A lattice car column load and fatigue prediction method and system based on reinforcement learning

Abstract

The present application relates to the field of vehicle structure health monitoring and artificial intelligence, in particular to a kind of warehouse grid car column load and fatigue prediction method and system based on reinforcement learning;It contains multi-source sensor data acquisition and independent verification module;System obtains predicted response signal and real physical signal, determines instantaneous drift error;Its core is to calculate cumulative drift index according to drift error, and evaluate model credibility;When model credibility is lower than safety threshold, the system will be adaptively switched to robust safety strategy;The present application solves the problem that black box model cannot perceive and predict failure under unknown working conditions, and realizes real-time monitoring of model drift.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vehicle structure health monitoring and artificial intelligence, in particular to a method and system for predicting the load and fatigue of a vertical column of a van and rack truck based on reinforcement learning. BACKGROUND

[0002] In the operation of engineering vehicles such as van and rack trucks, the vertical column structure bears complex dynamic loads for a long time, and accurate prediction of its fatigue damage is crucial to ensure safe operation. Currently, data-driven models represented by deep reinforcement learning are widely used for fatigue prediction and operation decision optimization. However, these models usually have black box characteristics, and their prediction performance is highly dependent on the completeness of the training data. When the vehicle encounters operating conditions not covered by the training data in actual operation, such as the use of inferior bolts due to supply chain pollution and other unmodeled factors, it will cause a serious model-reality drift between the model prediction and the physical reality. If the system cannot perceive this drift and continues to rely on the failed prediction model to output optimal decisions, it may lead to disastrous misjudgments and cause structural damage or even safety accidents.

[0003] Therefore, how to fully utilize data-driven models to pursue operational efficiency while quantifying their credibility under unknown disturbances in real time and avoiding potential safety risks caused by model failure is a technical problem that needs to be solved in this field. SUMMARY

[0004] To solve the above technical problems, the present application provides a method and system for predicting the load and fatigue of a vertical column of a van and rack truck based on reinforcement learning. Specifically, the technical solution of the present application is as follows:

[0005] A method for predicting the load and fatigue of a vertical column of a van and rack truck based on reinforcement learning, comprising:

[0006] S1, collecting vibration signals of main input sensors;

[0007] S2, collecting load signals of load sensors;

[0008] S3, collecting vehicle attitude signals of on-board units;

[0009] S4, collecting real physical signals of independent verification sensors;

[0010] S5, combining the vibration signals, load signals and vehicle attitude signals, and generating predicted response signals through a pre-set reference dynamic model;

[0011] S6, combining the predicted response signals and real physical signals to calculate instantaneous drift errors;

[0012] S7, based on the instantaneous drift errors, calculating the cumulative drift index through exponential moving average processing.

[0013] S8, determining a model confidence based on the accumulated drift index and a preset drift critical threshold value;

[0014] S9, in response to the model confidence being greater than a preset confidence safety threshold value, executing an optimal efficiency mode using a preset strategy of the deep reinforcement learning fatigue prediction model;

[0015] S10, in response to the model confidence being less than or equal to the preset confidence safety threshold value, executing a predefined robust safety strategy.

[0016] Preferably, the instantaneous drift error is calculated, comprising:

[0017] standardizing the predicted response signal;

[0018] standardizing the real physical signal;

[0019] calculating the absolute error between the standardized predicted response signal and the standardized real physical signal to obtain the instantaneous drift error.

[0020] Preferably, the accumulated drift index is calculated, comprising:

[0021] based on the current instantaneous drift error, the last time's accumulated drift index and a preset smoothing factor, the current accumulated drift index is recursively calculated through an exponential moving average algorithm.

[0022] Preferably, the model confidence is determined, comprising:

[0023] calculating the ratio of the accumulated drift index to the preset drift critical threshold value to obtain a drift ratio value;

[0024] subtracting the drift ratio value from 1 to obtain an intermediate confidence value;

[0025] comparing the intermediate confidence value with 0 and taking the maximum value as the model confidence.

[0026] Preferably, the optimal efficiency mode is executed, comprising:

[0027] using the preset strategy of the deep reinforcement learning fatigue prediction model to maximize a standard reward function and output optimal operation instructions.

[0028] Preferably, the predefined robust safety strategy is executed, comprising:

[0029] outputting a forced degradation operation instruction to place the vehicle in a low-speed-good-road degradation operation mode;

[0030] outputting a forced recall inspection instruction to generate a high-priority physical inspection work order in a vehicle fleet management system.

