Battlefield situation prediction and interpretability evaluation method fusing EMD (empirical mode decomposition) and LSTM (long short term memory) models

Abstract

The invention belongs to the technical field of military information, and particularly relates to a battlefield situation prediction and interpretability evaluation method fusing EMD and LSTM models. According to the method, the EMD technology is introduced for data preprocessing, complex battlefield data is decomposed into feature information of different time scales through EMD, the representativeness of input data is improved, noise interference is reduced, and therefore high-quality input is provided for an LSTM prediction model. According to the method, the LSTM network is fused for dynamic prediction, the dynamic change of the battlefield situation can be accurately captured based on the time sequence modeling capability of the LSTM network, and the prediction accuracy of the model is improved. According to the method, result explanation is performed in combination with the SHAP, the model is explained and analyzed through the SHAP framework, contributions of different input variables to a battlefield situation prediction result are disclosed, a clear decision basis is provided for a commander, and the interpretability and the application value of the model are enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of military information technology, and specifically relates to a battlefield situation perception and assessment method, which realizes efficient processing and dynamic prediction of multi-source heterogeneous battlefield data by combining signal processing, deep learning and explanatory analysis technology. Background Art

[0002] With the rapid development of information technology, military informatization has entered a new stage of development. Modern warfare is rapidly evolving from traditional mechanized warfare to information-based and intelligent warfare. This trend not only changes combat methods and command approaches but also places higher demands and challenges on battlefield situational awareness and assessment technologies. Battlefield situation, as a comprehensive reflection of the dynamic interaction between enemy and friendly forces within the battlefield environment, lies at its core in real-time understanding of current conditions and accurate prediction of future development trends. Faced with a rapidly changing battlefield situation, accurate analysis and assessment of battlefield situation is key to achieving strategic advantage and a crucial guarantee for seizing the initiative in future wars.

[0003] The battlefield situation is a dynamic system composed of multi-dimensional and multi-layered information elements, primarily comprising real-time updates on troop deployments, environmental changes, and combat events. These elements together form the core of the battlefield landscape and provide commanders with critical decision-making foundations. On the modern battlefield, the comprehensiveness, real-time nature, and accuracy of information acquisition directly determine combat effectiveness. However, faced with massive, diverse, and rapidly changing battlefield data, extracting key information has become a primary challenge in command and decision-making. Traditional combat methods cannot meet the high demands of situational awareness on the modern battlefield. Efficient and intelligent situational awareness technology is becoming a core pillar of the next generation of military strategy.

[0004] Traditional battlefield situation analysis techniques are typically based on objective data, relying primarily on expert experience and intuitive analysis. While this approach has some practicality in static scenarios, it exhibits significant limitations in dynamic, changing, and complex battlefield environments. On the one hand, traditional methods lack the ability to analyze systematic correlations among multiple targets, making it difficult to reveal potential relationships between enemy and friendly targets. On the other hand, they are limited by computing power and model complexity, making them ineffective in predicting enemy action trends and assessing the development of war. These shortcomings are becoming increasingly prominent as battlefield environments become more complex and operational demands grow.

[0005] Existing situational awareness and assessment methods, such as Bayesian networks, case-based reasoning (CBR), fuzzy logic, genetic algorithms, and DS evidence theory, are often studied. For example, in Reference 1 (Jia Q, Hu J, Zhang W, et al. A new situation assessment method for aerial targets based on linguistic fuzzy sets and trapezium clouds [J]. Engineering Applications of Artificial Intelligence, 2023, 117:105610), this method takes a multi-criteria group decision-making perspective. It uses three linguistic fuzzy sets to represent threat information, proposes the related concept of trapezium clouds and its aggregation operator, and proposes a weight calculation method based on multi-objective planning that comprehensively considers both subjective and objective factors. While these methods have improved the scientific nature of battlefield situational awareness to a certain extent, they largely rely on expert experience and subjective judgment. This not only increases uncertainty in the reasoning process but also limits the intelligent level of situational assessment. Furthermore, these methods typically require extensive human input and exhibit low processing efficiency and insufficient accuracy when processing large-scale and dynamic battlefield data. They are no longer able to meet the requirements of rapid response and precise decision-making in information warfare.

