Method and device for early warning of climbing instability of tracked unmanned vehicle, medium

Abstract

The application relates to a method and device for early warning of climbing instability of a tracked unmanned vehicle, and a medium, and relates to the technical field of climbing control of unmanned vehicles. The early warning method comprises the following steps: a digital twin model of the tracked unmanned vehicle is constructed, the digital twin model is fused with a motor dynamic model and a whole-vehicle dynamics model, and virtual scene data is generated. Real-time data of a plurality of sources of the tracked unmanned vehicle is acquired, the real-time data comprises vehicle attitude data, environment data and driving system data. A time convolution network prediction model based on transfer learning is constructed, the virtual scene data is used as source domain data for pre-training, the real-time data is used as target domain data for fine-tuning, and a target domain prediction model is formed. Current data is input into the target domain prediction model, and a predicted vehicle state parameter prediction sequence is output. According to the vehicle state parameter prediction sequence and a preset safety threshold, the instability risk is evaluated, and a warning signal is sent out, wherein the generation logic of the warning signal is evaluated and optimized through a warning performance loss function.

Description

Technical Field

[0001] This invention relates to the field of unmanned vehicle hill-climbing control technology, and more specifically, to a method, device, and medium for early warning of hill-climbing instability of a tracked unmanned vehicle. Background Technology

[0002] In outdoor operations of high-speed tracked unmanned vehicles, complex working conditions and harsh road conditions significantly increase the difficulty of driving control. Due to their high drive force requirements, these vehicles often employ multi-motor distributed drive systems. However, when climbing steep inclines, factors such as sudden acceleration and road slippage can easily cause the center of gravity to shift backward or one track to lift off the ground, leading to longitudinal rollover or side rollover risks. Therefore, there is an urgent need for a method that can monitor and provide real-time warnings of instability during inclines to improve the active safety and operational reliability of vehicles in complex environments.

[0003] In existing technologies, several solutions have attempted to address the problem of vehicle instability during hill climbing. For example, patent CN104340116A proposes a real-time measurement method for the maximum gradeability of a vehicle based on tire pressure to estimate the load. This method obtains tire pressure using a tire pressure monitoring device and calculates the vehicle's load, then combines this with a mechanical model to estimate the safe gradeability, thereby generating a warning. Another patent, CN118596993A, addresses the differential steering process of a dual-motor driven tracked vehicle. It employs digital twin and machine learning technologies, fusing data from physical sensors and virtual models to predict key dynamic parameters and provide tiered warnings.

[0004] However, these existing methods still have significant shortcomings. On the one hand, the instability warning mechanism for high-speed tracked unmanned vehicles going uphill is not yet perfect, lacking the ability to predict risks in scenarios with instantaneous high power and high torque. On the other hand, existing sensor fusion technologies are unable to make forward-looking judgments on instability and rollover risks, resulting in delayed warnings or frequent false alarms, failing to meet the real-time and accurate safety requirements in complex environments. Summary of the Invention

[0005] The present invention provides a method, device, and medium for early warning of instability during hill climbing of a tracked unmanned vehicle, in order to improve at least one of the above-mentioned technical problems.

[0006] In a first aspect, the present invention provides a method for early warning of instability of a tracked unmanned vehicle climbing a slope, which includes steps S1 to S5.

[0007] S1. Construct a digital twin model of the tracked unmanned vehicle. The digital twin model integrates the motor dynamic model and the vehicle dynamic model to generate virtual scene data covering different working conditions.

[0008] S2. Acquire real-time data from multiple sensors of the tracked unmanned vehicle during the actual hill climbing process. The real-time data includes vehicle attitude data, environmental data, and drive system data.

[0009] S3. Construct a temporal convolutional network prediction model based on transfer learning, use the virtual scene data as source domain data to pre-train the prediction model, and use the real-time data as target domain data to fine-tune the prediction model to form a target domain prediction model.

[0010] S4. Input the current real-time data into the target domain prediction model and output the predicted sequence of vehicle state parameters within the future set time window.

[0011] S5. Based on the comparison result between the predicted sequence of vehicle state parameters and the preset safety threshold, assess the risk of instability and issue a warning signal, wherein the generation logic of the warning signal is evaluated and optimized through a warning performance loss function.

[0012] Specifically, S3 includes:

[0013] A multi-channel temporal convolutional network (TCN) is used as the basic architecture to build the TCN prediction model. The TCN includes causal convolutional layers, dilated convolutional layers, and residual modules, and a multi-head temporal attention layer is connected after the residual blocks.

[0014] The TCN prediction model is trained using the virtual scene data to learn the ability to extract temporal features. The TCN prediction model also includes a fully connected layer and a multi-head attention mechanism layer sequentially connected after the temporal convolutional network.

[0015] When migrating to the target domain, the underlying general feature layer of the TCN network is frozen by quantifying the difference in feature distribution between the source domain and the target domain, and only the mid-to-high-level working condition adaptation feature layer is fine-tuned using real data.

[0016] As a further aspect of the present invention, S1 specifically includes:

[0017] Simulink is used to establish the dynamic equations of the motor, the control system, and the sensor simulation module, generating the motor's motion equations and output characteristics.

[0018] RecurDyn was used to build the mechanical structure model, transmission system model, and ramp environment model of the tracked vehicle.

[0019] The motor model defined in Simulink is added as a torque input source to the drive wheel of the RecurDyn model for co-simulation, generating virtual vehicle attitude change curves, drive wheel torque and power consumption data.

[0020] As a further aspect of the present invention, S2 specifically includes:

[0021] The vehicle's three-axis acceleration and angular velocity information are obtained through the IMU.

