Class information-containing ML-ELM-AE target motion pattern recognition method

Abstract

The invention discloses an ML-ELM-AE target motion pattern recognition method containing class information, and belongs to the technical field of situation cognition. An ELM-AE model is constructed, and an ML-ELM-AE model is constructed by stacking the ELM-AE model; carrying out the mapping of the motion features of the target through the ML-ELM-AE; a CELM classification model is constructed, and according to a difference vector set formed by inter-class samples, the weight of an input layer of the CELM classification model to a hidden layer and a hidden layer node bias item are optimized, so that the mapping of the samples from a feature space to a class space has regularity, and the precision and generalization ability of target motion pattern classification are improved.

Description

Technical Field

[0001] This invention belongs to the field of situational awareness technology, specifically relating to a method for recognizing target motion patterns using ML-ELM-AE containing class information. Background Technology

[0002] Target motion pattern recognition is one of the core technologies in the field of situational awareness. It usually refers to the automatic classification of target motion patterns based on target motion feature data (such as kinematic parameters such as longitude, latitude, altitude, speed, and acceleration) collected by sensors. Through three steps, namely feature extraction, model training, and pattern matching, it is widely used in fields such as intelligent transportation and security monitoring.

[0003] With the development of deep learning technology in the field of target motion pattern recognition, data-driven target motion pattern recognition methods based on target trajectory sequences have gradually become mainstream. These methods mainly employ models such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Long Short-Term Memory Networks (LSTM) to mine latent features in target time-series data and combine them with classification ideas to achieve online segmentation of target motion patterns. However, these methods all use gradient descent for model training, requiring repeated iterations to adjust parameters to obtain optimal weights. This results in low training efficiency, long iteration cycles, and a tendency for gradient explosion or vanishing phenomena, leading to poor model generalization performance and difficulty in meeting the target motion pattern recognition needs in complex scenarios.

[0004] In recent years, the Extreme Learning Machine (ELM) and Extreme Learning Machine Autoencoder (ELM-AE) have addressed the training efficiency problem of traditional deep learning models. As a single-hidden-layer neural network, it can randomly generate input-hidden layer weights and neuron biases during training. By simply setting the number of hidden layer nodes appropriately, the optimal output weights can be obtained quickly. Compared with models such as CNN and RNN, it has higher learning efficiency and better generalization performance, and is gradually being applied to the fields of target feature extraction and target motion pattern recognition. However, traditional ELM and ELM-AE models still have significant technical shortcomings, mainly in that the hidden layer weights and biases of ELM are randomly generated and do not utilize category information, making it difficult to adapt to complex and ever-changing target motion pattern recognition scenarios.

[0005] The closest existing technical solution to this invention is: the doctoral dissertation "Semantic Mining of Target Activities Driven by Data and Knowledge" mentions a target maneuver classification method with a memory learning mechanism. In the process of "motion semantic memory learning" to perform behavioral semantic perception of trajectory segments, in order to further extract the semantic information of trajectory segments divided by motion patterns, the semantic discrimination problem is transformed into a trajectory classification problem using the kinematic information of the target. That is, firstly, the target motion position and velocity information contained in the trajectory segments are used, and then the semantic categories are divided by motion patterns through a classifier (that is, these trajectory segments are labeled with maneuver tags), and finally the target motion pattern is extracted.

[0006] The shortcomings of this similar technical solution are that it only extracts the basic shallow features of the target motion, without performing layer-by-layer mining of abstract features, and does not utilize data category information to optimize classifier parameters, resulting in insufficient classification accuracy and generalization ability, and failing to meet the needs of target motion pattern recognition in complex scenarios.