[0006] CN117291475A, a battlefield situation assessment method and system based on deep learning, establishes a battlefield database, and after data fusion and information conversion, uses intelligence correlation analysis to use military knowledge and battlefield environmental conditions as heuristic knowledge to conduct battlefield intelligence information correlation mining and establish a battlefield intelligence information database; applies uncertainty reasoning and comprehensive evaluation methods to establish a battlefield situation evaluation library; uses a deep learning network model to obtain a battlefield target group situation forecast, and applies integrated reasoning and decision fusion methods to obtain a global forecast of the battlefield situation.

[0007] CN110472296B, a method for air combat target threat assessment based on a standardized fully connected residual network, constructs a standardized fully connected residual network within a TensorFlow database and creates a TensorFlow session. This method uses a fully connected network to map input data into a high-dimensional space and ultimately into a sample label space, achieving data classification. It can self-learn the distribution of input data and uncover patterns hidden within the data.

[0008] CN110490422B, a target combat effectiveness situation assessment method based on a game cloud model, constructs a target combat effectiveness assessment index system; secondly, constructs a target combat effectiveness situation assessment cloud model, forms a quantitative decision matrix, and obtains the corresponding cloud expectation and cloud entropy vector matrices; finally, the game cloud model is used to obtain the optimal combination weight and the game cloud center of gravity vector, thereby determining the weighted deviation of the game cloud, activating the game cloud generator, and judging the performance status of the target combat effectiveness.

[0009] Overall, the lack of dynamic prediction capabilities is a significant shortcoming of existing methods. The modern battlefield environment is rapidly changing, and situational assessment requires not only real-time perception of the current state but also effective prediction of future trends. However, many traditional methods are still limited to static analysis and have limited ability to model the dynamic evolution of situational factors. The root of this problem lies in the fact that traditional methods often rely on simple time weighting or trend fitting, which cannot cope with the high dynamics and uncertainty of the battlefield environment. Furthermore, while some emerging machine learning technologies have shown some potential, their "black box" nature makes the results lack interpretability, making it difficult for commanders to make reasonable decisions based on the assessment results. These technical shortcomings significantly limit their value in practical applications. Summary of the Invention

[0010] In response to the defects of the existing technology, the purpose of the present invention is to provide a battlefield situation prediction and interpretability evaluation method that integrates the EMD and LSTM models, provides a new approach to the battlefield situation prediction problem, combines the empirical mode decomposition algorithm in the field of signal decomposition with the deep learning algorithm, and improves the accuracy and credibility of battlefield situation prediction.

[0011] In order to achieve the above objectives, the present invention integrates the battlefield situation prediction and interpretability evaluation method of EMD and LSTM models, referring to Figure 1 The model structure diagram is shown in Figure 2. The model is constructed for experimentation. The steps are as follows:

[0012] Step 1: Conduct a simulation experiment to label data. Features extracted from the raw data are labeled and categorized, and seven situational features are defined and calculated: Enemy and friendly troop strength comparison: measures the difference in troop strength between the two sides; Enemy and friendly defensive capability comparison: comprehensively evaluates defensive fortifications and defensive equipment; Enemy and friendly offensive capability comparison: based on firepower coverage and firepower output per unit time; Enemy and friendly terrain position comparison: comprehensively evaluates differences in terrain control and high ground occupation; Enemy and friendly weapon firepower comparison: evaluates the precision strike and long-range firepower capabilities of equipment; Enemy and friendly target distance comparison: based on the proximity of target locations; Enemy and friendly artillery threat comparison: analyzes the threat posed to the target by artillery type, range, and tactical flexibility. Using a combination of expert annotation and cluster analysis, samples are labeled, resulting in five situational levels: absolute advantage, advantage, parity, disadvantage, and absolute disadvantage, with corresponding label values of 1, 2, 3, 4, and 5.

[0013] Step 2: The original situation data is enhanced using a sliding window method combined with noise perturbation to improve the diversity and robustness of the time series data.