[0022] The angular velocity and displacement information of the tracks on both sides are measured by a track encoder installed at the motor output end.

[0023] The distribution of contact load on the vehicle is monitored by contact load sensors located at the track track rollers or tensioner rollers.

[0024] The motor current changes are monitored using a current sensor.

[0025] Obtain the slope angle and terrain parameters in front using LiDAR or a camera.

[0026] As a further aspect of the present invention, the early warning performance loss function is a composite loss function, which includes a missed reporting penalty, a false alarm penalty, an early warning timing penalty, and a decision consistency penalty.

[0027] Early warning performance loss function: In the formula This is a penalty item for failure to report. This is a penalty item for false alarms. This is a penalty item for early warning. This is a penalty for ensuring consistency in decision-making. for The weight. for The weight. for The weight. for The weight.

[0028] in .

[0029] As a further aspect of the present invention, the underreporting penalty item The calculation formula is as follows.

[0030] .

[0031] .

[0032] In the formula This is the function for indicating missed reports. This is a risk severity function. This is a theoretical early warning label.

[0033] This is the duration weighting function. Deadline The number of consecutive missed steps. Virtual early warning signals generated for the model.

[0034] The false alarm penalty item The calculation formula is as follows.

[0035] .

[0036] .

[0037] In the formula This is a false alarm indicator function. The base penalty value is fixed.

[0038] The penalty item for the timing of the warning The calculation formula is as follows.

[0039] .

[0040] .

[0041] .

[0042] In the formula This is the minimum duration threshold. This is the maximum duration threshold. This provides advance warning. This is the optimal lead time. This is the tolerance parameter. This is the time when the risk begins. This is the warning trigger time. The sampling interval is denoted as .

[0043] The decision consistency penalty item The calculation formula is as follows.

[0044] .

[0045] .

[0046] In the formula This indicates the change in warning level between adjacent time steps. This is the penalty function.

[0047] when The penalty value is 0.

[0048] when Increase the penalty value when it occurs.

[0049] As a further aspect of the present invention, S5 specifically includes at least one of the following:

[0050] Longitudinal instability detection: If the predicted pitch angle is greater than or equal to the set proportion of the critical angle within a set time in the future, or the predicted pitch velocity exceeds the threshold, a longitudinal risk warning is triggered.

[0051] Lateral instability detection: Combining the roll angle measured by the IMU with the support stability index calculated by the contact load sensor, if the predicted roll angle exceeds the threshold or the support stability index is less than the set value, a lateral risk warning is triggered.

[0052] Adhesion instability detection: Combining motor current and track displacement data, if it is predicted that the motor current will continue to exceed the steady-state requirements and the slippage rate will be greater than the set threshold, an adhesion instability risk warning will be triggered.

[0053] Secondly, the present invention provides an early warning device for the instability of a tracked unmanned vehicle climbing a slope, which includes a virtual data module, a real-time data module, a training module, a prediction module and an evaluation module.

[0054] The virtual data module is used to construct a digital twin model of the tracked unmanned vehicle. The digital twin model integrates the motor dynamic model and the vehicle dynamic model to generate virtual scene data covering different working conditions.

[0055] The real-time data module is used to acquire real-time data from multiple sensors of the tracked unmanned vehicle during the actual hill climbing process. The real-time data includes vehicle attitude data, environmental data, and drive system data.

[0056] The training module is used to construct a temporal convolutional network prediction model based on transfer learning. The prediction model is pre-trained using the virtual scene data as source domain data, and fine-tuned using the real-time data as target domain data to form a target domain prediction model.

[0057] The prediction module is used to input the current real-time data into the target domain prediction model and output a predicted sequence of vehicle state parameters within a future set time window.

[0058] The evaluation module is used to evaluate the instability risk and issue a warning signal based on the comparison result between the predicted sequence of vehicle state parameters and the preset safety threshold. The generation logic of the warning signal is evaluated and optimized through a composite loss function.

[0059] Thirdly, the present invention provides a tracked unmanned vehicle, which includes a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement a method for early warning of instability during hill climbing of a tracked unmanned vehicle, as described in any paragraph of the first aspect.

[0060] Fourthly, the present invention provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform a method for early warning of instability during hill climbing as described in any paragraph of the first aspect.

[0061] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0062] This invention constructs an integrated technology chain of "digital twin - perception - prediction - early warning," fusing the Simulink motor dynamic model with the RecurDyn vehicle dynamic model to achieve high-fidelity simulation of multi-domain coupling between the motor and the whole vehicle. Combined with real-time perception data from multi-source heterogeneous sensors, it forms a continuous characterization and closed-loop analysis of the vehicle's attitude and dynamic state. This enables the prediction of key parameters such as pitch angle and roll angle in the future and triggers graded early warnings accordingly. Ultimately, it achieves adaptive, forward-looking, and accurate early warning for various instability modes such as pitch, roll, and adhesion, significantly improving the active safety and operability of the climbing process.

[0063] Meanwhile, this invention, with transfer learning at its core, enables the model to be pre-trained using diverse virtual extreme working condition data generated in batches within a digital twin environment, and to achieve efficient adaptation with limited real-world working condition data. This reduces reliance on high-risk real-vehicle data collection and enhances cross-working-condition generalization capabilities. At the warning and decision-making level, a multi-sensor cross-validation and physical-logic fusion mechanism is employed to improve robustness and accuracy under sensor noise, single-point failures, and complex environmental interference. Furthermore, by quantifying risk coefficients, using multi-level thresholds, and employing a composite loss function that integrates penalties for missed detections, false alarms, warning timing, and decision consistency, objectives such as "safety priority, timely warning, and stable decision-making" are directly transformed into optimizable decision criteria. This improves the timeliness and reliability of warnings while suppressing unnecessary false alarms and control jitter caused by oscillations near thresholds, thus balancing safety, operational continuity, and system lifespan. Attached Figure Description

[0064] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some specific embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0065] Figure 1 This is a logic diagram of an early warning method for hill climbing instability of tracked unmanned vehicles based on transfer learning.