[0007] The core requirement of target motion pattern recognition is to balance recognition efficiency and recognition accuracy. Existing technologies (including traditional deep learning models and traditional ELM-EA models) cannot simultaneously meet this core requirement. Their technical shortcomings and causal reasoning process are as follows:

[0008] (1) Traditional deep learning models (CNN, RNN, LSTM) are trained using gradient descent, which requires repeated iterations to adjust parameters to obtain the optimal weights, resulting in slow training speed, long iteration cycle, and easy gradient explosion / gradient vanishing phenomenon, which in turn leads to poor generalization performance of the model and inability to quickly and accurately extract the abstract features of target motion.

[0009] (2) Traditional ELM and ELM-AE models randomly initialize the input-hidden layer weights and bias terms and do not utilize data category information → resulting in scattered sample mapping and poor feature category separability → leading to low target motion pattern recognition accuracy and inability to adapt to complex scene requirements.

[0010] In view of the technical shortcomings of the prior art, the objective of the present invention is as follows:

[0011] (1) Construct the ML-ELM-AE multi-layer feature extraction framework to realize the layer-by-layer mining of the target motion abstract features, improve the model training efficiency, avoid the model getting stuck in local optima, and solve the problems of slow training and poor generalization of traditional deep learning models.

[0012] (2) Introduce target motion pattern category information, optimize the input-hidden layer weights and hidden layer node bias terms of the CELM classifier, and solve the problems of scattered sample mapping and poor feature class separability;

[0013] (3) Ultimately, the efficiency and accuracy of target motion pattern recognition are improved, and a target motion pattern recognition method with both high efficiency and robustness is constructed to meet the application needs of complex scenarios such as intelligent transportation and security monitoring.

[0014] This patent proposes a multilayer extreme learning machine autoencoder (ML-ELM-AE) target motion pattern recognition method with class information. The main technical problems to be solved are as follows: (1) Construct a multilayer extreme learning machine autoencoder (ML-ELM-AE) to extract the abstract feature information of target motion layer by layer, which solves the problems of slow training speed, long iteration cycle and easy gradient explosion / gradient vanishing of the feature extraction method based on traditional deep learning networks (such as convolutional neural network CNN, recurrent neural network RNN, long short-term memory network LSTM), while avoiding the possibility of the model training getting stuck in local optima; (2) Based on the target motion pattern category information, the weights of the input layer to the hidden layer and the bias of the hidden layer nodes of CELM are optimized in a targeted manner, which solves the problems of traditional ELM being unable to utilize category information, scattered sample mapping, and insufficient accuracy of target motion pattern recognition. Summary of the Invention

[0015] To address the shortcomings of existing technologies, this invention proposes a ML-ELM-AE target motion pattern recognition method incorporating class information, belonging to the field of target motion pattern recognition technology. It constructs an Extreme Learning Machine Auto Encoder (ELM-AE) model, and then builds a multi-layer Extreme Learning Machine Auto Encoder (ML-ELM-AE) by stacking ELM-AEs. The ML-ELM-AE is used to map target features. A CELM (Constrained Extreme Learning Machine) classification model is constructed. Based on the difference vector set composed of inter-class samples, the weights of the input layer to the hidden layer and the bias terms of the hidden layer nodes in the CELM classification model are optimized, making the mapping of samples from the feature space to the class space more regular, thereby improving the accuracy and generalization ability of target motion pattern classification.

[0016] A target motion pattern recognition method with class information, such as Figure 1 As shown; the specific steps are as follows:

[0017] Step 1: Collect multivariate time series data of the target's motion process through the data acquisition module to obtain the target motion time series data;

[0018] Step 2: The target motion time series data is preprocessed using the data preprocessing module to obtain a normalized multivariate time series dataset;

[0019] Step 3: Construct and train the ML-ELM-AE feature extraction module;

[0020] Step 4: Extract features from the normalized multivariate time series data using the trained ML-ELM-AE feature extraction module to obtain the target motion abstract feature matrix;

[0021] Step 5: Classify the target motion pattern using the CELM target motion pattern classification optimization module on the target motion abstract feature matrix;

[0022] Step 6: Output the recognition results through the target motion pattern output module.