[0014] Step 3, refer to Figure 2 The flowchart uses empirical mode decomposition (EMD) to perform time-frequency decomposition on the original battlefield situation data, extracts the inherent laws and multi-scale feature information from the complex non-stationary data, and obtains the inherent time-frequency information structure and change law of the battlefield situation data;

[0015] Step 4: For the multi-scale components generated by EMD decomposition, spectrum analysis and wavelet threshold method are combined to achieve effective noise reduction, and the signal quality and reliability are improved by reconstructing and integrating feature information.

[0016] In step 5, the gated recurrent unit is selected as the basic unit, the spatiotemporal feature information obtained in the previous two steps is aggregated, the processed battlefield situation data is deeply modeled, and finally the prediction result is obtained through the output of the fully connected layer.

[0017] In step 6, based on the Shapley value principle in game theory, the difference in predicted output with and without a certain feature is compared. The marginal contribution value of a single feature is calculated using the SHAP formula. The feature contribution values of all samples are summarized to generate a global feature importance distribution map.

[0018] Preferably, step 010 further includes the following steps:

[0019] Step 2.1, sliding window processing: select the optimal values through experimental tuning and determine the key parameters such as the sliding window size W(s) and step size S(s).

[0020] A starting point is randomly selected from the original dataset. Based on the selected starting point, consecutive time steps are grouped into a window segment W(s). The data within the window is a set of feature vectors used to generate new data segments. Linear interpolation is performed between the first and last data points of the window to capture transitional changes in the time series data and generate new intermediate data points.

[0021] To increase the diversity of data segments, the window is moved with a fixed step size S(s). The above process is repeated to generate more window segments at different positions, thereby expanding the dataset. The generated new data samples are stored and merged with the original data, retaining the original feature columns and corresponding label columns to ensure the integrity and consistency of the new data.

[0022] Step 2.2, noise perturbation processing: Superimpose Gaussian noise on the generated new data samples, and define the noise intensity to simulate the uncertainty and sensor error in the real environment; assign the generated new samples a label consistent with the starting point of the window, and finally save the generated enhanced dataset for subsequent analysis or model training.

[0023] Preferably, step 3 further comprises the following steps:

[0024] In step 3.1, perform a comprehensive scan of the raw battlefield situation data at each time step, examining the changing trends of each value. Determine the precise locations of all local maxima and minima and record these locations as an extreme point sequence. Use interpolation to connect all local maxima to generate an upper envelope, and concatenate all local minima to generate a lower envelope. Ensure smoothness and data integrity during the interpolation process to avoid envelope distortion caused by spurious extremes. Verify that the generated upper and lower envelopes cover the entire oscillation range of the signal, providing an accurate benchmark for subsequent calculations.

[0025] In step 3.2, take the average of the corresponding points in each time step of the upper and lower envelopes to obtain the envelope mean sequence. Subtract this mean from the original signal to obtain a new signal component. The following conditions are checked for this new signal component: whether the number of extreme points and zero crossings differs by no more than 1; and whether the upper and lower envelopes are nearly symmetrical. If these conditions are not met, use the current signal component as the new input and repeat steps 021 and 022 until the conditions are met, ultimately extracting the first set of intrinsic mode functions (IMFs).

[0026] In step 3.3, after extracting each component that meets the IMF criteria, remove it from the signal to obtain a residual signal. This process is repeated until the residual signal becomes monotonic. Ultimately, the complex original signal is decomposed into several IMF components and a residual term, thereby achieving the decomposition and extraction of the signal's multi-scale features.

[0027] Preferably, step 4 specifically includes the following steps:

[0028] Step 4.1: Calculate the spectral density distribution of each IMF component based on the short-time Fourier transform (STFT) to identify the high-frequency noise-dominant component with low energy content and large frequency fluctuations, thereby avoiding excessive filtering of low-frequency feature information.

[0029] Step 4.2: Perform wavelet decomposition on the high-frequency IMF components, set an adaptive soft threshold, and dynamically adjust the noise reduction threshold according to the continuous mean square error criterion to accurately remove random noise components while retaining the useful signal to the greatest extent possible;

[0030] In step 4.3, the denoised IMF components and residual terms are sorted by spectral density and weighted integrated to generate the denoised signal. Combining a multi-scale feature fusion strategy optimizes the correlation and integrity of time series, providing highly robust input data for model training.