[0066] Figure 2 This is a schematic diagram of the Temporal Convolutional Network (TCN).

[0067] Figure 3 This is a residual link network diagram of TCN.

[0068] Figure 4 This is a flowchart of an early warning method for tracked unmanned vehicles experiencing instability while climbing a hill.

[0069] Figure 5 This is a diagram illustrating the process of optimizing early warning performance. Detailed Implementation

[0070] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0071] Example 1, please refer to Figures 1 to 5 The first embodiment of the present invention provides a method for early warning of instability during hill climbing by a tracked unmanned vehicle, which can be executed by the tracked unmanned vehicle. Specifically, it is executed by one or more processors in the tracked unmanned vehicle to implement steps S1 to S5.

[0072] S1. Construct a digital twin model of the tracked unmanned vehicle. The digital twin model integrates the motor dynamic model and the vehicle dynamic model to generate virtual scene data covering different working conditions.

[0073] S2. Acquire real-time data from multiple sensors of the tracked unmanned vehicle during the actual hill climbing process. The real-time data includes vehicle attitude data, environmental data, and drive system data.

[0074] S3. Construct a temporal convolutional network prediction model based on transfer learning, use the virtual scene data as source domain data to pre-train the prediction model, and use the real-time data as target domain data to fine-tune the prediction model to form a target domain prediction model.

[0075] S4. Input the current real-time data into the target domain prediction model and output the predicted sequence of vehicle state parameters within the future set time window.

[0076] S5. Based on the comparison result between the predicted sequence of vehicle state parameters and the preset safety threshold, assess the risk of instability and issue a warning signal, wherein the generation logic of the warning signal is evaluated and optimized through a warning performance loss function.

[0077] Figure 1 This is a logic diagram of an early warning method for hill climbing instability of a tracked unmanned vehicle based on transfer learning, which mainly shows the working principle and process of the early warning method.

[0078] Specifically, to meet the proactive early warning needs of unmanned high-speed tracked vehicles under complex working conditions and improve their active safety during hill climbing, this embodiment uses the unmanned high-speed tracked vehicle as the physical entity and the Simulink model of the motor and the RecurDyn model of the tracked vehicle as virtual entities. It constructs an integrated technical solution of digital twin model-entity perception-prediction-early warning, using transfer learning as the core method. This allows the model to accurately predict the driving trend of the unmanned high-speed tracked vehicle even with limited target scene data, achieving adaptive early warning. The digital twin architecture integrates Simulink and RecurDyn to construct multi-domain models of motor dynamics and vehicle dynamics. The perception module collects real-time state data of the vehicle-environment-drive system through multi-source heterogeneous sensors. The prediction module adopts a time-series prediction model based on TCN and uses transfer learning to pre-train the TCN prediction model using source domain virtual data, transferring low-level features to the target domain. After validating the effectiveness of the transfer layer, if there is a significant difference from reality, the diversity of the virtual scene dataset needs to be increased. If the transfer layer is deemed effective, it is transferred to the target domain and combined with the feature layer finely adjusted according to real scene working condition data to form the target domain TCN prediction model. The sensed data is input into the target domain model to predict parameters such as pitch and roll angles. The early warning module not only provides multi-level warnings based on the comparison of predicted values ​​with thresholds, but also evaluates the early warning performance by calculating the warning performance in real time using a composite loss function that integrates penalties for missed detections, false alarms, warning timing, and decision consistency. This guides and optimizes early warning decisions, forming an integrated technical solution from high-fidelity simulation and virtual-real transfer learning to intelligent decision optimization. Ultimately, this achieves adaptive, forward-looking, and accurate early warnings for various instability modes such as pitch, roll, and adhesion.

[0079] Based on the above embodiments, in an optional embodiment of the present invention, S1 specifically includes S11 to S13.

[0080] S11. Use Simulink to establish the dynamic equations of the motor, the control system and sensor simulation module, and generate the motor motion equations and output characteristics.

[0081] S12. Use RecurDyn to build the mechanical structure model, transmission system model, and ramp environment model of the tracked vehicle.

[0082] S13. Add the motor model defined in Simulink as a torque input source to the drive wheel of the RecurDyn model for joint simulation, and generate virtual vehicle attitude change curves, drive wheel torque and power consumption data.

[0083] Specifically, the digital twin model combines Simulink and Recurdyn simulation models. Simulink is used to build the dynamic equations of the motor, control system, and sensor simulation modules. The Recurdyn simulation model is used to simulate the physical characteristics and dynamic behavior of the tracked vehicle, such as its mass, center of gravity, forces during hill climbing, and attitude changes. The digital twin model enables multi-domain, multi-scale fusion modeling of the motor and the overall system, combining actual operation with theoretical simulation to improve the accuracy of the model.

[0084] First, the dynamic equations and control system of the motor are established using Simulink software. Appropriate modules are selected to establish the motor's motion equations, sensor simulation modules, and corresponding control logic. The mathematical operations module is used to establish the dynamic equations related to the motor's motion. The "Sensor Simulation" module in Simulink is used to simulate the output of the motor's internal sensors, including output power and frequency. The "Transfer Function" module in Simulink is used to establish the motor's control system, ensuring that the motor can be controlled according to the predetermined motion equations. The "Scope" module is used to monitor the output of the motor model in real time to ensure the effectiveness of the model's motion equations and control logic.