[0023] Furthermore, in step one, the data acquisition process is as follows:

[0024] The data acquisition module collects multivariate time series data during the target's motion process. ,in, The number of multivariate time series data. Representing the A multivariate time series variable; the multivariate time series variable It contains M (M=7) dimensions; the 7 dimensions are the target's longitude, latitude, altitude, velocity, acceleration, Doppler velocity, and distance between the target and the observation point; Representing the Victor One multivariate time series variable; the sampling frequency is 10Hz to ensure the continuity and integrity of the data;

[0025] After data collection is completed, the multivariate time series dataset will be... Transmitted to the data preprocessing module; multivariate time series dataset The sliding window is dynamically updated, with the length of the sliding window set to a (a≥10, adjusted according to the amount of sample data).

[0026] Furthermore, in step two, the data preprocessing process is as follows:

[0027] Multivariate time series datasets transmitted by the data acquisition module Z-score normalization is performed to avoid the impact of data fluctuations on subsequent feature extraction and classification accuracy.

[0028] The Z-score normalization formula is as follows:

[0029] ;

[0030] In the formula, Represents the normalized ; Representing the All variables The mean; Representing the All variables The variance; It is a minimum value ( This is used to prevent the denominator from being zero and to ensure the stability of normalized calculations;

[0031] right After normalizing the multivariate time series variables, we obtain multivariate time series data. .

[0032] Furthermore, the multivariate time series dataset Generated using simulation tools, containing 100 multivariate time series data points, represented as follows: ;

[0033] After normalization and labeling, the training set for linear motion mode is obtained. , Represents linear motion mode;

[0034] Similarly, after normalization and labeling, training sets for climbing, diving, turning, and U-turn motion modes are obtained respectively. , , and ;

[0035] The linear motion mode This indicates motion in which both the direction of motion and the track angle remain constant, with no horizontal or vertical trajectory deviation; the climb... The diving motion indicates a continuous increase in altitude; The movement indicates a continuous decrease in altitude; the turning... This refers to a movement that changes the direction of flight within a horizontal or inclined plane; the turning around... This indicates a movement that completes a 180° reversal in a short period of time;

[0036] Normalized and labeled training set , , , , The data is then transferred to the ML-ELM-AE feature extraction module.

[0037] Furthermore, in step three, the ML-ELM-AE feature extraction module has a multi-level structure, which is formed by stacking ELM-AE models layer by layer;

[0038] The ML-ELM-AE feature extraction module includes an input layer and a K-layer hidden layer (…). (Adjusted according to actual recognition needs) and output layer;

[0039] The input layer is used to receive preprocessed target motion time series data;

[0040] Each of the hidden layers is set with an activation function to extract the target motion abstract features layer by layer. The output of the previous hidden layer is used as the input of the next hidden layer, and the weights of adjacent hidden layers are connected by a weight matrix.

[0041] The output layer is the output of the Kth hidden layer, used to output the abstract feature matrix after multi-layer extraction.

[0042] Furthermore, the first ELM-AE model includes One input layer node, Hidden layer nodes and 'N output layer nodes, L > N'; the 2nd to the 3rd Each ELM-AE model contains L input layer nodes, L hidden layer nodes, and L output layer nodes.

[0043] Furthermore, in step three, the training process of the ML-ELM-AE feature extraction module is as follows:

[0044] Step 3.1, train the first ELM-AE module:

[0045] First, randomly initialize the orthogonal weight matrix of the first input layer of the ELM-AE model. and orthogonal bias vector Combine the sigmoid activation function of the hidden layer to calculate the output matrix of the first hidden layer. :

[0046] ;

[0047] Where g(·) represents the Sigmoid activation function;

[0048] Next, the weight matrix of the first output layer is calculated. :

[0049] The optimal output weight matrix is ​​obtained by estimating the cost function using regularized least squares. :

[0050] ;

[0051] in, This is a parameter that balances empirical risk and structural risk (C=0.01, adjusted based on training performance), obtained by taking... right If the partial derivatives are zero, then the output weight matrix can be obtained. :

