Intelligent chariot training state monitoring method and system

Abstract

The invention discloses an intelligent chariot training state monitoring method and system. The method comprises the steps of multi-source data acquisition, chariot training state monitoring model construction, monitoring model performance optimization and chariot training state intelligent monitoring. The invention relates to the technical field of data processing, in particular to an intelligent combat vehicle training state monitoring method and system.According to the scheme, training process morphological characteristics are innovatively introduced to serve as static constraint factors, the sensing ability of morphological changes in the combat vehicle training process is enhanced, and the training state of a combat vehicle is monitored. The accuracy and the stability of offside and border-crossing risk prediction are improved; an improved structured auto-encoder is adopted, so that the recognition precision and stability of the model on offside and border-crossing risks are improved; a double-risk time sequence loss function is designed, an improved optimization algorithm of a sine and cosine strategy of circular chaotic mapping, self-adaptive T distribution disturbance position and self-adaptive weight is introduced, and the accuracy, stability and generalization ability of a monitoring model on offside and border-crossing probability prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to an intelligent method and system for monitoring the training status of combat vehicles. Background Technology

[0002] The intelligent armored vehicle training status monitoring method and system refers to a method that uses artificial intelligence technology to collect multi-dimensional data in real time and combine it with machine learning algorithms to predict probabilistic risks such as offside and boundary crossing that may occur during armored vehicle training. This enables real-time monitoring of the armored vehicle training process, effectively improving the safety, standardization, and overall training efficiency of armored vehicle training, thereby providing intelligent early warning and auxiliary decision support for the training process.

[0003] However, traditional armored vehicle training status monitoring systems suffer from several technical problems. First, they rely solely on a single temporal feature, failing to adequately consider the varying impacts of overall morphological changes during training on the risks of exceeding boundaries and overshooting, leading to insufficient accuracy in overshoot and overshoot prediction. Second, existing models for armored vehicle training status monitoring employ a single temporal feature extraction method and do not differentiate between temporal patterns of different risk types, resulting in inaccurate model output. Third, existing models for armored vehicle training status monitoring suffer from limitations in performance improvement. Traditional loss functions are often designed in a single way, making it difficult to simultaneously consider temporal feature learning and probability prediction accuracy. Furthermore, conventional optimization algorithms are prone to getting trapped in local optima and have low hyperparameter search efficiency, resulting in insufficient accuracy and stability in the model's overshoot and overshoot probability predictions. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent method and system for monitoring the training status of armored vehicles. Addressing the problem in traditional armored vehicle training status monitoring systems that rely solely on single temporal features and fail to adequately consider the impact of overall morphological changes during training on the risks of off-boundary and overstepping, leading to insufficient accuracy in off-boundary and overstepping predictions, this solution innovatively introduces morphological features from the training process as a static constraint factor. These features are then fused with the temporal features of off-boundary and overstepping risks, enhancing the perception of morphological changes during armored vehicle training. This enables the system to maintain efficient and reliable prediction capabilities in complex training scenarios, effectively solving the problems of inaccurate predictions and poor scenario adaptability in traditional models during training. This improves the accuracy and stability of offside and boundary risk prediction, enabling intelligent monitoring during vehicle training. Addressing the technical problem of existing vehicle training status monitoring models using a single temporal feature extraction method and failing to differentiate between temporal patterns of different risk types, leading to inaccurate model output, this solution innovatively employs an improved structured autoencoder combined with a multi-type temporal window annotation method. The improved structured autoencoder extracts general basic temporal features through a three-layer stacked LSTM encoder with residual connections. Then, utilizing a compliance decoder and a risk dual-branch decoder, it differentiates the temporal evolution patterns under different states, significantly improving the discriminative power of temporal features and ensuring accurate separation of compliance baseline features from various risk features, thus enhancing... This solution enhances the model's learning ability in scarce dual-risk scenarios, improving the model's accuracy and stability in identifying offside and boundary risks. It provides highly discriminative temporal support for independent prediction of offside and boundary risks, achieving accurate identification and quantitative prediction of multiple types of risks during vehicle training. Addressing the technical problems of existing models for monitoring vehicle training status, such as the simplistic design of traditional loss functions that struggle to simultaneously consider temporal feature learning and probability prediction accuracy, and the tendency of conventional optimization algorithms to get trapped in local optima and suffer from low hyperparameter search efficiency, leading to insufficient accuracy and stability in offside and boundary probability prediction, this solution innovatively designs a dual-risk temporal loss function. This function weightedly fuses the temporal reconstruction loss term with the dual-probability calibration loss term, constraining the model... This study effectively models the temporal evolution of the vehicle training state, ensuring that the extracted temporal features are sufficiently expressive and distinguishable, making the output probability highly consistent with the actual risk occurrence. It also introduces an improved optimization algorithm using circular chaotic mapping, adaptive T-distribution perturbation positions, and adaptive weights with a sine and cosine strategy. This effectively avoids the problem of getting trapped in local optima during the optimization process and accelerates the convergence speed of the optimal hyperparameter combination. Through these techniques, the accuracy, stability, and generalization ability of the vehicle training state monitoring model in predicting offside and boundary violations are significantly improved. The model's adaptability to different training scenarios and data distribution changes is enhanced, ultimately achieving highly reliable quantitative prediction and intelligent dynamic monitoring of offside and boundary violations risks during vehicle training.

[0005] The technical solution adopted by this invention is as follows: This invention provides an intelligent method for monitoring the training status of combat vehicles, which includes the following steps:

[0006] Step S1: Acquisition of multi-source data;

[0007] Step S2: Construct a vehicle training status monitoring model;

[0008] Step S3: Monitor model performance optimization;

[0009] Step S4: Intelligent monitoring of vehicle training status.

[0010] Further, in step S1, the multi-source data acquisition specifically involves obtaining raw data for vehicle training status monitoring through data acquisition operations, and performing data optimization operations on the raw data to obtain optimized data for vehicle training status monitoring. The raw data for vehicle training status monitoring includes historical training status monitoring data and real-time training status monitoring data. The historical training status monitoring data and real-time training status monitoring data include vehicle operation data, maneuver control data, vehicle training data, training environment data, and personnel operation data.

[0011] The historical training status monitoring data also includes historical tank overrun prediction results and historical tank boundary crossing prediction results.

[0012] The data optimization operations include data cleaning, data normalization, data encoding, and data feature selection.

[0013] Furthermore, in step S2, the construction of the vehicle training status monitoring model specifically includes the following steps:

[0014] Step S21: Extract morphological features of the training process. Specifically, within a preset statistical period, aggregate the training state monitoring data to form a morphological input vector of the training process, and input the morphological input vector of the training process into a morphological feature extraction network based on a multilayer perceptron structure to obtain the morphological features of the training process.

[0015] Step S22: Extract the temporal evolution features of the training states. Specifically, the extraction of temporal evolution features of the training states is based on an improved structured autoencoder, including the following steps:

[0016] Step S221: Construct a time-series input sequence. Specifically, sort the various types of data in the training status monitoring data according to the vehicle identification and timestamp. Then, use the sliding time window method to segment the sorted data to obtain K time-series windows. Label each time-series window with data. Combine the K labeled time-series windows in chronological order to form a structured time-series input sequence.

[0017] Step S222: Construct an LSTM encoder, specifically by constructing an LSTM encoder with three stacked LSTM units with residual connections to obtain the basic temporal feature vector; the LSTM encoder includes an encoder input layer, an encoder hidden layer, and an encoder output layer;

[0018] Step S223: Construct a compliance status decoder, specifically by constructing a compliance status decoder structure that includes an LSTM temporal modeling layer, a feature purification layer, and a compliance reconstruction output layer connected in series, and inputting the basic temporal feature vector into the compliance status decoder to obtain the compliance temporal status features and the corresponding compliance reconstruction results;

[0019] The feature purification layer specifically employs a feature purification structure consisting of a fully connected layer, a layer normalization layer, and a ReLU activation function to map compliant time series features and obtain compliant time series state features.