[0085] Then, use RecurDyn's built-in track module to create a tracked vehicle model, establish a simplified model of the transmission system, and set up the slope environment, sensors, simulation parameters, etc., to simulate real road conditions.

[0086] Finally, the Simulink simulation model and the Recurdyn model are fused. The motor model defined in Simulink is added to the drive wheels of the tracked vehicle. By fusion simulation of the tracked vehicle climbing a slope, the vehicle's attitude change curve, torque and power consumption of the drive wheels are generated.

[0087] Based on the above embodiments, in an optional embodiment of the present invention, S2 specifically includes steps S21 to S25.

[0088] S21. Obtain the vehicle's three-axis acceleration and angular velocity information through the IMU.

[0089] S22. Measure the angular velocity and displacement information of the tracks on both sides by using the track encoder installed at the motor output end.

[0090] S23. Monitor the vehicle contact load distribution by using contact load sensors located at track track rollers or tension rollers.

[0091] S24. Monitor the changes in motor current using a current sensor.

[0092] S25. Obtain the slope angle and terrain parameters ahead using LiDAR or a camera.

[0093] Step S2, as the sensing module, is the core of the system's data input. It is responsible not only for collecting the vehicle's dynamic parameters, environmental parameters, and drive system parameters during the climb, but also for optimizing the raw data collected by each sensor to avoid data deviations caused by sensor malfunctions or transient interference, thus providing stable and reliable data for subsequent modules. The various sensors and their functions are shown in Table 1.

[0094] Table 1. Sensor categories and their functions.

[0095]

[0096] Specifically, the operating environment of high-speed unmanned tracked vehicles is quite complex, and different operating conditions can affect their driving status. In this embodiment, data is first acquired in real time through various sensors, and then the data collected under different operating conditions is correlated with the operating status. The possible operating conditions involved are shown in Table 2.

[0097] Table 2. Operating conditions of high-speed unmanned tracked vehicles.

[0098]

[0099] The driving status of a tracked vehicle climbing a slope can be determined by its pitch angle. yaw angle Motor torque Motor current Track slip ratio It is represented by five parameters.

[0100] Based on the above embodiments, in an optional embodiment of the present invention, S3 specifically includes steps S31 to S33.

[0101] S31. A multi-channel temporal convolutional network (TCN) is used as the basic architecture to build the TCN prediction model. The TCN includes causal convolutional layers, dilated convolutional layers, and residual modules, and a multi-head temporal attention layer is connected after the residual blocks.

[0102] S32. Train the TCN prediction model using the virtual scene data to enable it to learn temporal feature extraction capabilities. The TCN prediction model also includes a fully connected layer and a multi-head attention mechanism layer sequentially connected after the temporal convolutional network.

[0103] S33. When migrating to the target domain, the underlying general feature layer of the TCN network is frozen by quantifying the difference in feature distribution between the source domain and the target domain, and only the mid-to-high-level working condition adaptation feature layer is fine-tuned using real data.

[0104] As a prediction module, S3 is an important part of the tracked vehicle's safe operation under steep inclines and complex terrain conditions. It processes the data collected by the sensor module to provide early warning of instability risks.

[0105] The prediction module employs a temporal prediction model based on Temporal Convolutional Network (TCN) (referred to as the TCN prediction model). It utilizes transfer learning to pre-train the TCN prediction model using virtual data from the source domain, transferring low-level features to the target domain. After validating the effectiveness of the transfer layer, if there are significant discrepancies with reality, the diversity of the virtual scene dataset needs to be increased. If the transfer layer is deemed effective, it is transferred to the target domain and combined with a feature layer fine-tuned based on real-world scenario data to form the target domain TCN prediction model.

[0106] Transfer learning sets virtual scene parameters according to the application goal, collects the input features and output labels required by TCN, and uses the massive data generated by the virtual environment to train the core temporal feature extraction capability of the TCN prediction model, enabling the model to accurately learn temporal patterns and adapt to different virtual working conditions.

[0107] TCN learns cross-domain general features such as noise filtering and basic time series trends at the bottom layer, captures long-term dependency patterns in the middle layer, and uses a multi-head attention mechanism to dynamically assign weights to features at different historical moments in the upper layer, dynamically focusing on key historical periods.

[0108] During the model migration process, the range of features that need to be adapted is identified by quantifying statistical distribution differences, and the response differences of each layer of the model are identified at the feature layer to determine reusable layers and layers that need fine-tuning. Finally, through error analysis, the optimization direction is clarified, and efficient migration is achieved with minimal data cost.

[0109] Within the target domain, time-series dependencies of varying durations are modeled hierarchically, and predictions for the next 10 seconds are made sequentially in 0.5-second increments. The predicted distribution is then output and transmitted to the risk assessment and early warning module. The predicted values ​​are compared with a set threshold, and an early warning is triggered when the predicted curve continuously approaches or exceeds the threshold within a future time window.

[0110] TCN can effectively capture long-distance temporal dependencies. Figure 2 This is a schematic diagram of the structure of a Temporal Convolutional Network (TCN). Figure 2 As shown, its main structure includes causal convolution, dilated convolution, and residual modules. Causal convolution ensures that the output of the convolution process depends only on elements in the input sequence that occurred before it. Dilated convolution refers to interval sampling during the convolution process, with the sampling rate determined by the dilation coefficient. This expands the receptive field without increasing computational complexity, making it easier to capture longer dependencies.

