.

6. The method for monitoring the operation of unmanned heavy-duty vehicles based on big data according to claim 5, characterized in that, The control upper and lower thresholds of the self-adjusting cumulative statistic specifically include: Based on the residual sequence within the sliding time window constructed from the first residual evolution trajectory and the second residual evolution trajectory, the time difference score of the change in the mean residual in adjacent time periods is calculated, and the time difference score is defined as the change in the error steering rate. The change in error steering rate is used as a sensitivity parameter to construct an adaptive decision threshold for controlling the trigger threshold of the CUSUM cumulative statistic. The threshold is calculated using the following formula: h k =h0+γ·|Δθ k |; Among them, h k The adaptive threshold value at the current moment, h0 is the initial threshold value, γ is the sensitivity adjustment coefficient, and Δθ is the sensitivity adjustment coefficient. k This represents the change in the error steering rate within the k-th sliding time window.

7. The method for monitoring the operation of unmanned heavy-duty vehicles based on big data according to claim 1, characterized in that, The triggering conditions for outputting multi-level early warning signals and operation abnormality markers when the triggering conditions are met include: When the current cumulative statistic in the CUSUM judgment module is greater than the adaptive judgment threshold value calculated from the change in error guidance rate, the first warning condition is triggered. When the fluctuation amplitude and change frequency in the multidimensional residual feature vector simultaneously exceed the corresponding feature threshold within the sliding time window of the first residual evolution trajectory and the second residual evolution trajectory, and maintain the trend consistency for multiple consecutive time windows, the second warning condition is determined to be triggered. Among them, the feature threshold is preset based on the statistical maximum value of the multidimensional residual feature vector in the historical normal operation data during the initialization phase of the edge computing module; When either the first or second warning condition is met, the trend judgment logic control unit generates multi-level warning signals and operation abnormality markers.

8. A big data-based unmanned heavy-duty vehicle operation monitoring system, characterized in that, This includes an edge computing module deployed on the onboard platform of an autonomous heavy-duty vehicle, the edge computing module comprising: The running expression model processing unit is used to receive the power system operating parameters, braking system operating parameters, steering control operating parameters and environmental sensing operating parameters collected during the operation of the unmanned heavy-duty vehicle, and input the above operating parameters into the running expression model to generate predicted output values. The dual-baseline fusion processing unit is used to call the first operating baseline and the second operating baseline, calculate the residual difference for the predicted output value, and generate the first residual evolution trajectory and the second residual evolution trajectory. The multidimensional residual feature extraction unit is used to extract the mean value of residual change, fluctuation amplitude, change frequency and gradient direction from the first residual evolution trajectory and the second residual evolution trajectory based on a sliding time window, and construct a multidimensional residual feature vector; The CUSUM determination module is used to receive the multidimensional residual feature vector, update the cumulative statistics based on the error guidance rate, and perform degradation trend determination according to the adaptive determination threshold, wherein the adaptive determination threshold is constructed based on the change of the error guidance rate. The trend judgment logic control unit is used to determine whether the current cumulative statistics exceed the adaptive judgment threshold, or to determine whether the fluctuation amplitude and change frequency in the multidimensional residual feature vector exceed the initial feature threshold for multiple consecutive sliding time windows, and outputs multi-level early warning signals and operation abnormality markers when either condition is met.

Images