[0020] The compliance reconstruction output layer specifically adopts a fully connected layer structure, takes compliance time-series features as input, and outputs compliance reconstruction results through linear mapping.

[0021] Step S224: Construct a risk state decoder, specifically by constructing a risk state decoder including a shared risk feature extraction layer, an offside risk feature branch, a boundary risk feature branch, and a risk reconstruction output layer. Input the basic temporal feature vector into the risk state decoder. By combining shared risk features with branch reinforcement, offside risk temporal features and boundary risk temporal features are obtained respectively, and the corresponding offside risk reconstruction results and boundary risk reconstruction results are output.

[0022] The shared risk feature extraction layer specifically employs two stacked LSTM units to perform temporal modeling on the basic temporal feature vector. After the output of each LSTM unit is processed, it undergoes layer normalization and Dropout processing in sequence to enhance the ability to model the potential risk evolution patterns in the training state and obtain shared risk features.

[0023] The offside risk feature branch specifically employs an offside risk feature extraction structure consisting of a fully connected layer, a spatial attention mechanism layer, a layer normalization layer, and a ReLU activation function. Shared risk features are input into the offside risk feature extraction structure to obtain offside risk temporal features.

[0024] The boundary risk feature branch specifically adopts a boundary risk feature extraction structure consisting of a fully connected layer, a spatial attention mechanism layer, a layer normalization layer, and a ReLU activation function. The shared risk features are input into the boundary risk feature extraction structure to obtain the boundary risk temporal features.

[0025] The risk reconstruction output layer specifically consists of two parallel fully connected reconstruction output layers, which take the off-limit risk temporal features and the out-of-bounds risk temporal features as inputs, and output off-limit risk reconstruction results and out-of-bounds risk reconstruction results that are consistent with the original risk temporal window dimensions.

[0026] Step S23: Feature fusion of serial dual risks, specifically including the following steps:

[0027] Step S231: The first stage of feature fusion guided by offside risk involves concatenating the morphological features of the training process with the temporal features of offside risk to obtain the offside fusion input vector. Subsequently, the features of the offside fusion input vector are weighted and fused through an attention mechanism to obtain the offside fusion features.

[0028] Step S232: The second stage of feature fusion guided by the risk of exceeding the boundary is to concatenate the morphological features of the training process, the feature fusion of the excess and the temporal features of the risk of exceeding the boundary to obtain the feature fusion input vector. Then, the features in the feature fusion input vector of exceeding the boundary are weighted and fused through the attention mechanism to obtain the feature fusion of the excess.

[0029] Step S24: Output the dual independent probability results. Specifically, the offside probability and the outbound probability are predicted by using a fully connected layer and a Sigmoid activation function respectively, to obtain the tank offside probability prediction result and the tank outbound probability prediction result.

[0030] Furthermore, in step S3, the performance optimization of the monitoring model specifically includes the following steps:

[0031] Step S31: Monitor model training, specifically by using historical training status monitoring data as training data for the monitoring model, designing a dual-risk temporal loss function as the loss function, and training the model to obtain the trained vehicle training status monitoring model.

[0032] The dual-risk time-series loss function is specifically a weighted combination of two parts: a time-series reconstruction loss term and a dual-probability calibration loss term.

[0033] The temporal reconstruction loss term consists of a compliance reconstruction loss term, an off-limits risk reconstruction loss term, an out-of-bounds risk reconstruction loss term, and an unlabeled temporal adaptive loss term. The average of the biprobability calibration loss values ​​of all samples is taken as the final biprobability calibration loss term.

[0034] The dual-probability calibration loss term specifically involves calculating the cross-entropy loss for the out-of-bounds probability and the cross-entropy loss for the out-of-bounds probability for each sample, and summing the two cross-entropy losses as the dual-probability calibration loss value for that sample.

[0035] Step S32: Monitor model hyperparameter optimization, specifically, obtaining the optimal hyperparameter combination of the monitoring model through an improved optimization algorithm; including the following steps:

[0036] Step S321: Population initialization, specifically, encoding the hyperparameters of the vehicle training state monitoring model into individual position vectors, and generating them using circular chaotic mapping. The position vectors of each individual are used to complete the population initialization, resulting in the initialized population.

[0037] The formula used is as follows:

[0038] ;

[0039] In the formula, Indicates the first The individual initial position vector, Indicates the first The individual initial position vector, and Indicates control parameters;

[0040] Step S322: Confirm the global optimal position of an individual. Specifically, calculate the fitness value of an individual in the population, use the performance of the vehicle training status monitoring model established based on the individual position as the individual fitness value, sort the individuals from best to worst fitness value, and take the position vector of the individual with the best fitness value as the global optimal position of the individual.

[0041] Step S323: Update position based on adaptive T-distribution perturbation, specifically by updating the individual position using an adaptive T-distribution; the formula used is as follows:

[0042] ;

[0043] In the formula, This represents the time-varying degrees of freedom of the T-distribution in the t-th iteration. Indicates the first The perturbation position of the i-th individual in the next iteration. No. The position of the i-th individual in the next iteration. Indicates the base scaling factor. This indicates that the degree of freedom is T-distributed random variable, This represents the correction factor. Indicates the maximum number of iterations;

[0044] Step S324: Update the final position based on the sine and cosine strategy. Specifically, based on the perturbation position, update the final position of the current iteration individual using a sine and cosine strategy with adaptive weights; the formula used is as follows:

[0045] ;

[0046] In the formula, This represents the adaptive weight value in the t-th iteration. and express Randomness parameters within a range This represents the global optimal position of an individual. Indicates the first The position of the i-th individual in the next iteration;

[0047] Step S325: Iterative search terminates. Specifically, for all updated individuals, the individual fitness value is recalculated and compared with the fitness values ​​of all individuals. The global optimal position of the individual is updated. When the fitness value corresponding to the global optimal position of the individual is higher than the fitness threshold and the number of search iterations reaches the maximum number of search iterations, the search is terminated and the global optimal position of the individual is obtained. The global optimal position specifically refers to the optimal hyperparameter combination of the vehicle training status monitoring model.

[0048] Step S33: Optimize and adjust the hyperparameters of the monitoring model. Specifically, based on the optimal combination of hyperparameters of the vehicle training status monitoring model, adjust the hyperparameters of the vehicle training status monitoring model to obtain the optimized vehicle training status monitoring model.

[0049] Furthermore, in step S4, the intelligent monitoring of the vehicle training status specifically involves inputting real-time training status monitoring data into the optimized vehicle training status monitoring model to obtain real-time vehicle overstepping probability prediction results and real-time vehicle boundary crossing probability prediction results. Based on the prediction results, the corresponding vehicle response strategy is triggered to achieve real-time monitoring of vehicle training.

[0050] The technical solution adopted by the present invention is as follows: The present invention provides an intelligent vehicle training status monitoring system, including a multi-source data acquisition module, a vehicle training status monitoring model construction module, a monitoring model performance optimization module, and an intelligent vehicle training status monitoring module;

[0051] The multi-source data acquisition module obtains optimized data for vehicle training status monitoring through data acquisition and data optimization operations, and sends the data to the monitoring model performance optimization module and the intelligent monitoring module for vehicle training status.

[0052] The module for constructing the vehicle training status monitoring model first extracts the morphological features of the training process as static constraint factors for risk probability prediction. Then, it uses an improved structured autoencoder combined with type-time window annotation to extract the offside risk time-series features and the boundary risk time-series features. Subsequently, it performs serial dual-risk feature fusion, progressively fusing the morphological features of the training process, the offside risk time-series features, and the boundary risk time-series features to generate offside fusion features and boundary fusion features. Finally, it outputs the vehicle offside probability prediction result and the vehicle boundary probability prediction result through dual independent probability prediction, thus completing the construction of the vehicle training status monitoring model and obtaining the vehicle training status monitoring model. The data is then sent to the monitoring model performance optimization module.