[0111] Residual structures are used to connect layers in a TCN network, thus avoiding problems such as gradient explosion and network degradation. The principle of residual structures is as follows: Figure 3 As shown, the Temporal Convolutional Network (TCN) consists of a series of interconnected dilated causal convolutional layers, weight normalization layers, ReLU activation function layers, regularization layers, dilated causal convolutional layers, weight normalization layers, ReLU activation function layers, and regularization. The residual structure concatenates the input of the TCN with the output of the regularization layer. Specifically, a ReLU activation function is added to the top of each convolutional layer to introduce non-linearity, and each convolutional layer is followed by weight normalization and a random deactivation layer. The output of the residual block is obtained by adding the regularized output to the input across layers. The sum of the results obtained from the output of the previous residual block after a 1×1 convolution operation and two layers of dilated causal convolution is input into the next residual block.

[0112] The temporal convolutional network (TCN) in this embodiment adopts a multi-channel TCN network, which uses its convolution operation to capture the coupling and temporal dependency between temporal data of different channels at the same time. By setting different convolution kernel sizes, multi-time scale information extraction is achieved. Furthermore, a multi-head temporal attention layer is added after the last or several residual blocks of the TCN.

[0113] It receives temporal features extracted by TCN and outputs an attention-weighted feature sequence. This allows the model to not only capture long-range dependencies through dilated convolutions but also adaptively emphasize key historical evidence.

[0114] Based on the above embodiments, in an optional embodiment of the present invention, the early warning module is a key component to ensure the smooth operation of the tracked vehicle. By analyzing and predicting the data on the state trend of the tracked vehicle's operation, it provides judgment and early warning of potential instability risks.

[0115] This embodiment employs a multi-parameter fusion-based instability risk coefficient for quantitative assessment, achieving a tiered response from trend warning to emergency alarm. The basic logic follows a three-layer interaction mechanism: environment, vehicle, and dynamics. At the environmental level, slope, road surface unevenness, and adhesion coefficient directly affect the support and traction boundaries. At the vehicle level, pitch angle, roll angle, and their rate of change determine whether the center of gravity deviates from the stable polygon. At the dynamics level, motor current, torque demand, and track slip ratio reflect whether the adhesion limit has been exceeded. When the predicted curves of these quantities continuously approach or exceed thresholds within a future time window, a warning signal is generated and transmitted to the control system.

[0116] To further enhance the reliability and security of the early warning system, this embodiment constructs a composite loss function to optimize early warning performance. This loss function includes four components: missed detection penalty, false alarm penalty, early warning timing penalty, and decision consistency penalty. Through reasonable weight allocation and function design, the prediction accuracy of the early warning system is improved, leading more directly to the optimal early warning decision.

[0117] Penalties for underreporting. For periods in history where instability or high-risk conditions have actually occurred, severe penalties will be imposed if the sequence predicted by the model fails to trigger the corresponding level of warning.

[0118] The penalty for failure to report The calculation formula is as follows.

[0119] .

[0120] .

[0121] In the formula This is the function for indicating missed reports. This is a risk severity function. This is a theoretical early warning label.

[0122] This is the duration weighting function. Deadline The number of consecutive missed steps. Virtual early warning signals generated for the model.

[0123] . . .

[0124] Penalties for underreporting are the highest priority to ensure that any real risk is detected.

[0125] False alarm penalties. Unnecessary warnings should be moderately suppressed to improve system availability while ensuring safety. During safe driving periods, if a model erroneously triggers a high-level warning, a light penalty should be imposed.

[0126] The false alarm penalty item The calculation formula is as follows.

[0127] .

[0128] .

[0129] In the formula This is a false alarm indicator function. The base penalty value is fixed.

[0130] The fixed base penalty value ensures that the safety principle of "better to give false alarms than to miss alarms" is upheld, while avoiding setting the warning threshold too high due to excessive penalties for false alarms.

[0131] Timing penalty. Guiding the system to issue an early warning at the optimal time is valuable because of the just-right lead time, allowing the control system sufficient but not redundant reaction time.

[0132] The penalty item for the timing of the warning The calculation formula is as follows.

[0133] .

[0134] .

[0135] .

[0136] In the formula This is the minimum duration threshold. This is the maximum duration threshold. This provides advance warning. This is the optimal lead time. This is the tolerance parameter. This is the time when the risk begins. This is the warning trigger time. The sampling interval is denoted as . .

[0137] This function penalizes zero within the ideal window, and the penalty increases with the square of the deviation after the window is exceeded. This provides a clear direction for optimization and avoids training instability caused by hard boundaries.

[0138] Decision consistency penalty. This ensures that the early warning decisions generated by continuous forecast sequences are smooth and stable, avoiding oscillations around thresholds that could cause control command jitter, thus improving ride comfort and control system lifespan.

[0139] The decision consistency penalty item The calculation formula is as follows.

[0140] .

[0141] .

[0142] In the formula This indicates the change in warning level between adjacent time steps. This is the penalty function.

[0143] when The penalty value is 0. Increase the penalty value when it occurs.

[0144] Specifically, when hour, .when hour, .when hour, .when hour, .

[0145] This embodiment comprehensively considers penalties for missed detections, false alarms, early warning timing, and decision consistency, transforming abstract security requirements into concrete and optimizable mathematical objectives. This enables the system to make high-quality security decisions, resulting in the early warning performance loss function:

[0146] The early warning performance loss function is a composite loss function, which includes penalties for missed reports, false alarms, early warning timing, and decision consistency.

[0147] Early warning performance loss function: In the formula This is a penalty item for failure to report. This is a penalty item for false alarms. This is a penalty item for early warning. This is a penalty for ensuring consistency in decision-making. for The weight. for The weight. for The weight. for The weight.