[0031] A lattice car column load and fatigue prediction system based on reinforcement learning, comprising:

[0032] A data acquisition module for acquiring vibration signals of a main input sensor, load signals of a load sensor, vehicle posture signals of a vehicle-mounted unit, and real physical signals of an independent verification sensor;

[0033] A prediction response unit for combining the vibration signals, the load signals, and the vehicle posture signals, and generating a prediction response signal through a preset reference dynamic model;

[0034] A drift degree calculation unit for combining the prediction response signal and the real physical signal, and calculating an instantaneous drift error;

[0035] A cumulative drift unit for calculating a cumulative drift index based on the instantaneous drift error;

[0036] A credibility determination unit for determining a model credibility based on the cumulative drift index and a preset drift critical threshold;

[0037] A decision execution unit for switching between an optimal efficiency mode and a predefined robust safety strategy based on the model credibility and a preset credibility safety threshold.

[0038] Preferably, the drift degree calculation unit is specifically configured to:

[0039] standardize the prediction response signal;

[0040] standardize the real physical signal;

[0041] calculate an absolute error between the standardized prediction response signal and the standardized real physical signal to obtain the instantaneous drift error.

[0042] Preferably, the cumulative drift unit is specifically configured to:

[0043] recursively calculate the current cumulative drift index based on the current instantaneous drift error, the cumulative drift index at the last time, and a preset smoothing factor through an exponential moving average algorithm.

[0044] Preferably, the decision execution unit is specifically configured to:

[0045] in response to the model credibility being greater than the preset credibility safety threshold, adopt a strategy of a preset deep reinforcement learning fatigue prediction model to execute the optimal efficiency mode;

[0046] in response to the model credibility being less than or equal to the preset credibility safety threshold, execute the predefined robust safety strategy.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. This invention, by introducing independent physical signal verification and comparing with a reference dynamic model, can monitor and quantify the degree of drift between the data-driven model and physical reality in real time, thus solving the problem that black-box models cannot perceive their own prediction failure when faced with unknown working conditions not covered by training data.

[0049] 2. This invention extracts the true trend reflecting model failure by smoothing the instantaneous drift error through exponential moving average and maps it to a standardized model credibility score. This provides a stable, reliable and easy-to-understand quantitative basis for subsequent decision switching, avoiding frequent misjudgments by the decision system caused by instantaneous signal noise or glitches.

[0050] 3. This invention establishes a dynamic decision-making switching mechanism based on model credibility; when the credibility of the core prediction model is high, the system executes the optimal efficiency mode to pursue maximum transportation efficiency and economic benefits; when the credibility is low, the system actively abandons the model and switches to a predefined robust safety strategy, thus achieving antifragile control.

[0051] 4. By implementing robust safety strategies such as forced degradation operation and forced recall checks when the model fails, this invention can ensure that even if the core prediction model fails completely due to unmodeled factors such as supply chain contamination, the vehicle can still be placed in a physically safe state. This proactively avoids structural damage or catastrophic safety accidents that may result from relying on failed model decisions, thus ensuring the physical safety of operations. Attached Figure Description

[0052] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0053] Figure 1 This is a flowchart of the method of the present invention;

[0054] Figure 2 This is a structural diagram of the system of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0056] Example 1:

[0057] Please see Figure 1 A reinforcement learning-based method for predicting the load and fatigue of the uprights of a cargo box truck, comprising:

[0058] S1, collect the vibration signal from the main input sensor;

[0059] S2, collects the load signal from the load sensor;

[0060] S3, collects vehicle attitude signals from the on-board unit;

[0061] S4, acquires and independently verifies the real physical signals of the sensor;

[0062] S5 combines vibration signals, load signals, and vehicle attitude signals, and generates a predicted response signal through a preset reference dynamic model.

[0063] S6, combine the predicted response signal with the real physical signal to calculate the instantaneous drift error;

[0064] S7, based on instantaneous drift error, calculates the cumulative drift exponent through exponential moving average processing;

[0065] S8. Based on the cumulative drift index and the preset drift threshold, the model credibility is determined.

[0066] S9, in response to the model credibility being greater than the preset credibility safety threshold, adopts the preset strategy of deep reinforcement learning fatigue prediction model and executes the optimal efficiency mode.

[0067] S10, in response to the model's credibility being less than or equal to a preset credibility security threshold, execute a predefined robust security policy.

[0068] In this embodiment, the method performs multi-source time-series data acquisition S1-S4:

[0069] S1: Acquire vibration signals from the main input sensor. The main input sensor refers to the core sensor used to capture the dynamic response of the vehicle pillar. In this embodiment, it is preferably an accelerometer installed at a key structural point of the pillar, which collects vibration signals. Used to characterize the vibration state of the column under real operating conditions;

[0070] S2: Acquire the load signal from the load sensor. A load sensor is a sensor installed on the floor of the cargo box that collects load signals. Used to characterize the real-time weight of goods;

[0071] S3: Acquire vehicle attitude signals from the on-board unit. The onboard unit may specifically include an onboard GPS and an inertial measurement unit (IMU), which collect vehicle attitude signals. Used to characterize road gradient and vehicle driving posture;