[0053] The monitoring model performance optimization module receives data from the multi-source data acquisition module and the vehicle training status monitoring model construction module. Specifically, it designs a dual-risk temporal loss function with a temporal reconstruction loss term and a dual-probability calibration loss term as the objective function for training the monitoring model. It then improves the optimization algorithm through circular chaotic mapping, adaptive T-distribution perturbation position, and sine and cosine strategies with adaptive weights to obtain the optimal hyperparameter combination of the vehicle training status monitoring model. Finally, it adjusts the hyperparameters of the monitoring model according to the optimal hyperparameter combination to obtain the optimized vehicle training status monitoring model and sends the data to the intelligent monitoring module for vehicle training status.

[0054] The intelligent monitoring module for vehicle training status receives data from the multi-source data acquisition module and the monitoring model performance optimization module. Specifically, it inputs real-time data into the optimized vehicle training status monitoring model to obtain real-time vehicle offside and boundary probability prediction results, and adopts corresponding vehicle response strategies based on the prediction results.

[0055] The beneficial effects achieved by the present invention using the above solution are as follows:

[0056] (1) In view of the technical problem that traditional tank training status monitoring systems rely solely on a single temporal feature and fail to fully consider the differences in the impact of overall morphological changes during training on the risk of off-boundary and over-limit, resulting in insufficient accuracy in the prediction of off-boundary and over-limit, this solution innovatively introduces the morphological features of the training process as a static constraint factor and fuses them with the temporal features of off-boundary and over-limit risks. This enhances the ability to perceive morphological changes during tank training, enabling it to maintain efficient and reliable prediction capabilities in complex training scenarios. It effectively solves the problems of inaccurate prediction and poor scenario adaptability of traditional models during training, improves the accuracy and stability of off-boundary and over-limit risk prediction, and realizes intelligent monitoring during tank training.

[0057] (2) In response to the technical problem that the existing vehicle training status monitoring model uses a single temporal feature extraction method and does not distinguish between different risk types of temporal patterns, resulting in inaccurate monitoring model output results, this solution innovatively adopts an improved structured autoencoder combined with a multi-type temporal window annotation method. The improved structured autoencoder extracts general basic temporal features through a three-layer stacked LSTM encoder with residual connections, and then uses a compliance decoder and a risk dual-branch decoder to learn the temporal evolution law under different states in a differentiated manner, which significantly improves the discriminativeness of temporal features, ensures the accurate separation of compliance benchmark features and various risk features, enhances the model's learning ability for scarce dual-risk scenarios, improves the model's recognition accuracy and stability for offside and boundary risks, provides highly discriminative temporal support for the independent prediction of offside and boundary dual probabilities, and realizes the accurate identification and quantitative prediction of multiple types of risks under the vehicle training status.

[0058] (3) To address the technical problems of existing models applicable to vehicle training status monitoring, which suffer from insufficient accuracy and stability in predicting out-of-bounds probabilities due to the single design of traditional loss functions, difficulty in simultaneously considering temporal feature learning and probability prediction accuracy, and the tendency of conventional optimization algorithms to get trapped in local optima and have low hyperparameter search efficiency, this solution innovatively designs a dual-risk temporal loss function. This function weights and fuses the temporal reconstruction loss term with the dual-probability calibration loss term, constraining the model to effectively model the temporal evolution of vehicle training status and ensuring that the extracted temporal features have sufficient expressiveness and distinguishability. This approach ensures that the output probability closely matches the actual risk occurrence. An improved optimization algorithm incorporating circular chaotic mapping, adaptive T-distribution perturbation positions, and adaptive weights using a sine and cosine strategy effectively avoids getting trapped in local optima during optimization and accelerates the convergence speed of the optimal hyperparameter combination. Through these techniques, the accuracy, stability, and generalization ability of the vehicle training status monitoring model for predicting offside and boundary violations are significantly improved. The model's adaptability to different training scenarios and data distribution changes is enhanced, ultimately achieving highly reliable quantitative prediction and intelligent dynamic monitoring of offside and boundary violations risks during vehicle training. Attached Figure Description

[0059] Figure 1 A flowchart illustrating an intelligent combat vehicle training status monitoring method provided by the present invention;

[0060] Figure 2 A schematic diagram of a module for an intelligent combat vehicle training status monitoring system provided by the present invention;

[0061] Figure 3 This is a flowchart illustrating step S2;

[0062] Figure 4 This is a flowchart illustrating step S3;

[0063] Figure 5 This is a flowchart illustrating step S22;

[0064] Figure 6 This is a flowchart illustrating step S32;

[0065] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0066] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments 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 creative effort are within the scope of protection of the present invention.

[0067] In the description of this invention, it should be understood that the terms upper, lower, front, back, left, right, top, bottom, inner, and outer, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0068] Example 1, see Figure 1 The technical solution adopted by the present invention is as follows: The present invention provides an intelligent method for monitoring the training status of combat vehicles, the method comprising the following steps:

[0069] Step S1: Multi-source data acquisition, specifically, obtaining optimized data for vehicle training status monitoring through data acquisition and data optimization operations;

[0070] Step S2: Construct a vehicle training status monitoring model. Specifically, firstly, extract the morphological features of the training process as static constraint factors for risk probability prediction. Then, use an improved structured autoencoder combined with type temporal window annotation to extract offside risk temporal features and boundary risk temporal features. Subsequently, perform serial dual-risk feature fusion, progressively fusing the morphological features of the training process, the offside risk temporal features, and the boundary risk temporal features to generate offside fusion features and boundary fusion features. Finally, through dual independent probability prediction, output the vehicle offside probability prediction result and the vehicle boundary probability prediction result, thus completing the construction of the vehicle training status monitoring model and obtaining the vehicle training status monitoring model.

[0071] Step S3: Optimize the performance of the monitoring model. Specifically, a dual-risk temporal loss function with a temporal reconstruction loss term and a dual-probability calibration loss term is designed as the objective function for training the monitoring model. The optimization algorithm is improved by circular chaotic mapping, adaptive T-distribution perturbation position, and sine and cosine strategies with adaptive weights to obtain the optimal hyperparameter combination of the vehicle training status monitoring model. Finally, the hyperparameters of the monitoring model are adjusted according to the optimal hyperparameter combination to obtain the optimized vehicle training status monitoring model.

[0072] Step S4: Intelligent monitoring of vehicle training status. Specifically, real-time data is input into the optimized vehicle training status monitoring model to obtain real-time prediction results of vehicle overstepping and boundary crossing probabilities. Based on the prediction results, corresponding vehicle response strategies are adopted.

[0073] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the multi-source data acquisition specifically involves obtaining the original data of the vehicle training status monitoring through data acquisition operations, and performing data optimization operations on the original data of the vehicle training status monitoring to obtain optimized data of the vehicle training status monitoring. The original data of the vehicle training status monitoring includes historical training status monitoring data and real-time training status monitoring data.

[0074] The historical training status monitoring data and real-time training status monitoring data include vehicle operation data, maneuver control data, vehicle training data, training environment data, and personnel operation data;

[0075] The historical training status monitoring data also includes historical tank overrun prediction results and historical tank boundary crossing prediction results.

[0076] The data optimization operations include data cleaning, data normalization, data encoding processing, and data feature selection;

[0077] The data acquisition operation specifically involves issuing a data acquisition start command to the participating combat vehicles at the start of training. The participating combat vehicles sample the data from sensors and equipment through the on-board data acquisition unit and encapsulate it into data packets, which are then transmitted back to the monitoring system via a wireless communication link to obtain the raw data of the combat vehicle training status monitoring.

[0078] The vehicle offside prediction result and the vehicle boundary crossing prediction result are specifically the offside probability and boundary crossing probability output by the monitoring model, with values ​​ranging from... ;

[0079] The vehicle's operating data includes engine speed, throttle opening, vehicle speed, braking status, gear information, and mileage.

[0080] The motor control data includes steering wheel angle, steering angular velocity, acceleration / deceleration rate, and hill pass status;

[0081] The vehicle training data includes training subject number, vehicle positioning data, training ground space data, training mission route data, designated location data, and training vehicle reference data;

[0082] The training environment data includes terrain type, road conditions, slope information, and weather parameters;

[0083] The personnel operation data includes driver control commands, operation response time, and operation frequency;

[0084] The data cleaning is used to remove invalid and abnormal data, specifically by filling in missing values ​​and removing outliers from the original data.