[0148] The weighting coefficients follow the principle of prioritizing safety: Security >> Availability > Timing Optimization ≥ Output Stability. >> indicates much greater than. Preferably, Timing Optimization ≈ Output Stability.

[0149] In one specific embodiment This weighting configuration ensures that the penalty for missed reports guides the optimization direction, while effectively balancing false positives and early warning timing.

[0150] Based on the above embodiments, in an optional embodiment of the present invention, S5 includes at least one of the following:

[0151] Longitudinal instability detection: If the predicted pitch angle is greater than or equal to the set proportion of the critical angle within a set time in the future, or the predicted pitch velocity exceeds the threshold, a longitudinal risk warning is triggered.

[0152] Lateral instability detection: Combining the roll angle measured by the IMU with the support stability index calculated by the contact load sensor, if the predicted roll angle exceeds the threshold or the support stability index is less than the set value, a lateral risk warning is triggered.

[0153] Adhesion instability detection: Combining motor current and track displacement data, if it is predicted that the motor current will continue to exceed the steady-state requirements and the slippage rate will be greater than the set threshold, an adhesion instability risk warning will be triggered.

[0154] Specifically, the following are the criteria for judging slope instability involved in this invention.

[0155] 1. When the tracked vehicle is climbing a slope, the fused road gradient and IMU attitude data are input into the target domain TCN prediction model. The migrated TCN outputs pitch angle data for the next 0.5~10 seconds. If the current pitch angle of the tracked vehicle is within the safe operating threshold, the tracked vehicle's driving attitude is normal, and no warning needs to be triggered. If the predicted pitch angle velocity will exceed the threshold within the next 1 second, a rollover alarm is triggered, and emergency braking or attitude compensation is performed. If the pitch angle is greater than or equal to 80% of the critical angle within 3 seconds, it indicates that the tracked vehicle is currently entering the longitudinal risk zone, and a longitudinal risk warning is triggered in advance, with the system performing torque limiting or traction control. If the pitch angle exceeds the critical angle after 3 seconds, a longitudinal instability trend warning is output.

[0156] 2. When the tracked vehicle is climbing a slope with a lateral angle, the contact load sensor and IMU data are input into the target domain TCN prediction model. The migrated TCN outputs the roll angle data for the next 0.5~10 seconds. If the current roll angle of the tracked vehicle is within the safe operating threshold, the tracked vehicle's driving posture is normal, and no warning needs to be triggered. If the predicted roll angle velocity will exceed the threshold within 1 second, a rollover alarm is triggered, and emergency braking or attitude compensation is performed. If the roll angle is greater than or equal to 80% of the roll angle critical value or the support stability index is less than 0.15 within 3 seconds, the tracked vehicle is currently entering the lateral risk zone, and a lateral risk warning is triggered in advance. If the predicted roll angle exceeds the critical condition after 3 seconds, a lateral instability trend warning is output.

[0157] 3. Input the current data from the motor current sensor and track encoder, along with the track displacement data, into the target domain TCN prediction model. The TCN, after being trained in the source domain and migrated to a frozen layer, outputs time-series data of the motor current and track displacement reflected by the track encoder for the next 0.5~10 seconds. If the predicted current value is within the normal operating range, the tracked vehicle will travel normally on the slope without triggering a warning. If the TCN outputs a state sequence exceeding a threshold within the next 3 seconds, an instability alarm is triggered, and the tracked vehicle performs emergency braking or attitude compensation. If the output motor current for the next 3~10 seconds continuously exceeds the steady-state requirement by more than 20% and the slippage rate is >0.3, it is determined to be an adhesion instability risk, triggering an adhesion instability risk warning.

[0158] This embodiment utilizes high-fidelity digital twin technology to generate virtual and real data on the driving conditions of a high-speed electric-drive tracked chassis. By jointly simulating a Simulink motor model as the torque input source with a RecurDyn vehicle model, a digital twin system capable of accurately reflecting the coupling effect between motor dynamics and vehicle dynamics is constructed. This system can generate massive amounts of high-dimensional time-series data in a virtual environment, covering different gradients, loads, road adhesion coefficients, and extreme conditions, providing a rich, low-cost, and risk-free source domain dataset for subsequent model training.

[0159] This embodiment is based on the TCN architecture and hierarchical transfer learning for temporal prediction. This architecture utilizes causal dilated convolution to capture long-range temporal dependencies and dynamically weights features from different historical moments through a multi-head temporal attention mechanism, focusing on the historical periods most relevant to the future state. For model training, a hierarchical transfer learning approach is adopted: first, the model is pre-trained on virtual source domain data generated by a digital twin, enabling it to learn cross-domain general features such as noise filtering and basic attitude change trends. Subsequently, by quantitatively analyzing the feature distribution differences between the source and target domains, the underlying general feature layer of the TCN network is identified and frozen, with only the mid-to-high-level working condition adaptation feature layers being fine-tuned. This method achieves efficient and reliable model transfer from a virtual environment to a real-world scenario with minimal real-vehicle data cost, outputting predicted sequences of key attitude parameters such as pitch and roll angles for the next 0.5 to 10 seconds.

[0160] This embodiment proposes an instability detection mechanism based on multi-sensor cross-validation and physical model fusion. This mechanism is not a simple data weighting, but rather a collaborative verification based on different physical principles.