[0072] S4: Acquire the real physical signals of the independently verified sensor. An independent verification sensor refers to a completely independent sensor that is not used for training or input to the reference dynamic model in S5 or the DRL model in S9. Its technical purpose is to provide an independent physical benchmark for evaluating the model's predictive accuracy. In this embodiment, it is preferably deployed at critical connection points on the cargo truck's uprights, such as bolted connections susceptible to supply chain contamination. High-frequency acoustic emission (AE) sensors or high-frequency strain gauges can be used to capture the real physical signals generated by microscopic material damage, such as the propagation of microcracks caused by substandard bolts. ;

[0073] System Execution Model - Quantitative Evaluation of Reality Drift S5-S8:

[0074] S5: The system combines the vibration signals collected above. Load signal and vehicle attitude signals Through a preset reference dynamic model Generate predictive response signal ;

[0075] Reference dynamic model This refers to a supervised learning model, such as a recurrent neural network (RNN) or a state-space model, trained based on historical normal operating condition data, such as a dataset of qualified connecting bolts; this model It does not participate in the DRL strategy decision-making of S9; its technical purpose is to create a digital reference point for learning. , , and The normal physical mapping relationship between them;

[0076] Predicted response signal Refers to the model In the current input The signal that the sensor should output is predicted based on the normal physical laws it has learned and independently verified.

[0077] S6: The system incorporates the predicted response signal With real physical signals Calculate instantaneous drift error S7: To eliminate transient noise interference, the system uses the transient drift error obtained in the previous step. The cumulative drift index is calculated using the exponential moving average (EMA). S8: The system is based on the cumulative drift index obtained in the previous step. and the preset drift threshold Determine the model's credibility ;

[0078] The system executes antifragile decision control S9-S10:

[0079] The system defines a trustworthiness security threshold. This threshold refers to a parameter preset according to engineering risk preferences, such as 0.7, which represents the lowest model credibility that the system can tolerate; its source is the engineering parameters set according to risk preferences.

[0080] S9: When responding to model credibility Greater than the preset trustworthiness and security threshold When the system determines that the core prediction model, i.e., the DRL model, is reliable, the system then uses a pre-set deep reinforcement learning fatigue prediction model. The strategy is to execute the optimal efficiency mode;

[0081] S10: When responding to model confidence Less than or equal to the preset trustworthiness and security threshold When the system determines that the core prediction model is no longer reliable, for example, due to unmodeled supply chain factors such as substandard bolts causing severe model-reality drift, the system will execute a predefined robust safety strategy. ;

[0082] This embodiment introduces an independent model—the real-time drift monitoring closed loop S4-S8—to quantify the core prediction model in real time. Credibility in the face of unknown interference Based on this credibility, the system dynamically switches between the optimal efficiency mode S9 and the robust safety mode S10. This solves the problem caused by the black-box nature of the DRL model in S9, which relies solely on data-driven models. In real-world scenarios, the system may make catastrophic misjudgments if it encounters inferior components caused by supply chain contamination. This method enables the system to have antifragile properties, pursuing maximum efficiency while ensuring model credibility, and actively avoiding risks when the model fails, thus ensuring physical safety.

[0083] Example 2:

[0084] Calculating instantaneous drift error includes:

[0085] Standardize the predicted response signal;

[0086] Standardize the real physical signals;

[0087] The instantaneous drift error is obtained by calculating the absolute error between the standardized predicted response signal and the standardized real physical signal.

[0088] According to a reinforcement learning-based method for predicting the load and fatigue of a cargo box truck column in Example 1, the instantaneous drift error is calculated by: standardizing the predicted response signal; standardizing the real physical signal; and calculating the absolute error between the standardized predicted response signal and the standardized real physical signal to obtain the instantaneous drift error.

[0089] In this embodiment, the calculation of instantaneous drift error in step S6 of embodiment 1 is specifically defined;

[0090] To eliminate and Interference from potential inconsistencies in dimensions or differences in amplitude may affect the system's prediction of the response signal. Standardization processing is performed, and the actual physical signals are... Standardize the process;

[0091] Standardization refers to using the same mathematical transformation method, such as Z-score normalization, which involves subtracting the mean and dividing by the standard deviation to convert both signals into dimensionless signal amplitudes.

[0092] The predicted response signal after system calculation and standardization is denoted as The normalized real physical signal is denoted as The absolute error between them yields the instantaneous drift error. ;

[0093] The calculation method is as follows:

[0094]

[0095] in,

[0096] Instantaneous drift error, dimensionless, calculated in this step;

[0097] The standardized predicted response signal is dimensionless and is calculated and standardized by step S5.

[0098] : Standardized real physical signal, dimensionless, obtained by acquisition and standardization in step S4;

[0099] Technical motivation: This formula is used to capture model predictions in real time. With physical reality The instantaneous deviation between them;

[0100] This embodiment calculates the error before performing the error calculation. and Mandatory, identical standardization processes ensured the accuracy of subsequent calculations. It is a normalized error value that is unaffected by the original signal dimensions, providing a stable, reliable, and dimensionally consistent input for the subsequent calculation of the cumulative drift index S7.