[0085] The missing value imputation specifically involves filling in missing values ​​using the mean imputation method; the outlier removal specifically involves detecting extreme values ​​in the original data using the interquartile range method and removing data records that meet the outlier conditions.

[0086] The data normalization is used to eliminate differences in different data units and numerical scales. Specifically, the Z-Score standardization method is used to standardize all continuous variables so that each continuous variable meets a uniform numerical scale range.

[0087] The data encoding process is used to convert non-numerical data into a numerical format that can be processed by the model. Specifically, a label encoding method is used to map the category field in the original data to the corresponding integer value according to the category value to obtain the encoded category feature.

[0088] The data feature selection is used to remove redundant and weakly correlated features and reduce the impact of noise. Specifically, correlation analysis and variance thresholding methods are used to measure the correlation between candidate features and target variables and the information content of the features themselves. Candidate features are sorted and filtered, and a subset of features whose correlation and importance both meet the preset threshold conditions are retained to obtain the vehicle training status monitoring set.

[0089] Example 3, see Figure 1 , Figure 2 , Figure 3 and Figure 5 This embodiment is based on the above embodiment. In step S2, the construction of the vehicle training status monitoring model specifically includes the following steps:

[0090] Step S21: Extract training process morphological features to extract the overall behavioral features and process morphology reflecting the current tank training process from non-time series data, as an influencing factor affecting the output of offside probability and boundary probability; specifically, within a preset statistical period, the training status monitoring data is aggregated to form a training process morphological input vector, and the training process morphological input vector is input into a morphological feature extraction network based on a multilayer perceptron structure to obtain the training process morphological features.

[0091] Step S22: Extract the temporal evolution features of the training state to characterize the dynamic evolution process of the vehicle's training state over time, providing temporal features that reflect the deviation trend and risk evolution for the prediction of offside probability and boundary violation probability. Specifically, the extraction of the temporal evolution features of the training state is based on an improved structured autoencoder, including the following steps:

[0092] Step S221: Construct a time-series input sequence to provide a structured and continuous time-series data foundation for the subsequent extraction of time-series evolution features of training states, ensuring accurate capture of the dynamic change patterns of the training state of the combat vehicle. Specifically, according to the vehicle identification and timestamp, sort the various types of data in the training state monitoring data by time. Then, use the sliding time window method to segment the sorted data to obtain K time-series windows, and label the data in each time-series window. Combine the K labeled time-series windows in chronological order to form a structured time-series input sequence.

[0093] The annotation types of the time series window specifically include compliance windows. Window before offside Before crossing the boundary window Dual risk windows and unlabeled window ;

[0094] The compliance window specifically refers to a time window in which neither overstepping nor overstepping occurs;

[0095] The "before the offside window" specifically refers to a window where no offside occurred within the window, but an offside occurred within the next 3 seconds.

[0096] The window before the boundary crossing specifically refers to a window where no boundary crossing occurred within the window, but a boundary crossing occurred within the following 3 seconds;

[0097] The aforementioned double risk window specifically refers to a window in which no double risk occurs, but an offside and an out-of-bounds violation occur simultaneously within the following 3 seconds.

[0098] The unlabeled window specifically refers to a situation where the actual occurrence was not labeled.

[0099] The timing window is specifically a sliding time window with a timing window length of wt and a step size of 1;

[0100] Step S222: Construct an LSTM encoder to extract basic temporal features that fuse long-term dependencies and multi-feature correlations from each temporal window of the temporal input sequence. Specifically, construct an LSTM encoder with three stacked LSTM units with residual connections to obtain the basic temporal feature vector. The LSTM encoder includes an encoder input layer, an encoder hidden layer, and an encoder output layer.

[0101] Specifically, the encoder input layer receives the kth timing window in the timing input sequence as input data;

[0102] The encoder hidden layer comprises three LSTM layers. Features are transferred between adjacent LSTM layers via residual connections. After each layer's output, it undergoes layer normalization and... Processing is performed to enhance feature stability and model generalization ability;

[0103] The encoder output layer is specifically based on the fully connected layer outputting the third LSTM layer. By performing mapping, the basic time series feature vector is obtained. ;

[0104] The formula used is as follows:

[0105] ;

[0106] ;

[0107] ;

[0108] ;

[0109] In the formula, This represents the k-th timing window in the timing input sequence. , and These represent the LSTM cells in layers 1 to 3, , and These represent the output features of the LSTM units in layers 1 to 3, respectively. Represents the basic time series feature vector. This represents the layer normalization function, used to normalize the output of each LSTM layer. Represents the random deactivation function. and This represents the fully connected weight matrix and bias term parameters of the encoder output layer. express Activation function;

[0110] Step S223: Construct a compliance state decoder to learn the temporal evolution pattern of compliant training states during vehicle training when neither offside nor overside occurs. Reconstruct and refine the basic temporal features to output compliant temporal state features that characterize the compliant training states. Specifically, construct a compliance state decoder structure consisting of an LSTM temporal modeling layer, a feature refinement layer, and a compliance reconstruction output layer connected in series. Input the basic temporal feature vector into the compliance state decoder to obtain the compliant temporal state features and the corresponding compliance reconstruction results.

[0111] The LSTM temporal modeling layer is used to restore the continuous characteristics of the training state changes over time under compliant training conditions, and to characterize the stable and repeatable temporal evolution pattern during compliant training. Specifically, it uses two stacked LSTM units to perform temporal modeling processing on the basic temporal feature vector. After the output of each LSTM unit is processed, it undergoes layer normalization and Dropout processing in sequence to enhance the stability of temporal feature expression and suppress overfitting.

[0112] The feature purification layer is used to perform nonlinear mapping and feature enhancement on the compliant time series features output by the LSTM time series modeling layer, highlighting the core time series features under the compliant training state. Specifically, it adopts a feature purification structure composed of a fully connected layer, a layer normalization layer and a ReLU activation function to further map and purify the compliant time series features to obtain compliant time series state features.

[0113] The compliance reconstruction output layer is used to reconstruct the compliance time-series window based on compliance time-series features. By calculating the reconstruction error between the compliance reconstruction result and the original compliance time-series window, a reconstruction constraint is applied to the compliance state decoder to ensure that the extracted compliance time-series features can accurately depict the temporal evolution of the vehicle under compliance training conditions. Specifically, a fully connected layer structure is adopted, using compliance time-series features as input, and outputting a compliance reconstruction result that is consistent with the original compliance time-series window in dimension through linear mapping. The formula used is as follows:

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] In the formula, and These represent the intermediate compliant time-series features after processing by the first and second compliant LSTM units, respectively. and This represents the first and second layer LSTM units in the compliance status decoder. This represents the compliance time-series feature vector, used to characterize the compliance training state time-series benchmark pattern of the vehicle under conditions where no offside or boundary violation occurs. and This represents the weight matrix and bias parameters of the feature extraction layer. and This represents the fully connected weight matrix and bias parameters of the output layer in the compliant reconstruction. This represents the compliance restructuring result, and its dimension is consistent with the original k-th compliance time series window.

[0119] Step S224: Construct a risk state decoder to differentiate the temporal evolution patterns of the vehicle before offside, before boundary crossing, and before dual risks. Reconstruct risk patterns and extract specific features from the basic temporal features. This enables differentiated characterization of the temporal evolution patterns of offside and boundary crossing risks within the same basic temporal feature space, providing stable and separable dynamic feature support for the independent prediction of the two risk probabilities. Specifically, a risk state decoder is constructed, including a shared risk feature extraction layer, an offside risk feature branch, a boundary crossing risk feature branch, and a risk reconstruction output layer. The basic temporal feature vector is input into the risk state decoder. By combining shared risk features with branch reinforcement, the temporal features of offside and boundary crossing risks are obtained respectively, and the corresponding offside risk reconstruction results and boundary crossing risk reconstruction results are output.