[0161] For rollover risk, the geometric roll angle measured by the IMU and the support stability index (SSI) calculated by the contact load sensor are jointly analyzed. When the trends of the two are consistent (e.g., the roll angle increases and the SSI decreases), the rollover risk is confirmed as high. If they deviate, the sensor reliability assessment process is initiated to avoid misjudgment due to single sensor failure. For slippage risk, the motor torque demand reflected by the current sensor and the slip ratio calculated by the track encoder / vehicle speed sensor are simultaneously verified. By establishing a torque-slip coupling relationship model, the normal torque increase caused by the increase in slope and the abnormal slippage caused by adhesion failure are distinguished. For impact risk, the road surface unevenness perceived by LiDAR / visual perception is used as the excitation input and compared with the vehicle body angular acceleration response measured by the IMU. Through a preset road surface-attitude transfer function, it is determined whether the current vehicle body dynamic response is within a reasonable range, thereby achieving early identification of impact instability caused by sudden road surface changes.

[0162] This embodiment is based on intelligent early warning decision-making technology using quantified risk coefficients and composite loss functions. By setting multi-level thresholds, it achieves tiered responses for normal, trend-based, and emergency alarms. Furthermore, to directly optimize the decision-making quality of the early warning system, this scheme designs a composite loss function that includes penalties for missed alarms, false alarms, early warning timing, and decision consistency. This function mathematically concretizes abstract goals such as safety priority, timely early warning, and decision stability, guiding the model to automatically learn the optimal early warning strategy during training, rather than simply optimizing attitude prediction accuracy.

[0163] The early warning method for the unstable climbing slope of the tracked unmanned vehicle in this embodiment has the following advantages.

[0164] 1. Achieve truly proactive early warning and significantly extend safety response time. Through a technology chain combining digital twin-generated data, TCN time-series prediction models, and hierarchical transfer learning, the system can perform high-precision, long-term predictions of key attitude parameters such as vehicle pitch and roll angles within the next 0.5-10 seconds. This allows the system to issue trend warnings 1-3 seconds before instability actually occurs, providing the control system with ample time for decision-making and execution, fundamentally upgrading the early warning mode from "passive response" to proactive prevention.

[0165] 2. Enhance the model's adaptability and reliability under complex operating conditions. An innovative virtual-real fusion training paradigm is adopted, utilizing high-fidelity digital twins to generate massive amounts of diverse virtual extreme operating condition data for pre-training the model. Then, hierarchical transfer learning is used to fine-tune the model with a small amount of real vehicle data. This not only significantly reduces reliance on high-risk real vehicle data collection but also gives the model stronger cross-condition generalization capabilities. Simultaneously, through a multi-sensor cross-validation mechanism, sensor information from different physical principles is integrated for collaborative discrimination, significantly improving the robustness and accuracy of early warning judgments under sensor noise, single-point failures, or complex environmental interference.

[0166] 3. Smarter and more precise early warning decisions, achieving an optimal balance between safety and efficiency. A smart decision-making framework is proposed, combining quantitative risk coefficients, multi-level thresholds, and a composite loss function for optimization. By constructing a risk coefficient formula that integrates multiple state variables, continuous and quantitative assessment of instability risks is achieved. Differentiated early warning levels and control strategies are triggered by setting multi-level thresholds. Furthermore, a composite loss function is designed, incorporating penalties for missed detections, false alarms, timing, and consistency, guiding the model to learn and make optimal early warning decisions that are timely yet not excessive, safe yet stable, during the model training phase. This allows the system to minimize unnecessary interventions while ensuring absolute safety, thereby improving the continuity and efficiency of vehicle operations.

[0167] Example 2: The present invention provides an early warning device for the instability of a tracked unmanned vehicle climbing a slope, which includes a virtual data module, a real-time data module, a training module, a prediction module and an evaluation module.

[0168] The virtual data module is used to construct a digital twin model of the tracked unmanned vehicle. The digital twin model integrates the motor dynamic model and the vehicle dynamic model to generate virtual scene data covering different working conditions.

[0169] The real-time data module is used to acquire real-time data from multiple sensors of the tracked unmanned vehicle during the actual hill climbing process. The real-time data includes vehicle attitude data, environmental data, and drive system data.

[0170] The training module is used to construct a temporal convolutional network prediction model based on transfer learning. The prediction model is pre-trained using the virtual scene data as source domain data, and fine-tuned using the real-time data as target domain data to form a target domain prediction model.

[0171] The prediction module is used to input the current real-time data into the target domain prediction model and output a predicted sequence of vehicle state parameters within a future set time window.

[0172] The evaluation module is used to evaluate the instability risk and issue a warning signal based on the comparison result between the predicted sequence of vehicle state parameters and the preset safety threshold. The generation logic of the warning signal is evaluated and optimized through a composite loss function.

[0173] Example 3: This invention provides a tracked unmanned vehicle, which includes a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement a method for early warning of instability during hill climbing of a tracked unmanned vehicle, as described in any paragraph of Example 1.

[0174] Example 4: This invention provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform a method for early warning of instability during hill climbing, as described in any paragraph of Example 1.