[0031] Preferably, step 5 specifically includes the following steps:

[0032] In step 5.1, the spatiotemporal feature information obtained in the previous two steps is aggregated to extract key battlefield features from the noise reduction signal, such as troop strength comparison and firepower comparison. The feature values are normalized to zero mean based on the data distribution to reduce the impact of the feature dimension on model training and accelerate network convergence.

[0033] Step 5.2: Build a multi-layer LSTM network. Each unit uses a forget gate to control the retention ratio of the previous moment's state, an input gate to filter the current moment's information, and an output gate to determine the impact of the current state on the next moment. Combine this with a dropout strategy to avoid overfitting.

[0034] Step 5.3: Map the LSTM output into the probability distribution of five situation levels through the fully connected layer, use the Softmax function to output the final evaluation result, and combine the decision rules to realize the situation level classification, providing commanders with real-time evaluation basis.

[0035] Preferably, step 6 specifically includes the following steps:

[0036] In step 6.1, based on the SHAP value calculation framework, the marginal contribution of each feature to the prediction result is evaluated by the Shapley value theory, and the impact of different input features is quantified in combination with the weighting strategy.

[0037] Step 6.2: Calculate the SHAP value of each feature for a single sample and analyze how the feature positively or negatively affects the prediction result;

[0038] Step 6.3: Summarize the feature importance distribution of all samples to reveal the degree of dependence of the model on key features.

[0039] Step 6.4: Use SHAP force diagrams, decision diagrams, and feature interaction diagrams to demonstrate the sensitivity of features to prediction results and their positive and negative impacts, providing decision makers with an intuitive and clear basis for situation assessment.

[0040] Compared with the existing technical solutions, the present invention has the following beneficial effects:

[0041] 1) Introducing EMD technology for data preprocessing: EMD decomposes complex battlefield data into feature information at different time scales, improving the representativeness of the input data and reducing noise interference, thereby providing high-quality input for the LSTM prediction model.

[0042] (2) Integrating LSTM network for dynamic prediction: Based on the time series modeling capability of LSTM network, it can accurately capture the dynamic changes of battlefield situation and improve the prediction accuracy of the model.

[0043] (3) Combined with SHAP to interpret the results: The model is interpreted and analyzed through the SHAP framework to reveal the contribution of different input variables to the battlefield situation prediction results, providing commanders with a clear decision-making basis and enhancing the interpretability and application value of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a structural diagram of the technical model of the present invention;

[0045] Figure 2 It is a flow chart for EMD signal decomposition;

[0046] Figure 3 This is the decomposition result graph of the EMD algorithm;

[0047] Figure 4 It is the confusion matrix and result graph of battlefield situation prediction by the method of the present invention;

[0048] Figure 5 This is an explanatory diagram of the influence of the full sample characteristics on the classification results of the battlefield situation prediction results of the method of the present invention;

[0049] Figure 6 This is an explanatory diagram of the impact of single sample features on classification results when predicting battlefield situation using the method of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, other embodiments obtained by ordinary technicians in this field without making creative efforts are all within the scope of protection of the present invention.

[0051] The battlefield situation prediction and interpretability evaluation method integrating EMD and LSTM models includes the following steps:

[0052] Step 1: Conduct a simulation experiment to label data. Features extracted from the raw data are labeled and categorized, and seven situational features are defined and calculated: Enemy and Friendly Force Number Comparison (F1): measures the difference in troop numbers between the two sides; Enemy and Friendly Defense Capability Comparison (F2): a comprehensive assessment based on fortifications and defensive equipment; Enemy and Friendly Offensive Capability Comparison (F3): based on firepower coverage and firepower output per unit time; Enemy and Friendly Terrain Position Comparison (F4): a comprehensive assessment of differences in terrain control and high ground occupation; Enemy and Friendly Weapon Firepower Comparison (F5): an assessment of equipment precision strike and long-range firepower capabilities; Enemy and Friendly Target Distance Comparison (F6): based on proximity to target locations; and Enemy and Friendly Artillery Threat Comparison (F7): an analysis of the threat posed to the target by artillery type, range, and tactical flexibility. Samples are labeled using a combination of expert annotation and cluster analysis, resulting in five situational levels: absolute advantage, advantage, parity, disadvantage, and absolute disadvantage, with corresponding label values of 1, 2, 3, 4, and 5, respectively.