[0120] The shared risk feature extraction layer is used to uniformly map the basic temporal features and extract potential risk evolution information in the training state. Specifically, it uses two stacked LSTM units to perform temporal modeling processing on the basic temporal feature vectors. After the output of each LSTM unit is processed, it undergoes layer normalization and Dropout processing in sequence to enhance the ability to model potential risk evolution patterns in the training state and obtain shared risk features.

[0121] The offside risk feature branch is used to extract risk time-series features that are highly correlated with the occurrence of offside from the shared risk features, highlighting the risk evolution features when the vehicle deviates from the tactical position constraints during training; specifically, it adopts an offside risk feature extraction structure composed of a fully connected layer, a spatial attention mechanism layer, a layer normalization layer, and a ReLU activation function, and inputs the shared risk features into the offside risk feature extraction structure to obtain the offside risk time-series features;

[0122] The boundary crossing risk feature branch extracts risk time-series features that are highly correlated with the occurrence of boundary crossing from the shared risk features, highlighting the risk evolution features when the vehicle approaches or deviates from the geographical boundary of the training area during the training process; specifically, it adopts a boundary crossing risk feature extraction structure composed of a fully connected layer, a spatial attention mechanism layer, a layer normalization layer, and a ReLU activation function, and inputs the shared risk features into the boundary crossing risk feature extraction structure to obtain the boundary crossing risk time-series features;

[0123] The risk reconstruction output layer is used to reconstruct the offside risk window and the outside risk window respectively, to ensure the relevance and effectiveness of risk feature extraction. Specifically, it constructs two parallel fully connected reconstruction output layers, which take the offside risk temporal features and the outside risk temporal features as inputs, and output the offside risk reconstruction result and the outside risk reconstruction result with the same dimensions as the original risk temporal window. The formula used is as follows:

[0124] ;

[0125] ;

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] ;

[0131] ;

[0132] In the formula, This represents the intermediate temporal features output by the first LSTM unit of the shared risk feature extraction layer. This represents the shared risk feature, which is the intermediate temporal feature output by the second LSTM unit of the shared risk feature extraction layer. and These represent the first and second LSTM units in the shared risk feature extraction layer, respectively. and This represents the weight matrix and bias parameters of the fully connected layer in the offside risk feature branch. and This represents the weight matrix and bias parameters of the fully connected layer in the out-of-bounds risk feature branch. This represents the spatial attention mechanism function. and This represents the weight matrix and bias parameters of the output layer reconstructed to indicate offside risk. and This represents the weight matrix and bias parameters of the output layer reconstructed to indicate the risk of exceeding the boundary. This indicates the results of the offside risk reconstruction. This indicates the results of the boundary risk reconstruction. This indicates the temporal characteristics of offside risk, used to characterize the risk evolution trend before offside occurs. It represents the temporal characteristics of boundary crossing risk, used to characterize the risk evolution trend before boundary crossing occurs;

[0133] Step S23: Serial dual-risk feature fusion, used to fuse the morphological features of the training process with the temporal features of offside risk and the temporal features of boundary risk in stages and progressively according to the logical order of risk evolution. This allows boundary risk prediction to be modeled based on a full consideration of the overall morphology of the training process and the impact of offside risk, enhancing the independence and accuracy of offside probability and boundary probability. Specifically, it includes the following steps:

[0134] Step S231: The first-stage fusion feature guided by offside risk is used to fuse the morphological features of the training process with the temporal features of offside risk through attention, reflecting the degree of offside risk under the current training morphological constraints. Specifically, the morphological features of the training process and the temporal features of offside risk are concatenated to obtain the offside fusion input vector. Subsequently, the features of the offside fusion input vector are weighted and fused through an attention mechanism to obtain the offside fusion feature; the formula used is as follows:

[0135] ;

[0136] In the formula, This represents the offside fusion feature, used to reflect the comprehensive risk characteristics of offside occurring in the current training state. This represents the attention mechanism function. This indicates a feature concatenation operation. Represents the morphological features of the training process;

[0137] Step S232: The second-stage fusion feature guided by the risk of exceeding the boundary is used to further fuse the morphological features of the training process, the fusion features of the exceeding the boundary, and the temporal features of the exceeding the boundary risk, based on the impact of the exceeding the boundary risk, to reflect the comprehensive evolution trend of the exceeding the boundary risk. Specifically, the morphological features of the training process, the fusion features of the exceeding the boundary, and the temporal features of the exceeding the boundary are concatenated to obtain the exceeding the boundary fusion input vector. Subsequently, the features in the exceeding the boundary fusion input vector are weighted and fused through an attention mechanism to obtain the exceeding the boundary fusion features. The formula used is as follows:

[0138] ;

[0139] In the formula, This indicates the boundary crossing fusion characteristic, used to reflect the comprehensive risk characteristics of boundary crossing under the influence of training mode and offside risk;

[0140] Step S24: Output dual independent probability results, used to independently predict the offside probability and the outbound probability, ensuring that both reflect the potential risks the vehicle may face in the current training state. Specifically, the offside probability and the outbound probability are predicted using a fully connected layer and a sigmoid activation function, respectively, to obtain the vehicle offside probability prediction result and the vehicle outbound probability prediction result; the formulas used are as follows:

[0141] ;

[0142] ;

[0143] In the formula, This indicates the predicted probability of the tank being offside. This indicates the predicted probability of the tank crossing the boundary. and These represent the weight matrix and bias term for predicting the offside probability, respectively. and These represent the weight matrix and bias term of the out-of-bounds probability prediction output, respectively.

[0144] By performing the above operations, this solution addresses the technical problem in traditional armored vehicle training status monitoring systems that rely solely on a single temporal feature and fail to fully consider the varying impacts of overall morphological changes during training on off-limits and boundary risks, leading to insufficient accuracy in off-limits and boundary risk predictions. This solution innovatively introduces morphological features from the training process as a static constraint factor, fusing them with off-limits and boundary risk temporal features. This enhances the system's ability to perceive morphological changes during armored vehicle training, enabling it to maintain efficient and reliable prediction capabilities in complex training scenarios. It effectively solves the problems of inaccurate predictions and poor scenario adaptability in traditional models during training, improving the accuracy and stability of off-limits and boundary risk predictions, and achieving intelligent monitoring during armored vehicle training. Furthermore, it addresses the issue of existing models for armored vehicle training status monitoring using a single temporal feature… The previous method of feature extraction, which failed to differentiate between different risk types and time series patterns, led to inaccurate output results from the monitoring model. This solution innovatively adopts an improved structured autoencoder combined with a multi-type time series window annotation method. The improved structured autoencoder extracts general basic time series features through a three-layer stacked LSTM encoder with residual connections. Then, it uses a compliance decoder and a risk dual-branch decoder to learn the time series evolution rules under different states in a differentiated manner, which significantly improves the discriminativeness of time series features, ensures the accurate separation of compliance baseline features and various risk features, enhances the model's learning ability for scarce dual-risk scenarios, and improves the model's recognition accuracy and stability for offside and out-of-bounds risks. It provides highly discriminative time series support for independent prediction of offside and out-of-bounds probabilities, and realizes accurate identification and quantitative prediction of multiple types of risks in the vehicle training state.

[0145] Example 4, see Figure 1 , Figure 2 , Figure 4 and Figure 6 This embodiment is based on the above embodiment. In step S3, the monitoring model performance optimization is used to optimize the parameters of the vehicle training status monitoring model through end-to-end training, improve the accuracy, stability and generalization ability of the model in predicting the probability of vehicle offside and boundary crossing, adapt to the data characteristics of different training scenarios of the vehicle, and ensure that the output probability is highly matched with the actual risk occurrence. Specifically, it includes the following steps:

[0146] Step S31: Monitor model training, which is used to supervise model parameter updates through a multi-objective joint loss function, so that the model can accurately learn the risk evolution mode of compliant training and achieve performance optimization of the entire link from feature extraction to probability prediction. Specifically, historical training status monitoring data is used as training data for the monitoring model, and a dual-risk time series loss function is designed as the loss function to train the model and obtain the trained vehicle training status monitoring model.