[0175] Obviously, the embodiments described above are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0176] In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0177] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0178] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0179] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0180] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0181] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0182] The terms "first" and "second" used in the embodiments are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0183] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early warning of instability in a tracked unmanned vehicle climbing a slope, characterized in that, Includes the following steps: S1. Construct a digital twin model of the tracked unmanned vehicle. The digital twin model integrates the motor dynamic model and the vehicle dynamic model to generate virtual scene data covering different working conditions. S2. Acquire real-time data from multiple sensors of the tracked unmanned vehicle during the actual hill climbing process. The real-time data includes vehicle attitude data, environmental data, and drive system data. S3. Construct a temporal convolutional network prediction model based on transfer learning. The virtual scene data is used as source domain data to pre-train the prediction model, and the real-time data is used as target domain data to fine-tune the prediction model, forming a target domain prediction model. Specifically, this includes: constructing a temporal convolutional network prediction model using a multi-channel temporal convolutional network as the basic architecture; the temporal convolutional network includes causal convolutional layers, dilated convolutional layers, and residual modules, with a multi-head temporal attention layer connected after the residual modules; training the temporal convolutional network prediction model using the virtual scene data to enable it to learn temporal feature extraction capabilities; the temporal convolutional network prediction model also includes a fully connected layer and a multi-head attention mechanism layer sequentially connected after the temporal convolutional network; when transferring to the target domain, by quantifying the feature distribution differences between the source and target domains, the underlying general feature layer of the temporal convolutional network is frozen, and only the mid-to-high-level working condition adaptation feature layers are fine-tuned using real data; S4. Input the current real-time data into the target domain prediction model and output the predicted sequence of vehicle state parameters within a future set time window. S5. Based on the comparison result between the predicted sequence of vehicle state parameters and the preset safety threshold, assess the risk of instability and issue a warning signal, wherein the generation logic of the warning signal is evaluated and optimized through a warning performance loss function. The early warning performance loss function is a composite loss function, which includes a penalty for missed reporting, a penalty for false reporting, a penalty for early warning timing, and a penalty for decision consistency. Early warning performance loss function: In the formula Penalties for underreporting; This is a penalty item for false alarms; Penalties for early warning timing; This is a penalty for ensuring consistency in decision-making. for The weights; for The weights; for The weights; for The weights; in ; The penalty for failure to report The calculation formula is as follows; ； ； In the formula This is a function to indicate missed reports; For risk severity function; This is a theoretical early warning label; For duration weighting function; Deadline The number of consecutive missed steps; The virtual early warning signal generated for the model; among which, ; ; The false alarm penalty item The calculation formula is as follows; ； ； In the formula This is a false alarm indicator function; A fixed base penalty value; The penalty item for the timing of the warning The calculation formula is as follows; ； ； ； In the formula Minimum duration threshold; The maximum duration threshold; To provide advance warning; For optimal lead time; For tolerance parameters; The time when the risk begins; This refers to the warning trigger time; The sampling interval; The decision consistency penalty item The calculation formula is as follows; ； ； In the formula To indicate the change in warning level between adjacent time steps; For the penalty function; when The penalty value is 0. when Increase the penalty value when it occurs.

2. The early warning method for instability of a tracked unmanned vehicle climbing a slope according to claim 1, characterized in that, S1 specifically includes: Simulink was used to establish the dynamic equations of the motor, the control system and sensor simulation module, and to generate the motor motion equations and output characteristics. RecurDyn was used to build the mechanical structure model, transmission system model, and ramp environment model of the tracked vehicle; The motor model defined in Simulink is added as a torque input source to the drive wheel of the RecurDyn model for co-simulation, generating virtual vehicle attitude change curves, drive wheel torque and power consumption data.

3. The early warning method for instability of a tracked unmanned vehicle climbing a slope according to claim 1, characterized in that, S5 specifically includes at least one of the following situations: Longitudinal instability detection: If the predicted pitch angle is greater than or equal to the set proportion of the critical angle within a set time in the future, or the predicted pitch angle velocity exceeds the threshold, a longitudinal instability risk warning is triggered. Lateral instability detection: Combining the roll angle measured by the IMU and the support stability index calculated by the contact load sensor, if the predicted roll angle exceeds the threshold or the support stability index is less than the set value, a lateral instability risk warning is triggered. Adhesion instability detection: Combining motor current and track displacement data, if it is predicted that the motor current will continue to exceed the steady-state requirements and the slippage rate will be greater than the set threshold, an adhesion instability risk warning will be triggered.

4. The early warning method for instability of a tracked unmanned vehicle climbing a slope according to claim 1, characterized in that, S2 specifically includes: The vehicle's three-axis acceleration and angular velocity information are obtained through the IMU; The angular velocity and displacement information of the tracks on both sides are measured by a track encoder installed at the motor output end; The distribution of vehicle contact load is monitored by contact load sensors located at track track rollers or tension rollers; Monitor motor current changes using a current sensor; Obtain the slope angle and terrain parameters in front using LiDAR or a camera.

5. A warning device for instability during hill climbing in a tracked unmanned vehicle, characterized in that, Used to perform the early warning method for the instability of a tracked unmanned vehicle climbing a slope as described in any one of claims 1 to 4; The early warning device includes: The virtual data module is used to construct a digital twin model of the tracked unmanned vehicle. The digital twin model integrates the motor dynamic model and the vehicle dynamic model to generate virtual scene data covering different working conditions. The real-time data module is used to acquire real-time data from multiple sensors of the tracked unmanned vehicle during the actual hill climbing process. The real-time data includes vehicle attitude data, environmental data, and drive system data. The training module is used to construct a temporal convolutional network prediction model based on transfer learning. The prediction model is pre-trained using the virtual scene data as source domain data, and fine-tuned using the real-time data as target domain data to form a target domain prediction model. The prediction module is used to input the current real-time data into the target domain prediction model and output a predicted sequence of vehicle state parameters within a future set time window. The evaluation module is used to evaluate the instability risk and issue a warning signal based on the comparison result between the predicted sequence of vehicle state parameters and the preset safety threshold. The generation logic of the warning signal is evaluated and optimized through a composite loss function.

6. A tracked unmanned vehicle, characterized in that, It includes a processor, a memory, and a computer program stored in the memory; the computer program can be executed by the processor to implement a method for early warning of instability during hill climbing of a tracked unmanned vehicle as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform an early warning method for instability during hill climbing as described in any one of claims 1 to 4.

Images