[0101] Example 3:

[0102] Calculating the cumulative drift index includes:

[0103] Based on the current instantaneous drift error, the cumulative drift index of the previous moment, and the preset smoothing factor, the current cumulative drift index is recursively calculated using the exponential moving average algorithm.

[0104] According to a reinforcement learning-based method for predicting the load and fatigue of a cargo box column in Embodiment 1, the cumulative drift index is calculated, including: based on the current instantaneous drift error, the cumulative drift index of the previous moment, and a preset smoothing factor, the current cumulative drift index is recursively calculated using an exponential moving average algorithm.

[0105] In this embodiment, the calculation of the cumulative drift index in step S7 of embodiment 1 is specifically defined;

[0106] The system is based on the current instantaneous drift error Cumulative drift index from S6 and the previous time step The smoothing factor is stored in system memory and preset. The current cumulative drift index is obtained by recursively calculating the exponential moving average (EMA) algorithm. ;exist At the initial moment, the cumulative drift index It was initialized to 0;

[0107] The calculation method is as follows:

[0108]

[0109] in,

[0110] The cumulative drift exponent at time t is dimensionless and is recursively calculated in this step.

[0111] The cumulative drift exponent at the previous moment is dimensionless and stores the result from the previous calculation cycle.

[0112] The current instantaneous drift error is dimensionless and is calculated in step S6.

[0113] The preset smoothing factor is dimensionless and its value range is [value range missing]. An adjustable hyperparameter;

[0114] Parameter source and basis: The value is derived from offline backtesting of historical data, including normal operating conditions and known interference conditions such as substandard bolts. Its selection is based on achieving the optimal balance between response sensitivity and noise suppression; smaller values... The values ​​will produce a smoother result. However, the response is slow; larger ones It has a fast response time, but poor noise suppression capability;

[0115] Technical motivation: The purpose of this EMA algorithm is to filter... High-frequency noise in the signal is used to extract the potential trend of model-reality drift, avoiding frequent misjudgments in decision systems S9 and S10 due to instantaneous disturbances.

[0116] This embodiment introduces the EMA algorithm to mitigate the instantaneous error of high-frequency fluctuations. Smoothing is a cumulative drift index that reflects the true trend. This greatly enhances the robustness of drift assessment, prevents the system from erroneously switching to robust safe mode due to brief signal glitches or noise interference, and improves the stability of decision-making.

[0117] Example 4:

[0118] Determining model credibility includes:

[0119] Calculate the ratio of the cumulative drift index to the preset drift threshold to obtain the drift ratio;

[0120] Subtract the drift ratio from 1 to obtain the median confidence value;

[0121] Compare the median confidence level with 0, and take the maximum value as the model confidence level.

[0122] According to the reinforcement learning-based method for predicting the load and fatigue of the column of a cargo box truck in Example 1, the model credibility is determined by: calculating the ratio of the cumulative drift index to a preset drift critical threshold to obtain the drift ratio; subtracting the drift ratio from 1 to obtain the median credibility value; comparing the median credibility value with 0, taking the maximum value, and determining it as the model credibility.

[0123] In this embodiment, specific limitations are made to step S8 of embodiment 1, which determines the model's credibility.

[0124] The system defines a drift threshold. This threshold is a key preset threshold, representing the level of tolerance the system can tolerate. The maximum value is derived from simulation tests or historical data analysis of extreme working conditions such as inferior bolts. For example, it can be set to the value under known disturbance conditions. Reaching 80% of the stable peak value; Dimensions and Same dimensionless;

[0125] The calculation process is as follows:

[0126] Calculate the cumulative drift index From S7 and the preset drift threshold The ratio of the two values ​​is used to obtain the drift ratio. This ratio reflects the current degree of drift.

[0127] Subtracting the drift ratio from 1 yields the median confidence level. This step reverses high drift to low confidence.

[0128] Compare the median confidence level with 0, and take the maximum value to determine the model's confidence level. This step ensures that the credibility is not negative.

[0129] The calculation method is as follows, which is the mathematical synthesis of the above steps:

[0130]

[0131] in,

[0132] The model confidence level at time t is dimensionless, ranging from 0 to 1, and is calculated in this step.

[0133] The current cumulative drift index is dimensionless and is calculated in step S7.

[0134] Drift threshold, dimensionless, obtained from historical data or simulation calibration;

[0135] Technical motivation: The purpose of this formula is to account for a potentially unbounded cumulative drift exponent. This is mapped to a bounded 0-1 and intuitive credibility score; when drifting... When the value is 0, the credibility is 1, meaning 100% credible; when... Increase and exceed When the credibility drops to 0, it becomes completely untrustworthy;

[0136] This embodiment uses linear normalization and sets a lower limit. This method represents an abstract, unbounded drift index. This was converted into a standardized credibility score ranging from 0% to 100%, with clear engineering significance. This provides a clear and quantifiable basis for subsequent decision-making on switching between S9 and S10.