[0147] The model training specifically involves jointly optimizing the evaluation model parameters using the backpropagation algorithm, continuously monitoring and optimizing the loss function value until it converges, at which point the model training stops, resulting in the trained vehicle training status monitoring model.

[0148] The dual-risk temporal loss function is used to ensure the accuracy of probability prediction while taking into account the effectiveness of temporal feature extraction. Specifically, it is a weighted combination of two parts: a temporal reconstruction loss term and a dual-probability calibration loss term.

[0149] The temporal reconstruction loss term consists of a compliance reconstruction loss term, an off-limits risk reconstruction loss term, an out-of-bounds risk reconstruction loss term, and an unlabeled temporal adaptive loss term. The average of the biprobability calibration loss values ​​of all samples is taken as the final biprobability calibration loss term.

[0150] The dual-probability calibration loss term specifically involves calculating the cross-entropy loss for the out-of-bounds probability and the cross-entropy loss for the out-of-bounds probability for each sample, and summing the two cross-entropy losses as the dual-probability calibration loss value for that sample.

[0151] The compliance refactoring loss term is specifically calculated as the average refactoring error between the original data of all compliance time series windows and the corresponding compliance refactoring results;

[0152] The offside risk reconstruction loss term is specifically calculated as the average reconstruction error between the original data of all original pre-offside time-series windows and dual-risk windows and the offside risk reconstruction result output by the offside branch of the risk state decoder.

[0153] The out-of-bounds risk reconstruction loss term is specifically calculated as the average reconstruction error between the original data of all original out-of-bounds time-series windows and dual-risk windows and the out-of-bounds risk reconstruction result output by the out-of-bounds branch of the risk state decoder.

[0154] The unlabeled temporal adaptive loss term is specifically calculated by taking the minimum average reconstruction error between all unlabeled temporal windows and the three types of reconstruction results; the formula used is as follows:

[0155] ;

[0156] ;

[0157] ;

[0158] ;

[0159] ;

[0160] ;

[0161] In the formula, This represents the value of the two-risk time-series loss function. This represents the temporal reconstruction loss value. This represents the biprobability calibration loss value. and These represent the weight coefficients of the temporal reconstruction loss term and the biprobability calibration loss term, respectively. This indicates the value of the compliance restructuring loss. This indicates the reconstructed loss value for offside risk. This indicates the reconstructed loss value for the risk of exceeding the boundary. This represents the unlabeled time-series adaptive loss value. , , , and These represent the total number of compliance windows, pre-out-of-bounds windows, pre-out-of-bounds windows, dual-risk windows, and unmarked windows, respectively. This represents the square of the L2 norm, used to calculate the difference between two vectors;

[0162] Step S32: Monitor model hyperparameter optimization, used to further improve the model's predictive performance by optimizing the hyperparameters of the vehicle training state monitoring model, ensuring that the model has good generalization ability and stability in different training scenarios. Specifically, it involves obtaining the optimal hyperparameter combination of the monitoring model through an improved optimization algorithm; including the following steps:

[0163] Step S321: Population initialization, used to initialize the hyperparameter search space of the monitoring model. Specifically, the hyperparameters of the vehicle training state monitoring model are encoded into individual position vectors, and a circular chaotic mapping is used to generate... The position vectors of each individual are used to complete the population initialization, resulting in the initialized population.

[0164] The hyperparameters of the vehicle training status monitoring model include learning rate, temporal window length, LSTM hidden state dimension, loss function weight coefficient, hidden layer dimension, and Dropout ratio.

[0165] The formula used is as follows:

[0166] ;

[0167] In the formula, Indicates the first The individual initial position vector, Indicates the first The individual initial position vector, and This represents a control parameter used to control the generation of an individual's initial position; its value ranges from [value range missing]. ;

[0168] Step S322: Confirm the global optimal position of an individual. This step is used to quantify the optimization effect of the hyperparameter combination corresponding to the current position of an individual in the population, clarify the global optimal benchmark for hyperparameter search, and provide a core reference for subsequent targeted and precise search. Specifically, it involves calculating the fitness value of an individual in the population, using the performance of the vehicle training status monitoring model established based on the individual's position as the individual's fitness value, sorting individuals from best to worst fitness value, and taking the position vector of the individual with the best fitness value as the individual's global optimal position.

[0169] Step S323: Position update based on adaptive T-distribution perturbation, used to enhance the algorithm's ability to escape local optima and balance the dynamic needs of global exploration and local exploitation. Specifically, the individual positions are updated using an adaptive T-distribution perturbation; the formula used is as follows:

[0170] ;

[0171] ;

[0172] In the formula, This represents the time-varying degrees of freedom of the T-distribution in the t-th iteration. This represents the minimum number of degrees of freedom. This represents the maximum number of degrees of freedom. Indicates the current iteration number. This represents the maximum number of iterations, and p represents the exponent of the rate of change of degrees of freedom, with a range of values. Used for control As the number of iterations increases, Indicates the first The perturbation position of the i-th individual in the next iteration. No. The position of the i-th individual in the next iteration. This represents the base scaling factor, and its value range is... , This indicates that the degree of freedom is T-distributed random variable, This represents the correction factor, and its value range is... ;

[0173] Step S324: Update the final position based on the sine and cosine strategy. This is used to dynamically adjust the search direction and step size, accelerate algorithm convergence, and avoid later oscillations. Specifically, based on the perturbation position, the final position of the current iteration individual is updated by introducing an adaptive weighted sine and cosine strategy. The formula used is as follows:

[0174] ;

[0175] ;

[0176] In the formula, This represents the adaptive weight value in the t-th iteration. This represents the minimum adaptive weight. This represents the maximum value of the adaptive weight. and express Randomness parameters within a range This represents the global optimal position of an individual. Indicates the first The position of the i-th individual in the next iteration;

[0177] Step S325: Iterative search terminates. Specifically, for all updated individuals, the individual fitness value is recalculated and compared with the fitness values ​​of all individuals. The global optimal position of the individual is updated. When the fitness value corresponding to the global optimal position of the individual is higher than the fitness threshold and the number of search iterations reaches the maximum number of search iterations, the search is terminated and the global optimal position of the individual is obtained. The global optimal position specifically refers to the optimal hyperparameter combination of the vehicle training status monitoring model.

[0178] Step S33: Optimize and adjust the hyperparameters of the monitoring model. Specifically, based on the optimal combination of hyperparameters of the vehicle training status monitoring model, adjust the hyperparameters of the vehicle training status monitoring model to obtain the optimized vehicle training status monitoring model.

[0179] By performing the above operations, this solution addresses the technical problems of existing models applicable to vehicle training status monitoring, such as the limitations of traditional loss function design in simultaneously considering temporal feature learning and probability prediction accuracy, the susceptibility of conventional optimization algorithms to local optima, and low hyperparameter search efficiency, leading to insufficient accuracy and stability in the model's out-of-bounds probability predictions. This solution innovatively designs a dual-risk temporal loss function, which weights and fuses the temporal reconstruction loss term with the dual-probability calibration loss term. This constrains the model's effective modeling of the temporal evolution of vehicle training status, ensuring that the extracted temporal features are sufficiently expressive and distinguishable. The model employs a divisibility mechanism to ensure that the output probability closely matches the actual risk occurrence. It also introduces a circular chaotic mapping, an improved optimization algorithm using an adaptive T-distribution perturbation position and adaptive weights with sine and cosine strategies, effectively avoiding the problem of getting trapped in local optima during optimization and accelerating the convergence speed of the optimal hyperparameter combination. Through these techniques, the accuracy, stability, and generalization ability of the vehicle training status monitoring model in predicting offside and boundary violations are significantly improved. The model's adaptability to different training scenarios and data distribution changes is enhanced, ultimately achieving highly reliable quantitative prediction and intelligent dynamic monitoring of offside and boundary violations risks during vehicle training.