[0052] ;

[0053] Finally, The weight matrices of the input and hidden layers of the first ELM-AE model are used to complete the training of a single-layer ELM-AE model; for transpose;

[0054] Step 3.2, train the second ELM-AE module:

[0055] First, the output matrix of the first hidden layer. As input to the second ELM-AE module, repeat step 3.1 to initialize the second layer orthogonal bias vector. ,Will As the weight matrix between the input layer and the hidden layer of the second layer, calculate the output matrix of the second hidden layer. and weight matrix ;

[0056] Step 3.3, and so on, repeating steps 3.1 to 3.2, to the next step. Output matrix of hidden layer As the first The input to the hidden layer ELM-AE model, initializing the first... Layer orthogonal bias vector , As the first The weight matrices of the input layer and the hidden layer are used to calculate the weight matrix of the first layer. Output matrix of hidden layer The data is then used as an abstract feature matrix and transmitted to the CELM target motion pattern classification and optimization module.

[0057] Furthermore, in step five, the CELM target motion pattern classification optimization module is a CELM classifier, including... One input layer node, Hidden layer nodes and One output layer node;

[0058] The CELM classifier includes an input weight matrix. and bias vector ;

[0059] Furthermore, the weight matrix and bias vector The calculation process is as follows:

[0060] Step 5.1, calculate the weight vector of the hidden layer nodes. and bias ;

[0061] Step 5.1.1: Calculate the weight vector of the first hidden layer node. and bias ;

[0062] Select the first sample from any two motion pattern categories;

[0063] The features of the first sample in each of the two motion pattern categories were calculated using the ML-ELM-AE feature extraction model. and And calculate the difference characteristics. ;

[0064] Standardize the differential features and calculate the corresponding biases; then we have:

[0065] ;

[0066] ;

[0067] Step 5.1.2, repeat step 5.1.1, select the first from any two motion pattern categories. The sample number was used to obtain the 2nd to the 3rd samples. The weight vector of each hidden layer node and bias ,

[0068] Step 5.2, based on the weight vector of the hidden layer nodes obtained in Step 5.1 and bias Construct the input weight matrix of the CELM classifier and bias vector .

[0069] Furthermore, in step five, the process of classifying the target motion pattern from the target motion abstract feature matrix is ​​as follows:

[0070] Step 5-1, calculate the output matrix H of the hidden layer of the CELM classifier:

[0071] Abstract feature matrix The input is fed into the CELM classifier, and combined with the weight matrix W and the bias vector B, the hidden layer output matrix H of the CELM classifier is calculated:

[0072] ;

[0073] Step 5-2: Evaluate the training error using the objective function; the solution with the minimum objective function is the optimal solution.

[0074] The objective function Q is as follows:

[0075] ;

[0076] The weight β connecting the hidden layer and the output layer is solved by minimizing the approximate squared difference, resulting in β:

[0077] ;

[0078] in, yes The Moore-Penrose generalized inverse matrix, Obtained through singular value decomposition; T represents the motion pattern label matrix;

[0079] Step 5-3, motion pattern determination:

[0080] Based on the weight matrix of the output layer of the CELM classifier Calculate the target motion pattern category score for each sample, with the score range being [0,1]. The semantic label with the highest category score is the target motion pattern corresponding to the sample, thus completing the inference of the target motion pattern.

[0081] Furthermore, in step six, the process of outputting the recognition result is as follows:

[0082] The pattern output module outputs the recognition results of the target motion pattern; the output results include the semantic label of the target motion pattern of the sample, the corresponding category score, and the recognition accuracy.

[0083] If the recognition accuracy is less than the preset threshold, return to step three;

[0084] If the recognition accuracy is greater than or equal to the preset threshold, the task ends.