[0137] Example 5:

[0138] Execute the optimal efficiency mode, including:

[0139] The strategy employs a pre-defined deep reinforcement learning fatigue prediction model to maximize the standard reward function and output optimal operational instructions.

[0140] According to a reinforcement learning-based method for predicting the load and fatigue of a warehouse truck column in Embodiment 1, the optimal efficiency mode is executed, including: adopting a strategy of a preset deep reinforcement learning fatigue prediction model, with the goal of maximizing the standard reward function, and outputting the optimal operation instructions;

[0141] In this embodiment, the optimal efficiency mode for step S9 in embodiment 1 is specifically defined;

[0142] When the system determines that the model is credible, that is... At that time, the system will use a preset deep reinforcement learning fatigue prediction model. Strategies;

[0143] Deep reinforcement learning fatigue prediction model It is an agent model based on Deep Reinforcement Learning (DRL), which uses... The state inputs are S1-S3, and the goal is to predict the fatigue accumulation rate and maximize the standard reward function. To achieve the goal, output the optimal operational instructions;

[0144] Standard reward function yes The core objective function optimized by the model during training and decision-making; its calculation method can be defined as:

[0145]

[0146] in,

[0147] Standard reward, a scalar value, calculated using this formula;

[0148] Transportation efficiency indicators are derived from the ton-kilometers of transport recorded by the vehicle's onboard system.

[0149] M(t): Predicted maintenance costs, sourced from... The fatigue accumulation rate predicted by the model based on its internal state or output. The estimated future maintenance costs; for example, may be The model predicts the column damage value, while It can be specifically defined as a... Related functions, such as ,in, This is the maintenance cost coefficient. Damage index;

[0150] Weighting coefficients, their function is to unify and The dimensions of the units are adjusted and their relative importance is adjusted.

[0151] To clarify and The calibration process can construct a calibration dataset containing multiple sets of historical operational data, each set corresponding to a specific operational strategy. And the historical average efficiency index generated under this strategy. and historical average maintenance costs Here, and These are independent variables established for calibration purposes, and their signs are the same as those of variables during model runtime. and Differentiate between them; find a set of... through optimization algorithms, such as grid search or Bayesian optimization. and Values ​​that maximize the value on the calibration dataset. strategy The strategy that best aligns with expert experience or simulation results is used to determine the optimal strategy. and The value of ;

[0152] Technical Motivation: This reward function is for DRL agents. Provide a scalar optimization objective that enables conflicting transportation efficiencies. Maintenance costs Find a balance between goals;

[0153] In this mode, The optimal operational instructions output by the model may include: allowing transportation at maximum rated load, allowing selection of the shortest route even if road conditions are worse, etc., to achieve... Maximize;

[0154] This embodiment ensures the maximization of economic benefits of vehicle operation by executing this DRL strategy under the premise of model reliability; the system uses the DRL model to find the optimal balance between efficiency and cost, and achieves maximum transportation efficiency while ensuring that fatigue damage is controllable.

[0155] Example 6:

[0156] Implement predefined robust security policies, including:

[0157] Output a forced degraded operation command to put the vehicle into a low-speed, good-road degraded operation mode;

[0158] Output a mandatory recall inspection command to generate a high-priority physical inspection work order in the fleet management system.

[0159] According to a reinforcement learning-based method for predicting column load and fatigue of a cargo box vehicle in Embodiment 1, a predefined robust safety strategy is executed, including: outputting a forced degraded operation command to place the vehicle in a low-speed-good road degraded operation mode; and outputting a forced recall inspection command to generate a high-priority physical inspection work order in the fleet management system.

[0160] In this embodiment, the execution of a predefined robust security strategy in step S10 of embodiment 1 is specifically defined;

[0161] When the system determines that the model is no longer reliable, that is... At that time, the system determines The optimal solution from the S9 output may conceal significant physical risks; the system will automatically discard it. The strategy then shifts to executing a predefined robust security policy. ; It is a pre-programmed set of fixed rules with the primary goal of ensuring physical security, and it does not rely on the predictions of the DRL model;

[0162] This strategy includes the following mandatory downgrade operations:

[0163] Forced Degradation Operation Command Output: The system immediately sends a command to the vehicle's ECU to place the vehicle into a low-speed, good-road degraded operation mode. The technical purpose of this is to immediately reduce the dynamic load and fatigue accumulation on the vehicle, especially the pillars, prioritizing physical safety rather than transportation efficiency. ;

[0164] Output a mandatory recall inspection instruction: The system immediately generates a high-priority physical inspection work order in the fleet management system; the technical purpose of this is to require the vehicle to return to the maintenance station for a special physical flaw detection inspection after the current transportation task is completed, such as a special inspection of whether the column connecting bolts are defective due to supply chain contamination.

