CN-122020514-A - Intelligent target behavior prediction method based on multi-feature fusion

CN122020514ACN 122020514 ACN122020514 ACN 122020514ACN-122020514-A

Abstract

An intelligent target behavior prediction method based on multi-feature fusion belongs to the technical field of target behavior prediction. Introducing LSTM and CNN to realize target behavior prediction, using LSTM to acquire time sequence characteristics, constructing a dynamic periodic chart to acquire periodic characteristics, using CNN to acquire spatial characteristics, fusing multi-element time sequence data time sequence characteristics, periodic characteristics and spatial characteristics, using multi-characteristics to realize target behavior prediction, and predicting target behaviors through longitude, latitude, altitude, speed, acceleration, doppler speed, distance and target type information generated by target motion. According to the invention, the time sequence characteristics, the period characteristics and the space characteristics of the target motion are fused, the motion characteristics of the target are comprehensively analyzed and judged, and the accuracy of target behavior prediction is improved.

Inventors

Qiao Dianfeng
HAN CHUNLEI
ZHANG YANG
ZHAO WANG
LU YAO
GUO FENGJUAN
JIN ZHONGQIAN
YANG DI
AN RUI

Assignees

中国电子科技集团公司第二十研究所

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (10)

1. An intelligent target behavior prediction method based on multi-feature fusion is characterized by comprising the following steps: step 1, constructing a multi-element time sequence data set, and preprocessing the data of the multi-element time sequence data set; step 2, constructing a time feature extraction module; The time feature extraction model comprises a CNN layer, an attention layer and a full connection layer; Step 3, extracting periodic characteristics from the preprocessed multi-element time sequence to obtain periodic characteristics; step 4, extracting time sequence characteristics of the preprocessed multi-element time sequence through an LSTM network to obtain the time sequence characteristics; step 5, extracting time characteristics from the time characteristics and the period characteristics through a time characteristic extraction module to obtain a target behavior prediction sequence based on the time characteristics and the period characteristics, namely an instant characteristic sequence; Step 6, converting the one-dimensional sequence in the preprocessed multi-element time sequence into a two-dimensional characteristic map through data structure conversion; Step 7, inputting the two-dimensional feature map into a CNN network for convolution operation to obtain a target behavior prediction sequence based on spatial features, namely a spatial feature sequence; step 8, carrying out weighted average on the time feature sequence and the space feature sequence to obtain a predicted target behavior sequence; Step 9, evaluating a prediction model through a loss function and an evaluation index, wherein the prediction model comprises an LSTM network, a time feature extraction module, a space feature extraction module and a CNN network; And step 10, predicting target behaviors of the multi-element time sequence generated by the target motion through a prediction model.
2. The intelligent target behavior prediction method based on multi-feature fusion according to claim 1, wherein in step 1, the multiple time series data sets Wherein X t represents the T moment, and T represents the total time length of the time series data, and the time series data has 8 dimensions, wherein the 8 dimensions are longitude, latitude, altitude, speed, acceleration, doppler speed, distance and target type respectively; Dynamically updating a plurality of time series data sets X through a sliding window, and moving a subset of time series of known target movements As the input of a time sequence prediction model, w is the window length, and the behavior sequence of the future motion of the target is predicted, wherein the time window length of the behavior sequence is w 𝜎 , the window moves forwards along with the time, the old sequence is replaced by the new multi-element time sequence, and the time sequence of the motion of the target belongs to space-time data.
3. The intelligent target behavior prediction method based on multi-feature fusion according to claim 1, wherein in step1, the data preprocessing is as follows: the value ranges of different variables have larger difference, so that the data is normalized by adopting a minimum-maximum normalization method, and the calculation formula of the minimum-maximum normalization is as follows: (14); Wherein, the A value representing the jth feature at the t-th time Represents the normalized value, and N is the number of features.
4. The intelligent target behavior prediction method based on multi-feature fusion according to claim 1, wherein in step 2, the CNN layer is a double-layer 1D-CNN layer, and the CNN layer sequentially includes a one-dimensional convolution layer, a Relu activation layer, a Dropout pooling layer, a one-dimensional convolution layer, a Relu activation layer, and a Dropout pooling layer.
5. The intelligent target behavior prediction method based on multi-feature fusion according to claim 1, wherein in step 3, the periodic feature extraction process is as follows: step 3.1, converting the time sequence data of the Doppler speed into a frequency domain through fast Fourier transform to form a spectrogram of the time sequence; Step 3.2, identifying the frequency corresponding to the first zero peak by interpreting the spectrogram, and then dividing the total number of sampling points M by the number of sampling points K in each period to calculate an estimated period T ', period T' =M/K; Step 3.3, dividing the preprocessed data set by taking the period T' as a unit, and dividing the time sequence of the monthly target movement into a series of active period subsequences to obtain a multi-relation diagram, namely period characteristics; The multiple relationship graph The structure of the multi-relationship graph is represented by an adjacency matrix a i,j , comprising a node set V and an edge set E, when node V kou and node V kou are connected to form an edge, If not, the first part of the first part is connected with the second part, 。