[0053] Step 2: Data enhancement processing:

[0054] The original situation data is enhanced by combining the sliding window method with noise disturbance, and the enhanced data is stored in a CSV file.

[0055] Step 2.1: Sliding window processing: Select the optimal values through experimental tuning and determine the key parameters such as the sliding window size W(s) and step size S(s).

[0056] Randomly from the original data set X orig = Select a starting position idx from [F1, F2, F3, F4, F5, F6, F7] to ensure that the generated window fragment has good randomness and coverage. The starting range is set to idx∈[0, len(X orig )-window_size] to ensure that the sliding window does not exceed the data boundary.

[0057] Based on the selected starting point, the continuous time steps are grouped into a window segment W(s) of length window_size. The data inside the window is a set of feature vectors used to generate new data segments.

[0058] Between the first and last data points of the window, a random interpolation factor interp_factor∈[0,1] is applied for linear interpolation. The interpolation result is: X interpolated [t] = X start [t]+interp_factor×(X end [t]-X start[t]), the result of linear interpolation can effectively capture the smooth transition and trend of time series data and generate new intermediate data points.

[0059] To increase the diversity of data segments, the window is moved with a fixed step size S(s). The above process is repeated to generate more window segments at different positions, thereby expanding the dataset. The generated new data samples are stored and merged with the original data, retaining the original feature columns and corresponding label columns to ensure the integrity and consistency of the new data.

[0060] Step 2.2: Noise perturbation processing: Superimpose Gaussian noise on the generated new data samples, and define the noise intensity to simulate the uncertainty and sensor error in the real environment; assign the generated new samples a label consistent with the starting point of the window, and finally save the generated enhanced dataset as a new CSV file for subsequent analysis or model training.

[0061] Step 3: EMD signal decomposition:

[0062] according to Figure 2 The steps shown are to perform time-frequency decomposition of battlefield situation data through empirical mode decomposition (EMD) to extract its inherent laws and characteristic information. The specific steps are as follows:

[0063] Step 3.1: Extract signal extreme points:

[0064] The original battlefield situation data of each time step is fully scanned, and the changing trend of its value is checked one by one. The following method is used to calculate the local extreme point: LocalMaxima:{t k :F i (t k )<F i (t k-1 ),F i (t k )<F i (t k+1 )}LocalMaxima:{t k :F i (t k )>F i (t k-1 ),F i (t k )>F i (t k+1 )}

[0065] Determine the exact locations of all local maxima and minima k , and record these positions as extreme point sequences.

[0066] Use spline interpolation to connect all local maxima to generate the upper envelope, and connect all local minima to generate the lower envelope. This ensures smoothness and data integrity during the interpolation process, and avoids envelope distortion caused by pseudo-extremes. The upper and lower envelopes are calculated as follows:

[0067]

[0068]

[0069] Check that the generated upper and lower envelopes cover the full range of the signal's oscillations, providing an accurate baseline for subsequent calculations. Ideally, the envelopes should completely encompass the original signal to ensure they accurately represent its full range.

[0070] Step 3.2: Calculate the mean of the upper and lower envelopes:

[0071] Take the average value of the corresponding points of each time step in the upper and lower envelopes, Envelope Mean(t), to obtain the envelope mean sequence. The mean calculation method is as follows:

[0072]

[0073] Then subtract the corresponding envelope mean from the original signal to obtain the new signal component R(t): R(t) = F i (t)-Envelope Mean(t).

[0074] The following conditions are checked for the new signal components: whether the difference between the number of extreme points and zero crossing points does not exceed 1; whether the upper and lower envelopes are nearly symmetrical.

[0075] If the above conditions are not met, the current signal component is used as a new input, and steps S3.1 and S3.2 are repeated until the conditions are met, and finally the first set of intrinsic mode functions (IMFs) are extracted.

[0076] Step 3.3: IMF component extraction and residual calculation:

[0077] The signal components that meet the IMF conditions are removed from the original signal and saved as independent IMF components.

[0078] The residual signal is obtained by subtracting the extracted IMF component from the original signal.

[0079] Repeat the process from S3.1 to S3.3 for the residual signal, and gradually extract all IMF components until the residual signal is monotonic.