[0165] This embodiment achieves antifragile control by defining a set of mandatory, conservative degradation operations; when a significant discrepancy is detected between the model and reality... If the value is too low, the system no longer trusts the model's predictions and instead proactively switches to a suboptimal strategy with high safety redundancy. This effectively avoids the risk of structural damage or safety accidents caused by model failure, such as unmodeled supply chain contamination or sensor malfunction.

[0166] Example 7:

[0167] Please see Figure 2 A reinforcement learning-based system for predicting the load and fatigue of column supports in a cargo box truck includes:

[0168] The data acquisition module is used to acquire vibration signals from the main input sensor, load signals from the load sensor, vehicle attitude signals from the on-board unit, and real physical signals from the independent verification sensor.

[0169] The predictive response unit is used to combine vibration signals, load signals and vehicle attitude signals, and generate predictive response signals through a preset reference dynamic model.

[0170] The drift calculation unit is used to combine the predicted response signal with the actual physical signal to calculate the instantaneous drift error;

[0171] The cumulative drift unit is used to calculate the cumulative drift exponent based on the instantaneous drift error;

[0172] The credibility determination unit is used to determine the credibility of the model based on the cumulative drift index and a preset drift threshold.

[0173] The decision execution unit is used to switch between the optimal efficiency mode and the predefined robust security policy based on the model's credibility and a preset credibility security threshold.

[0174] A reinforcement learning-based system for predicting the load and fatigue of a cargo box truck pillar, based on any one of the reinforcement learning-based methods for predicting the load and fatigue of a cargo box truck pillar in Examples 1-6, includes: a data acquisition module for acquiring vibration signals from a main input sensor, load signals from a load sensor, vehicle attitude signals from an onboard unit, and real physical signals from an independent verification sensor; a prediction response unit for generating a prediction response signal by combining the vibration signal, load signal, and vehicle attitude signal through a preset reference dynamic model; a drift calculation unit for calculating the instantaneous drift error by combining the prediction response signal and the real physical signal; a cumulative drift unit for calculating the cumulative drift exponent based on the instantaneous drift error; a credibility determination unit for determining the model credibility based on the cumulative drift exponent and a preset drift critical threshold; and a decision execution unit for switching between an optimal efficiency mode and a predefined robust safety strategy based on the model credibility and a preset credibility safety threshold.

[0175] In this embodiment, a system is provided, configured to execute the methods of embodiments 1-6 described above; the system can be deployed on an edge computing unit of a vehicle or on a cloud server; the system includes:

[0176] Data acquisition module: It is used to acquire signals S1-S4 in Example 1; specifically including:

[0177] Vibration signal from main input sensor Such as from an accelerometer;

[0178] Load signal of load sensor ;

[0179] Vehicle attitude signals of the on-board unit Such as from GPS / IMU;

[0180] Independently verify the true physical signal of the sensor Such as from an AE sensor;

[0181] Predictive response unit: electrically connected to the data acquisition module for receiving... It is used to execute S5: combining vibration signals, load signals, and vehicle attitude signals, through a preset reference dynamic model. Generate predictive response signal ;

[0182] Drift calculation unit: electrically connected to the data acquisition module for receiving data. and predictive response unit receive It is used to perform S6: combining the predicted response signal with the real physical signal to calculate the instantaneous drift error. ;

[0183] Accumulated drift unit: electrically connected to the drift calculation unit to receive... It is used to perform S7: calculating the cumulative drift exponent based on the instantaneous drift error. ;

[0184] Credibility determination unit: electrically connected to the cumulative drift unit for receiving... It is used to execute S8: based on the cumulative drift index and a preset drift threshold. Determine the model's credibility ;

[0185] Decision execution unit: Electrically connected to the credibility determination unit to receive... It is used to execute S9 and S10: based on model confidence. and preset trust and security thresholds Activated in optimal efficiency mode Activation of policies and predefined robust security policies Switching between strategies; the unit's output is electrically connected to the vehicle's ECU or fleet management system;

[0186] To further clarify, the aforementioned unit or module may be a logical functional module implemented by executing computer program instructions stored in non-volatile memory such as hard disk or flash memory through a central processing unit (CPU) or a graphics processing unit (GPU).

[0187] This embodiment constructs a complete automated system of perception-prediction-evaluation-decision through the collaborative work of the above modules. The system can automatically execute the antifragile prediction and control method of Embodiment 1, providing reliable, adaptive, and closed-loop system-level technical protection for the safe operation of the cargo truck.

[0188] Example 8:

[0189] The drift calculation unit is specifically used for:

[0190] Standardize the predicted response signal;

[0191] Standardize the real physical signals;

[0192] The instantaneous drift error is obtained by calculating the absolute error between the standardized predicted response signal and the standardized real physical signal.

[0193] According to Embodiment 7, a reinforcement learning-based prediction system for the load and fatigue of a cargo truck column includes a drift calculation unit, which is specifically used for: standardizing the predicted response signal; standardizing the real physical signal; and calculating the absolute error between the standardized predicted response signal and the standardized real physical signal to obtain the instantaneous drift error.