6. The intelligent target behavior prediction method based on multi-feature fusion according to claim 1, wherein in step 5, the process of extracting the temporal feature extraction module is as follows: step 5.1, G hidden states are obtained through an LSTM network, and convolved through a CNN layer to generate a matrix; step 5.2, the output of the CNN layer and the output of the LSTM layer are simultaneously input to the attention layer, and the current hidden state is compared with the matrix through a scoring function in an attention mechanism so as to obtain attention weight; step 5.3, entering a full-connection layer after being output by the attention layer, and outputting a target behavior prediction sequence based on time sequence characteristics and period characteristics by the full-connection layer; The LSTM updates the state at the current moment through a forget gate and a memory gate; discarding information no longer needed through a forget gate The method comprises the following steps: (1); Wherein, sigma represents a sigmoid function, sigma is between [0,1], and the value is closer to 0, the information is less important, the value is closer to 1, and the information is more important; Representing hidden layer state information; And Respectively representing forgetting gate weight and bias; the new state needed to be stored to the current moment is reserved through the memory gate Information of (a) memory gate The method comprises the following steps: (2); (3); wherein the new state at the current time Comprising the state of the previous moment 、 And current time input B i and b c represent the input gate and hidden layer state bias, respectively, W i and W c represent the weights of the input gate and candidate state, respectively, tanh is an activation function, and i t represents the output of the input gate at time t; after updating the state at the current time, the state information at the current time is transferred to the state at the next time through the output gate, as in formulas (4) and (5): (4); (5); Wherein, the Output at time t; B o denotes output gate bias; Is a forget gate, as indicated by element level multiplication; setting k 1D-CNN filters, wherein w is the size of a sliding window, and the convolution operation formula is expressed as follows: (6); Wherein, the Representing the convolution value of the b filter in row a ; Representing the weight of the b filters at time t, Is shown in R represents the index in the window; defining a scoring function Determining correlations between input vectors : (7); Wherein, the Is of line a ; The attention weight The calculation formula is as follows: (8); By aligning Weighting the row vectors of (2) to obtain a vector Vector(s) Representing information related to the current instant t, The method comprises the following steps: (9); θ represents the number of H C line vectors involved in the weight calculation.
7. The intelligent target behavior prediction method based on multi-feature fusion according to claim 1, wherein in step 7, the CNN network is a two-dimensional convolutional neural network; The two-dimensional convolutional neural network sequentially comprises a two-dimensional convolutional layer, a Relu activating layer, a Dropout pooling layer, a two-dimensional convolutional layer, a Relu activating layer, a Dropout pooling layer, a two-dimensional convolutional layer, a Relu activating layer, a Dropout pooling layer, a full connection layer, a Relu activating layer and a Dropout pooling layer.
8. The intelligent target behavior prediction method based on multi-feature fusion according to claim 7, wherein the convolution operation of the two-dimensional convolution layer is as follows: (10); wherein l represents a first layer of the CNN network; Is the beta-th feature map in the first layer, y () is Relu activation functions; Is the first to Layer 1 A beta-th feature map connected with the feature maps; Is the first A convolution kernel parameter of the layer feature map; Is the first ∗ Represents a convolution operation; Representing the index of the input feature elements involved in the computation; the downsampling operation of the Dropout pooling layer is expressed as: (11); Wherein m is the pooling width, and p and q are the input and output of the pooling layer respectively; Is a downsampling function, down () [ D n ] is Is the D th element; g is a parameter; The operation of the fully connected layer is expressed as: (12); Wherein, the Is the ith neuron output; And The weights and biases of the ith neuron, respectively; Is the output of the previous layer.
9. The intelligent target behavior prediction method based on multi-feature fusion according to claim 1, wherein in step 8, the weighted average method assigns different weights to the temporal feature sequence and the spatial feature sequence, in order to find the weight u minimizing the error, the weight vector u is solved by setting the partial derivative of the weight u to zero, the temporal feature sequence and the spatial feature sequence are combined by calculating the weight so that the final prediction best accords with the actual value, the final prediction result is the fusion feature, and the calculation is as follows: (13); Wherein, the Is a fusion feature, is a prediction result of the ith model, The weight of the ith module is that N 'is the number of models, the value of N' is 2, and the weight represents a time feature extraction module and a space feature extraction module.
10. The intelligent target behavior prediction method based on multi-feature fusion according to claim 1, wherein in step 9, the loss function uses a smooth L1 loss function : (15); Wherein x and y represent the output and label respectively, Representing the difference between the output and the tag, using a square error when the difference is less than 1, otherwise using a linear error; The evaluation index comprises a Root Mean Square Error (RMSE) and a decision coefficient ; The root mean square error RMSE is used for measuring the deviation between the observed value and the actual value, and the calculation formula is as follows: (16); n' represents the total length of the output predicted target behavior; the decision coefficient For reflecting the fitting degree of the model to the sample data, expressed as: (17); Wherein, the Is the actual value of the data and, Is the average value of the actual values of the data, Is a predictive value of the data.