[0180] Example 5, see Figure 1 and Figure 2This embodiment is based on the above embodiment. In step S4, the intelligent monitoring of the vehicle training status specifically involves inputting real-time training status monitoring data into the optimized vehicle training status monitoring model to obtain real-time vehicle overstep probability prediction results and real-time vehicle boundary crossing probability prediction results. Based on the prediction results, the corresponding vehicle response strategy is triggered to achieve real-time monitoring of vehicle training.

[0181] The specific tank response strategy is as follows: if the real-time tank offside probability prediction result is higher than the set offside risk threshold, the tank control command is automatically adjusted to avoid offside; if the real-time tank boundary crossing probability prediction result is higher than the set boundary crossing risk threshold, the tank driving route is adjusted and an alarm is issued to remind the training operator.

[0182] Example 6, see Figure 1 and Figure 2 Based on the above embodiments, the technical solution adopted by the present invention is as follows: The present invention provides an intelligent vehicle training status monitoring system, including a multi-source data acquisition module, a vehicle training status monitoring model construction module, a monitoring model performance optimization module, and an intelligent vehicle training status monitoring module;

[0183] The multi-source data acquisition module obtains optimized data for vehicle training status monitoring through data acquisition and data optimization operations, and sends the data to the monitoring model performance optimization module and the intelligent monitoring module for vehicle training status.

[0184] The module for constructing the vehicle training status monitoring model first extracts the morphological features of the training process as static constraint factors for risk probability prediction. Then, it uses an improved structured autoencoder combined with type-time window annotation to extract the offside risk time-series features and the boundary risk time-series features. Subsequently, it performs serial dual-risk feature fusion, progressively fusing the morphological features of the training process, the offside risk time-series features, and the boundary risk time-series features to generate offside fusion features and boundary fusion features. Finally, it outputs the vehicle offside probability prediction result and the vehicle boundary probability prediction result through dual independent probability prediction, thus completing the construction of the vehicle training status monitoring model and obtaining the vehicle training status monitoring model. The data is then sent to the monitoring model performance optimization module.

[0185] The monitoring model performance optimization module receives data from the multi-source data acquisition module and the vehicle training status monitoring model construction module. Specifically, it designs a dual-risk temporal loss function with a temporal reconstruction loss term and a dual-probability calibration loss term as the objective function for training the monitoring model. It then improves the optimization algorithm through circular chaotic mapping, adaptive T-distribution perturbation position, and sine and cosine strategies with adaptive weights to obtain the optimal hyperparameter combination of the vehicle training status monitoring model. Finally, it adjusts the hyperparameters of the monitoring model according to the optimal hyperparameter combination to obtain the optimized vehicle training status monitoring model and sends the data to the intelligent monitoring module for vehicle training status.

[0186] The intelligent monitoring module for vehicle training status receives data from the multi-source data acquisition module and the monitoring model performance optimization module. Specifically, it inputs real-time data into the optimized vehicle training status monitoring model to obtain real-time vehicle offside and boundary probability prediction results, and adopts corresponding vehicle response strategies based on the prediction results.

[0187] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "include," "contain," 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.

[0188] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0189] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for monitoring the training status of intelligent combat vehicles, characterized in that: The method includes the following steps: Step S1: Multi-source data acquisition, specifically, obtaining optimized data for vehicle training status monitoring through data acquisition and data optimization operations; Step S2: Construct a vehicle training status monitoring model. Specifically, firstly, extract the morphological features of the training process as static constraint factors for risk probability prediction. Then, use an improved structured autoencoder combined with type temporal window annotation to extract offside risk temporal features and boundary risk temporal features. Subsequently, perform serial dual-risk feature fusion, progressively fusing the morphological features of the training process, the offside risk temporal features, and the boundary risk temporal features to generate offside fusion features and boundary fusion features. Finally, through dual independent probability prediction, output the vehicle offside probability prediction result and the vehicle boundary probability prediction result, thus completing the construction of the vehicle training status monitoring model and obtaining the vehicle training status monitoring model. Step S3: Optimize the performance of the monitoring model. Specifically, a dual-risk temporal loss function with a temporal reconstruction loss term and a dual-probability calibration loss term is designed as the objective function for training the monitoring model. The optimization algorithm is improved by circular chaotic mapping, adaptive T-distribution perturbation position, and sine and cosine strategies with adaptive weights to obtain the optimal hyperparameter combination of the vehicle training status monitoring model. Finally, the hyperparameters of the monitoring model are adjusted according to the optimal hyperparameter combination to obtain the optimized vehicle training status monitoring model. Step S4: Intelligent monitoring of vehicle training status. Specifically, real-time data is input into the optimized vehicle training status monitoring model to obtain real-time prediction results of vehicle overstepping and boundary crossing probabilities. Based on the prediction results, corresponding vehicle response strategies are adopted.

2. The intelligent combat vehicle training status monitoring method according to claim 1, characterized in that: In step S2, the construction of the vehicle training status monitoring model specifically includes the following steps: Step S21: Extract morphological features of the training process. Specifically, within a preset statistical period, aggregate the training state monitoring data to form a morphological input vector of the training process, and input the morphological input vector of the training process into a morphological feature extraction network based on a multilayer perceptron structure to obtain the morphological features of the training process. Step S22: Extract the temporal evolution features of the training states; Step S23: Feature fusion of serial dual risks, specifically including the following steps: Step S231: The first stage of feature fusion guided by offside risk involves concatenating the morphological features of the training process with the temporal features of offside risk to obtain the offside fusion input vector. Subsequently, the features of the offside fusion input vector are weighted and fused through an attention mechanism to obtain the offside fusion features. Step S232: The second stage of feature fusion guided by the risk of exceeding the boundary is to concatenate the morphological features of the training process, the feature fusion of the excess and the temporal features of the risk of exceeding the boundary to obtain the feature fusion input vector. Then, the features in the feature fusion input vector of exceeding the boundary are weighted and fused through the attention mechanism to obtain the feature fusion of the excess. Step S24: Output the dual independent probability results. Specifically, the offside probability and the outbound probability are predicted by using a fully connected layer and a Sigmoid activation function respectively, to obtain the tank offside probability prediction result and the tank outbound probability prediction result.

3. The intelligent combat vehicle training status monitoring method according to claim 2, characterized in that: In step S22, the extraction of training state temporal evolution features specifically involves the extraction of training state temporal evolution features based on an improved structured autoencoder, including the following steps: Step S221: Construct a time-series input sequence. Specifically, sort the various types of data in the training status monitoring data according to the vehicle identification and timestamp. Then, use the sliding time window method to segment the sorted data to obtain K time-series windows. Label each time-series window with data. Combine the K labeled time-series windows in chronological order to form a structured time-series input sequence. Step S222: Construct an LSTM encoder, specifically by constructing an LSTM encoder with three stacked LSTM units with residual connections to obtain the basic temporal feature vector; the LSTM encoder includes an encoder input layer, an encoder hidden layer, and an encoder output layer; Step S223: Construct a compliance status decoder, specifically by constructing a compliance status decoder structure that includes an LSTM temporal modeling layer, a feature purification layer, and a compliance reconstruction output layer connected in series, and inputting the basic temporal feature vector into the compliance status decoder to obtain the compliance temporal status features and the corresponding compliance reconstruction results; The feature purification layer specifically employs a feature purification structure consisting of a fully connected layer, a layer normalization layer, and a ReLU activation function to map compliant time series features and obtain compliant time series state features. The compliance reconstruction output layer specifically adopts a fully connected layer structure, takes compliance time-series features as input, and outputs compliance reconstruction results through linear mapping. Step S224: Construct the risk state decoder.