[0194] In this embodiment, the drift degree calculation unit in Embodiment 7 is specifically defined to correspond to the method in Embodiment 2;

[0195] This drift calculation unit integrates a standardized processor and an error calculator; specifically, it is used for:

[0196] The standardized processor processes the prediction response signal received from the prediction response unit. Perform standardization processing, such as Z-score normalization;

[0197] Standardized processors handle the real physical signals received from the data acquisition module. Standardization processing and adoption The same standardization method;

[0198] The error calculator calculates the absolute error between the standardized predicted response signal and the standardized real physical signal. The instantaneous drift error is obtained. and output it to the cumulative drift unit;

[0199] This embodiment ensures that the data entering the error calculator has consistent dimensions and comparable amplitudes by incorporating a standardization function into the drift calculation unit, thereby improving the instantaneous drift error. The accuracy and reliability of the calculations provide higher-quality data input for subsequent system decisions.

[0200] Example 9:

[0201] Accumulated drift unit, specifically used for:

[0202] Based on the current instantaneous drift error, the cumulative drift index of the previous moment, and the preset smoothing factor, the current cumulative drift index is recursively calculated using the exponential moving average algorithm.

[0203] According to Embodiment 7, a reinforcement learning-based system for predicting the load and fatigue of a cargo box column includes an accumulated drift unit, which is specifically used to: recursively calculate the current accumulated drift index based on the current instantaneous drift error, the accumulated drift index of the previous moment, and a preset smoothing factor, using an exponential moving average algorithm.

[0204] In this embodiment, the cumulative drift unit in Embodiment 7 is specifically defined to correspond to the method of Embodiment 3;

[0205] The cumulative drift unit internally implements an exponential moving average (EMA) filter and includes a storage mechanism. The register or memory location; specifically used for:

[0206] Based on the current instantaneous drift error input from the drift degree calculation unit The cumulative drift index read from its own memory at the previous moment and preset smoothing factor Using the exponential moving average algorithm The current cumulative drift index is obtained by recursive calculation. After the calculation is completed, the new Store in memory, overwrite And output it to the credibility determination unit;

[0207] This embodiment achieves the processing of instantaneous drift signals by embedding the EMA algorithm into the cumulative drift unit. Effective smoothing and noise reduction; this makes the cumulative drift index of the output... It can more realistically reflect the long-term trend of model drift, prevent the system from making unstable decisions due to transient noise interference, and improve the overall decision robustness of the system.

[0208] Example 10:

[0209] The decision execution unit is specifically used for:

[0210] In response to the model's credibility exceeding a preset credibility safety threshold, the strategy of the preset deep reinforcement learning fatigue prediction model is adopted to execute the optimal efficiency mode.

[0211] In response to a model's credibility being less than or equal to a preset credibility security threshold, a predefined robust security policy is executed.

[0212] According to Embodiment 7, a reinforcement learning-based system for predicting the load and fatigue of a cargo box column is provided. The decision execution unit is specifically used to: respond to a model confidence level greater than a preset confidence level safety threshold, adopt a preset deep reinforcement learning fatigue prediction model strategy, and execute the optimal efficiency mode; and respond to a model confidence level less than or equal to the preset confidence level safety threshold, execute a predefined robust safety strategy.

[0213] In this embodiment, the decision execution unit in embodiment 7 is specifically defined so that it corresponds to S9 and S10 in embodiment 1 and its specific implementation in embodiments 5 and 6;

[0214] The decision execution unit contains one decision arbitrator and two strategy executors, respectively corresponding to... Strategies and Strategy; specifically used for:

[0215] The decision arbitrator compares the model confidence levels input from the confidence determination unit in real time. and the preset trust and security threshold stored locally. ;

[0216] Response to model credibility Greater than the preset trustworthiness and security threshold Decision arbitrator activated Policy executor; this executor employs a pre-defined deep reinforcement learning fatigue prediction model. The strategy is to execute the optimal efficiency mode, such as maximizing output. The operational instructions are sent to the ECU;

[0217] Response to model credibility Less than or equal to the preset trustworthiness and security threshold Decision arbitrator activated Strategy executor; this executor executes predefined robust safety policies such as outputting forced degradation and forced recall commands to the ECU and fleet management system;

[0218] This embodiment sets a credibility-based [function / mechanism] in the decision execution unit. The switching switch, or decision arbitrator, enables the system to operate at maximum efficiency. and absolute safety Intelligent and automatic arbitration is performed between the two strategies; this realizes the core logic of antifragile control, ensuring that the system can make the most reasonable response under various operating conditions, regardless of whether the model is reliable or not, which greatly improves the system's operational security and adaptability to operating conditions.