[0080] Finally, the original signal is decomposed into a set of IMF components {IMF1, IMF2, ···, IMF n} and a residual term rn .Right now:

[0081] Step 4: EMD noise reduction processing

[0082] In order to improve the signal-to-noise ratio of the situation data, the high-frequency noise-dominated IMF component after EMD decomposition is subjected to denoising. The specific steps are as follows:

[0083] Step 4.1: Identify high-frequency components: Calculate the spectral energy density Ω of the IMF component i , select the IMF component dominated by high-frequency noise for processing.

[0084]

[0085] Step 4.2: Wavelet threshold denoising: Apply wavelet transform to the selected high-frequency IMF components and remove noise using the soft thresholding method. Use the continuous mean square error criterion to determine the order K0 of the IMF components that require wavelet thresholding:

[0086]

[0087] Perform threshold estimation on the first k-order components to determine the standard deviation Δ of the component disturbance amplitude k and component dynamic threshold θ k :

[0088]

[0089]

[0090] Denoising of high frequency components:

[0091]

[0092] Step 4.3: Signal reconstruction: Reconstruct the denoised high-frequency components, low-frequency components and residual terms to obtain the denoised situation feature signal.

[0093]

[0094] Step 5: LSTM-based situation assessment modeling

[0095] The long short-term memory neural network (LSTM) is used to model the processed situation data. The specific steps are as follows:

[0096] Step 5.1: Input feature selection and normalization:

[0097] Features are extracted from the enhanced situation data and normalized to accelerate model training.

[0098] Step 5.2: LSTM unit design:

[0099] The forget gate, input gate, and output gate mechanisms are used to control the flow of information. The meanings of the parameters are as follows:

[0100] x' t : The characteristic variable input at the current moment, representing the situation characteristics observed at the current moment

[0101] S t-1 : The situation level status at the previous moment, used to indicate the impact of the historical status on the current moment.

[0102] H t-1 : The implicit situation feature output of the previous moment, indicating the memory of the situation feature at the previous moment.

[0103] S t : The current situation level status.

[0104] H t : Situation feature output at the current moment.

[0105] G ft : Situation forgetting weight, controls the retention of the previous moment information.

[0106] G it : Situation input weight, which controls the feature information of the current input.

[0107] G ot : Situation output weight, which controls the output result of situation level status.

[0108] : Candidate situation status at the current moment.

[0109] σ corresponds to the Sigmoid activation functions of the three gates respectively.

[0110] Forget Gate: Filters the historical situation level status that needs to be retained.

[0111] G ft =σ(W f [H t-1 ,x' t ]+b f )

[0112] Here G ft Control the state S of the previous moment t-1 The degree of retention is to retain historical information related to the current situation characteristics.

[0113] Input Gate: Select to receive the current situation feature input.

[0114] G it =σ(W i [H t-1,x' t ]+b i )

[0115]

[0116] G it Control the influence of the current input situation characteristics on the current situation level, and It is the candidate state at the current moment obtained through nonlinear mapping.

[0117] Update situational status information.

[0118]

[0119] S t Integrates the situation level state S of the previous moment t-1 and the current input situation feature set, F t Reflects the comprehensive situation level at the current moment.

[0120] Output gate: Determine the output of the current situation characteristics

[0121] G ot =σ(W0[H t-1 ,x' t ]+b0)

[0122] H t =G ot *tanh(S t )

[0123] Step 5.3: Model training and optimization: Use cross-validation to evaluate model performance and use the Adam optimization algorithm to update model parameters.

[0124] Step 6: SHAP model interpretation

[0125] In order to improve the interpretability and credibility of the model, SHAP is applied to interpret the output of the EMD-LSTM model:

[0126] Step 6.1: Calculation of feature contribution value:

[0127] The marginal contribution of each feature to the prediction result is calculated according to the SHAP method:

[0128] Step 6.2: Local and global interpretation analysis:

[0129] Local analysis: Decomposing the feature contribution value of the prediction results of a single sample;

[0130] Global analysis: summarizes the importance distribution of features in all samples.

[0131] Step 6.3: Visualization:

[0132] Use SHAP value charts to show the positive or negative impact of features and the sensitivity of output to changes in feature values.