Description

Intelligent target behavior prediction method based on multi-feature fusion Technical Field The invention belongs to the technical field of target behavior prediction, and particularly relates to an intelligent target behavior prediction method based on multi-feature fusion. Background Some current intelligent target behavior prediction methods only utilize a single temporal or spatial feature, ignoring the rich features contained in the time series periodicity. In an open environment, the data detected by the sensor on the target is not limited to radar, but also data such as information, and compared with the behavior prediction based on unit data (track information), the problem of low resource utilization rate exists, and the behavior prediction based on multi-element data (longitude, latitude, altitude, speed, acceleration, doppler speed, distance, target type and the like) can more objectively describe the real behavior of the target in a future time period from different aspects. Many studies have been conducted to date on unit data predictions based on traditional mathematical methods, such as autoregressive integrated moving average (ARIMA) models and Hidden Markov Models (HMMs). The ARIMA analyzes the time sequence, predicts the optimal distribution curve rate of the most cities of the Indian active cases, and the improved data mining method of the ARIMA model is more accurate in predicting future stock prices. The HMM can capture complex dynamic modes and non-stationarity of the time sequence, so that the model can adapt to the non-stationarity in prediction. The Support Vector Machine (SVM) model predicts financial data indicating that SVM significantly improves the prediction accuracy of time series. The above method works well in unit data prediction. However, they fail to meet the complex requirements of multivariate time series data prediction, making mining the spatio-temporal dependencies of multivariate data challenging. With the development of deep learning, such as Convolutional Neural Network (CNN), cyclic neural network (RNN), long-short-time memory network (LSTM) and the like, new possibilities are opened up for multi-element time sequence data prediction. The multi-component time sequence data prediction framework based on the RNN has a certain effect, but the RNN is easy to cause gradient explosion or gradient disappearance during learning and long-term sequence data dependency. LSTM in combination with a decomposition algorithm can predict a time series with multiple seasonal periods. The prediction method based on the multilayer LSTM network predicts the demand according to the inherent characteristics of the non-stationary time series data, and shows the superior performance in experiments. Some students have attempted to mix multiple deep neural network models, for example, using a combination of CNNs and RNNs to capture long-term and short-term data of sequences. Because RNNs and variants thereof are less efficient at processing large-scale historical data, some researchers use causal convolution to learn deeper time series features and use time convolution networks for multivariate time series data prediction. To obtain long-term dependencies of time series, students extend the transducer model to multivariate time series data predictions. The prediction model gradually processes data in the time dimension, so that time dependence can be fully explored. However, it is not easy to acquire the spatiotemporal relationship of the multivariate time series data. Aiming at the problems that the data acquired in the current complex battlefield environment is easy to be interfered and deceptively influenced, the movement forms of the targets are various, the targets have large concealment and the like, the behavior characteristics and rules of the targets need to be analyzed, the future movement behaviors of the targets are predicted, and the support is provided for air traffic management, high-value target early warning and the like. However, some existing target behavior prediction algorithms extract only a single temporal or spatial feature, ignoring the problem of rich features contained in the time series periodicity. The patent provides a target behavior prediction model based on multi-feature fusion, which utilizes a time sequence to perform multi-feature fusion on time, space and period. Firstly, a time sequence feature extraction module is constructed to extract time sequence features of multi-element time sequence data, long-term short-time memory (LSTM) with attention mechanism is used to extract time dependence of sequence data, in addition, a dynamic periodic chart is embedded to extract periodicity of time sequence data, then, a space feature extraction module is constructed to extract space features of multi-element time sequence data, a plurality of one-dimensional sequences are converted into a two-dimensional graph structure through data structure conversion, and a C