[0219] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the load and fatigue of the column of a cargo box truck based on reinforcement learning, characterized in that, include: S1, collect the vibration signal from the main input sensor; S2, collects the load signal from the load sensor; S3, collects vehicle attitude signals from the on-board unit; S4, Collect the real physical signals from the independent verification sensor; the independent verification sensor is deployed at the connection part of the column of the cargo truck, and adopts a high-frequency acoustic emission sensor or a high-frequency strain gauge to capture the real physical signals generated by the micro-damage of the material; S5 combines vibration signals, load signals, and vehicle attitude signals, and generates a predicted response signal through a preset reference dynamic model. S6, combine the predicted response signal with the real physical signal to calculate the instantaneous drift error; S7, based on instantaneous drift error, calculates the cumulative drift exponent through exponential moving average processing; S8. Based on the cumulative drift index and the preset drift threshold, the model credibility is determined. S9, in response to the model credibility being greater than the preset credibility safety threshold, adopts the preset strategy of deep reinforcement learning fatigue prediction model and executes the optimal efficiency mode. S10, in response to the model's credibility being less than or equal to a preset credibility security threshold, execute a predefined robust security policy.

2. The method for predicting the load and fatigue of the column of a cargo box truck based on reinforcement learning according to claim 1, characterized in that, Calculating instantaneous drift error includes: Standardize the predicted response signal; Standardize the real physical signals; The instantaneous drift error is obtained by calculating the absolute error between the standardized predicted response signal and the standardized real physical signal.

3. The method for predicting the load and fatigue of the column of a cargo box truck based on reinforcement learning according to claim 1, characterized in that, Calculating the cumulative drift index includes: Based on the current instantaneous drift error, the cumulative drift index of the previous moment, and the preset smoothing factor, the current cumulative drift index is recursively calculated using the exponential moving average algorithm.

4. The method for predicting the load and fatigue of the column of a cargo box truck based on reinforcement learning according to claim 1, characterized in that, Determining model credibility includes: Calculate the ratio of the cumulative drift index to the preset drift threshold to obtain the drift ratio; Subtract the drift ratio from 1 to obtain the median confidence value; Compare the median confidence level with 0, and take the maximum value as the model confidence level.

5. The method for predicting the load and fatigue of the column of a cargo box truck based on reinforcement learning according to claim 1, characterized in that, Execute the optimal efficiency mode, including: The strategy employs a pre-defined deep reinforcement learning fatigue prediction model to maximize the standard reward function and output optimal operational instructions.

6. The method for predicting the load and fatigue of the column of a cargo box truck based on reinforcement learning according to claim 1, characterized in that, Implement predefined robust security policies, including: Output a forced degraded operation command to put the vehicle into a low-speed, good-road degraded operation mode; Output a mandatory recall inspection command to generate a high-priority physical inspection work order in the fleet management system.

7. A reinforcement learning-based system for predicting the load and fatigue of a cargo box truck column, based on the reinforcement learning-based method for predicting the load and fatigue of a cargo box truck column as described in any one of claims 1-6, characterized in that... include: The data acquisition module is used to acquire vibration signals from the main input sensor, load signals from the load sensor, vehicle attitude signals from the vehicle unit, and real physical signals from the independent verification sensor. The independent verification sensor is deployed at the connection part of the column of the cargo truck and uses a high-frequency acoustic emission sensor or a high-frequency strain gauge to capture real physical signals generated by microscopic damage to the material. The predictive response unit is used to combine vibration signals, load signals and vehicle attitude signals, and generate predictive response signals through a preset reference dynamic model. The drift calculation unit is used to combine the predicted response signal with the actual physical signal to calculate the instantaneous drift error; The cumulative drift unit is used to calculate the cumulative drift exponent based on the instantaneous drift error; The credibility determination unit is used to determine the credibility of the model based on the cumulative drift index and a preset drift threshold. The decision execution unit is used to switch between the optimal efficiency mode and the predefined robust security policy based on the model's credibility and a preset credibility security threshold.

8. The reinforcement learning-based system for predicting the load and fatigue of a cargo box truck column as described in claim 7, characterized in that, The drift calculation unit is specifically used for: Standardize the predicted response signal; Standardize the real physical signals; The instantaneous drift error is obtained by calculating the absolute error between the standardized predicted response signal and the standardized real physical signal.

9. A reinforcement learning-based system for predicting the load and fatigue of a cargo box truck column, as described in claim 7, is characterized in that... Accumulated drift unit, specifically used for: Based on the current instantaneous drift error, the cumulative drift index of the previous moment, and the preset smoothing factor, the current cumulative drift index is recursively calculated using the exponential moving average algorithm.

10. A reinforcement learning-based system for predicting the load and fatigue of a cargo box truck column, as described in claim 7, is characterized in that... The decision execution unit is specifically used for: In response to the model's credibility exceeding a preset credibility safety threshold, the strategy of the preset deep reinforcement learning fatigue prediction model is adopted to execute the optimal efficiency mode. In response to a model's credibility being less than or equal to a preset credibility security threshold, a predefined robust security policy is executed.

Images