[0133] The multi-step data processing and modeling strategy proposed in the present invention enables it to handle complex nonlinear and non-stationary signals in practical applications, thereby improving the reliability and accuracy of prediction results.

[0134] The present invention does not require simulation and numerical simulation software to establish a complex model and does not require a large amount of computing resources.

[0135] This invention combines the Shapley value principle based on game theory with other deep learning methods and has a certain degree of explainability.

[0136] With the above-described preferred embodiments of the present invention as a guide, those skilled in the art will readily be able to make various changes and modifications without departing from the technical spirit of the present invention. The technical scope of the present invention is not limited to the contents of the specification and must be determined in accordance with the scope of the claims.

Claims

1. A battlefield situation prediction and interpretability evaluation method integrating EMD and LSTM models, characterized by: Here are the steps: Step 1: Conduct simulation experiment to mark data, mark and classify the features extracted from the original data, and define and calculate seven situational features: enemy and friendly force number comparison: measure the difference in the number of forces on both sides; enemy and friendly defense capability comparison: a comprehensive evaluation based on fortifications and defensive equipment; enemy and friendly attack capability comparison: based on firepower coverage and firepower output capability per unit time; enemy and friendly terrain position comparison: a comprehensive evaluation of the differences between the two sides in terrain control and high ground occupation; enemy and friendly weapon firepower comparison: an evaluation of the equipment's precision strike and long-range firepower strike capabilities; enemy and friendly target distance comparison: based on the proximity of target positions; Comparison of enemy and friendly artillery threats: Analyze the threat posed to targets by artillery types, ranges, and tactical flexibility; We use a combination of expert annotation and cluster analysis to label samples and classify them into five levels of situation: absolute advantage, advantage, balance, disadvantage, and absolute disadvantage. The corresponding label values are 1, 2, 3, 4, and 5, respectively. The processed dataset is called the original battlefield situation dataset. Step 2: The original battlefield situation dataset is enhanced using a sliding window method combined with noise perturbation to improve the diversity and robustness of the time series data, thereby obtaining an enhanced battlefield situation dataset. Step 3: Use empirical mode decomposition (EMD) to perform time-frequency decomposition on the enhanced battlefield situation dataset obtained through data enhancement in step 2, extract inherent laws and multi-scale feature information from the complex non-stationary data, and obtain the inherent time-frequency information structure and change laws of the battlefield situation data; Step 4: For the multi-scale components generated by EMD decomposition, effective noise reduction is achieved by combining spectrum analysis and wavelet threshold method, and the signal quality and reliability are improved by reconstructing and integrating feature information; Step 5: Select the gated recurrent unit as the basic unit, aggregate the spatiotemporal feature information obtained in the previous two steps (i.e., steps 3 and 4), perform deep modeling on the processed battlefield situation data, and finally output it through the fully connected layer to obtain the prediction result; In step 6, based on the Shapley value principle in game theory, the difference in predicted output with and without a certain feature is compared. The marginal contribution value of a single feature is calculated using the SHAP formula. The feature contribution values of all samples are summarized to generate a global feature importance distribution map.

2. The battlefield situation prediction and interpretability evaluation method integrating EMD and LSTM models as claimed in claim 1, characterized in that: The specific operations of step 2 are as follows: Step 2.1, sliding window processing: select the optimal values through experimental tuning and determine the sliding window size W(s) and step size S(s); A starting position is randomly selected from the original battlefield situation dataset. Based on the selected starting point, the continuous time steps are grouped into a window segment W(s). The data within the window is a set of feature vectors used to generate new data segments. Linear interpolation is performed between the first and last data points of the window to capture the transitional changes in the time series data and generate new intermediate data points. The window is moved repeatedly with a step size of S(s) to generate more window segments at different positions, thereby expanding the data set. The generated new data samples are stored and merged with the original data, retaining the original feature columns and corresponding label columns to ensure the integrity and consistency of the new data. Step 2.2, noise perturbation processing: Superimpose Gaussian noise on the generated new data samples, and define the noise intensity to simulate the uncertainty and sensor error in the real environment; assign the generated new samples a label consistent with the starting point of the window, and finally save the generated enhanced dataset for subsequent analysis or model training.