4. The intelligent combat vehicle training status monitoring method according to claim 3, characterized in that: In step S224, the construction of the risk state decoder specifically involves constructing a risk state decoder including a shared risk feature extraction layer, an offside risk feature branch, a boundary risk feature branch, and a risk reconstruction output layer. The basic temporal feature vector is input into the risk state decoder, and the offside risk temporal features and boundary risk temporal features are obtained by combining shared risk features with branch reinforcement, and the corresponding offside risk reconstruction results and boundary risk reconstruction results are output. The shared risk feature extraction layer specifically employs two stacked LSTM units to perform temporal modeling on the basic temporal feature vector. After the output of each LSTM unit is processed, it undergoes layer normalization and Dropout processing in sequence to enhance the ability to model the potential risk evolution patterns in the training state and obtain shared risk features. The offside risk feature branch specifically employs an offside risk feature extraction structure consisting of a fully connected layer, a spatial attention mechanism layer, a layer normalization layer, and a ReLU activation function. Shared risk features are input into the offside risk feature extraction structure to obtain offside risk temporal features. The boundary risk feature branch specifically adopts a boundary risk feature extraction structure consisting of a fully connected layer, a spatial attention mechanism layer, a layer normalization layer, and a ReLU activation function. The shared risk features are input into the boundary risk feature extraction structure to obtain the boundary risk temporal features. The risk reconstruction output layer specifically consists of two parallel fully connected reconstruction output layers, which take the off-limits risk temporal features and the out-of-bounds risk temporal features as inputs, and output off-limits risk reconstruction results and out-of-bounds risk reconstruction results that are consistent with the original risk temporal window dimensions.

5. The intelligent combat vehicle training status monitoring method according to claim 1, characterized in that: In step S3, the performance optimization of the monitoring model specifically includes the following steps: Step S31: Monitor model training, specifically by using historical training status monitoring data as training data for the monitoring model, designing a dual-risk temporal loss function as the loss function, and training the model to obtain the trained vehicle training status monitoring model. The dual-risk time-series loss function is specifically a weighted combination of two parts: a time-series reconstruction loss term and a dual-probability calibration loss term. The temporal reconstruction loss term consists of a compliance reconstruction loss term, an off-limits risk reconstruction loss term, an out-of-bounds risk reconstruction loss term, and an unlabeled temporal adaptive loss term. The average of the biprobability calibration loss values ​​of all samples is taken as the final biprobability calibration loss term. The dual-probability calibration loss term specifically involves calculating the cross-entropy loss for the out-of-bounds probability and the cross-entropy loss for the out-of-bounds probability for each sample, and summing the two cross-entropy losses as the dual-probability calibration loss value for that sample. Step S32: Monitor model hyperparameter optimization; Step S33: Optimize and adjust the hyperparameters of the monitoring model. Specifically, based on the optimal combination of hyperparameters of the vehicle training status monitoring model, adjust the hyperparameters of the vehicle training status monitoring model to obtain the optimized vehicle training status monitoring model.

6. The intelligent combat vehicle training status monitoring method according to claim 5, characterized in that: In step S32, the hyperparameter optimization of the monitoring model specifically includes the following steps: Step S321: Population initialization, specifically, encoding the hyperparameters of the vehicle training state monitoring model into individual position vectors, and generating them using circular chaotic mapping. The position vectors of each individual are used to initialize the population, resulting in the initialized population; the formula used is as follows: ； In the formula, Indicates the first The individual initial position vector, Indicates the first The individual initial position vector, and Indicates control parameters; Step S322: Confirm the global optimal position of an individual. Specifically, calculate the fitness value of an individual in the population, use the performance of the vehicle training status monitoring model established based on the individual position as the individual fitness value, sort the individuals from best to worst fitness value, and take the position vector of the individual with the best fitness value as the global optimal position of the individual. Step S323: Update position based on adaptive T-distribution perturbation, specifically by updating the individual position using an adaptive T-distribution; the formula used is as follows: ； In the formula, This represents the time-varying degrees of freedom of the T-distribution in the t-th iteration. Indicates the first The perturbation position of the i-th individual in the next iteration. No. The position of the i-th individual in the next iteration. Indicates the base scaling factor. This indicates that the degree of freedom is T-distributed random variable, This represents the correction factor. Indicates the maximum number of iterations; Step S324: Update the final position based on the sine and cosine strategy. Specifically, based on the perturbation position, update the final position of the current iteration individual using a sine and cosine strategy with adaptive weights; the formula used is as follows: ； In the formula, This represents the adaptive weight value in the t-th iteration. and express Randomness parameters within a range This represents the global optimal position of an individual. Indicates the first The position of the i-th individual in the next iteration; Step S325: Iterative search terminates. Specifically, for all updated individuals, the individual fitness value is recalculated, and the fitness values ​​of all individuals are compared. The global optimal position of the individual is updated. When the fitness value corresponding to the global optimal position of the individual is higher than the fitness threshold and the number of search iterations reaches the maximum number of search iterations, the search is terminated and the global optimal position of the individual is obtained. The global optimal position specifically refers to the optimal hyperparameter combination of the vehicle training state monitoring model.

7. The intelligent combat vehicle training status monitoring method according to claim 1, characterized in that: In step S4, the intelligent monitoring of the vehicle training status specifically involves inputting real-time training status monitoring data into the optimized vehicle training status monitoring model to obtain real-time vehicle overstepping probability prediction results and real-time vehicle boundary crossing probability prediction results. Based on the prediction results, the corresponding vehicle response strategy is triggered to achieve real-time monitoring of vehicle training.

8. The intelligent combat vehicle training status monitoring method according to claim 1, characterized in that: In step S1, the multi-source data acquisition specifically involves obtaining raw data for vehicle training status monitoring through data acquisition operations, and then performing data optimization operations on the raw data to obtain optimized data for vehicle training status monitoring. The raw data for vehicle training status monitoring includes historical training status monitoring data and real-time training status monitoring data. The historical and real-time training status monitoring data include vehicle operation data, maneuver control data, vehicle training data, training environment data, and personnel operation data. The historical training status monitoring data also includes historical vehicle overstepping prediction results and historical vehicle boundary crossing prediction results. The data optimization operations include data cleaning, data normalization, data encoding processing, and data feature selection.

9. The intelligent armored vehicle training status monitoring system according to claim 1, used to implement the intelligent armored vehicle training status monitoring method according to any one of claims 1-8, characterized in that: It includes a multi-source data acquisition module, a vehicle training status monitoring model construction module, a monitoring model performance optimization module, and a vehicle training status intelligent monitoring module; The multi-source data acquisition module obtains optimized data for vehicle training status monitoring through data acquisition and data optimization operations, and sends the data to the monitoring model performance optimization module and the intelligent monitoring module for vehicle training status. The module for constructing the vehicle training status monitoring model first extracts the morphological features of the training process as static constraint factors for risk probability prediction. Then, it uses an improved structured autoencoder combined with type-time window annotation to extract the offside risk time-series features and the boundary risk time-series features. Subsequently, it performs serial dual-risk feature fusion, progressively fusing the morphological features of the training process, the offside risk time-series features, and the boundary risk time-series features to generate offside fusion features and boundary fusion features. Finally, it outputs the vehicle offside probability prediction result and the vehicle boundary probability prediction result through dual independent probability prediction, thus completing the construction of the vehicle training status monitoring model and obtaining the vehicle training status monitoring model. The data is then sent to the monitoring model performance optimization module. The monitoring model performance optimization module receives data from the multi-source data acquisition module and the vehicle training status monitoring model construction module. Specifically, it designs a dual-risk temporal loss function with a temporal reconstruction loss term and a dual-probability calibration loss term as the objective function for training the monitoring model. It then improves the optimization algorithm through circular chaotic mapping, adaptive T-distribution perturbation position, and sine and cosine strategies with adaptive weights to obtain the optimal hyperparameter combination of the vehicle training status monitoring model. Finally, it adjusts the hyperparameters of the monitoring model according to the optimal hyperparameter combination to obtain the optimized vehicle training status monitoring model and sends the data to the intelligent monitoring module for vehicle training status. The intelligent monitoring module for vehicle training status receives data from the multi-source data acquisition module and the monitoring model performance optimization module. Specifically, it inputs real-time data into the optimized vehicle training status monitoring model to obtain real-time vehicle offside and boundary probability prediction results, and adopts corresponding vehicle response strategies based on the prediction results.

Images