3. The battlefield situation prediction and interpretability evaluation method integrating EMD and LSTM models as claimed in claim 1, characterized in that: The specific operations of step 3 are as follows: Step 3.1: Fully scan the enhanced battlefield situation dataset obtained in Step 2, observing each time step and examining the changing trends of its values one by one. Determine the precise locations of all local maxima and minima and record these locations as an extreme point sequence. Use interpolation to connect all local maxima to generate an upper envelope, and all local minima to generate a lower envelope. Ensure smoothness and data integrity during the interpolation process to avoid envelope distortion caused by spurious extreme values. Check whether the generated upper and lower envelopes cover the entire oscillation range of the signal, providing an accurate benchmark for subsequent calculations. In step 3.2, take the average value of the corresponding points in each time step of the upper and lower envelopes to obtain the envelope mean sequence; subtract this mean value from the original signal to obtain a new signal component; and perform the following condition checks on the new signal component: whether the difference between the number of extreme points and zero crossings does not exceed 1; and whether the upper and lower envelopes are nearly symmetrical. If these conditions are not met, use the current signal component as the new input and repeat steps 3.1 and 3.2 until the conditions are met, finally extracting the first set of intrinsic mode functions (IMFs). In step 3.3, after extracting each component that meets the IMF conditions, remove it from the signal to obtain a residual signal; repeat this process until the residual signal tends to be monotonic, and finally decompose the complex original signal into several IMF components and a residual term, thereby realizing the decomposition and extraction of the multi-scale features of the signal.

4. The battlefield situation prediction and interpretability evaluation method integrating EMD and LSTM models as claimed in claim 1, characterized in that: The specific operations of step 4 are as follows: Step 4.1: Calculate the spectral density distribution of each IMF component based on the short-time Fourier transform (STFT) to identify the high-frequency noise-dominant component with low energy content and large frequency fluctuations, thereby avoiding excessive filtering of low-frequency feature information. Step 4.2: Perform wavelet decomposition on the high-frequency IMF components, set an adaptive soft threshold, and dynamically adjust the noise reduction threshold according to the continuous mean square error criterion to accurately remove random noise components while retaining the useful signal to the greatest extent possible; In step 4.3, the denoised IMF components and residual terms are sorted according to spectral density and then weighted integrated to generate a denoised signal; combined with the multi-scale feature fusion strategy, the correlation and integrity between time series are optimized to provide highly robust input data for model training.

5. The battlefield situation prediction and interpretability evaluation method integrating EMD and LSTM model as claimed in claim 1, characterized in that: The specific operations of step 5 are as follows: Step 5.1: Aggregate the spatiotemporal feature information obtained in the previous two steps to extract key battlefield features from the noise-reduced signal, such as troop strength comparison and firepower comparison. Based on the data distribution, perform zero-mean normalization on the feature values to reduce the impact of feature dimensions on model training and accelerate network convergence. Step 5.2: Build a multi-layer LSTM network. Each unit uses a forget gate to control the retention ratio of the previous moment's state, uses an input gate to filter the current moment's information, and uses an output gate to determine the impact of the current state on the next moment. Combined with the Dropout strategy to avoid overfitting; Step 5.3: Map the LSTM output into the probability distribution of five situation levels through the fully connected layer, use the Softmax function to output the final evaluation result, and combine the decision rules to realize the situation level classification, providing commanders with real-time evaluation basis.

6. The battlefield situation prediction and interpretability evaluation method integrating EMD and LSTM models as claimed in claim 1, characterized in that: The specific operations of step 6 are as follows: Step 6.1: Based on the SHAP value calculation framework, the marginal contribution of each feature to the prediction result is evaluated using the Shapley value theory, and the impact of different input features is quantified by combining the weighting strategy; Step 6.2: Calculate the SHAP value of each feature for a single sample and analyze how the feature positively or negatively affects the prediction result; Step 6.3: Summarize the feature importance distribution of all samples to reveal the degree of dependence of the model on key features; Step 6.4: Use SHAP force diagrams, decision diagrams, and feature interaction diagrams to demonstrate the sensitivity of features to prediction results and their positive and negative impacts, providing decision makers with an intuitive and clear basis for situation assessment.

Images