[0085] The technical effects of this invention are as follows:

[0086] By constructing a multi-layer feature extraction model (ML-ELM-AE), a stacked ELM-AE model approach is adopted to achieve layer-by-layer mining of abstract features of target motion patterns. Unlike traditional deep learning models, it eliminates the need for iterative parameter fine-tuning. The optimal weight matrices of the input and hidden layers of the ELM-AE model can be quickly obtained through orthogonal weight initialization and regularized least squares estimation, thus effectively overcoming the shortcomings of slow training speed and long iteration cycles of traditional deep learning models. Furthermore, the convex optimization characteristics of the ELM-AE model prevent the model from getting trapped in local optima, significantly improving the model's learning efficiency and generalization ability; it also enhances the sufficiency of feature extraction.

[0087] Based on the category information of target motion data, the input-hidden layer weights and hidden layer node bias terms of the CELM classifier are optimized through differential feature calculation and standardization. This effectively solves the problems of traditional CELM, which cannot utilize category information and has scattered sample mapping, making the extracted features more class separable and thus improving the classification accuracy of target motion patterns.

[0088] This invention can adapt to the target motion pattern recognition needs of various scenarios such as intelligent transportation and security monitoring, and takes into account recognition efficiency, accuracy and practicality. Attached Figure Description

[0089] Figure 1 This is a schematic diagram of the ML-ELM-AE recognition model structure containing class information. Detailed Implementation

[0090] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0091] A target motion pattern recognition method with class information, such as Figure 1 As shown; the specific steps are as follows:

[0092] Step 1: Collect multivariate time series data of the target's motion process through the data acquisition module to obtain the target motion time series data;

[0093] Step 2: The target motion time series data is preprocessed using the data preprocessing module to obtain a normalized multivariate time series dataset; for sample data, a normalized and labeled sample dataset is obtained through preprocessing.

[0094] Step 3: Construct and train the ML-ELM-AE feature extraction module;

[0095] Step 4: Extract features from the normalized multivariate time series data using the trained ML-ELM-AE feature extraction module to obtain the target motion abstract feature matrix;

[0096] Step 5: Classify the target motion pattern using the CELM target motion pattern classification optimization module on the target motion abstract feature matrix;

[0097] Step 6: Output the recognition results through the target motion pattern output module.

[0098] Furthermore, in step one, the data acquisition process is as follows:

[0099] The data acquisition module collects multivariate time series data during the target's motion process. ,in, The number of multivariate time series data. Representing the A multivariate time series variable; the multivariate time series variable It contains M (M=7) dimensions; the 7 dimensions are the target's longitude, latitude, altitude (i.e., spatial parameters), velocity, acceleration, Doppler velocity (i.e., motion state parameters), and distance between the target and the observation point (i.e., relative position parameters). Representing the Victor One multivariate time series variable; the sampling frequency is 10Hz to ensure the continuity and integrity of the data;

[0100] After data collection is completed, the multivariate time series dataset will be... Transmitted to the data preprocessing module; multivariate time series dataset The sliding window is dynamically updated, with the length of the sliding window set to a (a≥10, adjusted according to the amount of sample data).

[0101] Furthermore, in step two, the data preprocessing process is as follows:

[0102] Multivariate time series datasets transmitted by the data acquisition module Z-score normalization is performed to avoid the impact of data fluctuations on subsequent feature extraction and classification accuracy.

[0103] The Z-score normalization formula is as follows:

[0104] ;

[0105] In the formula, Represents the normalized ; Representing the All variables The mean; Representing the All variables The variance; It is a minimum value ( This is used to prevent the denominator from being zero and to ensure the stability of normalized calculations;

[0106] right After normalizing the multivariate time series variables, we obtain multivariate time series data. .

[0107] Furthermore, the multivariate time series dataset Generated using simulation tools (such as MATLAB), containing 100 multivariate time series data points, represented as follows: ;

[0108] After normalization and labeling, the training set for linear motion mode is obtained. , Represents linear motion mode;

[0109] Similarly, after normalization and labeling, training sets for climbing, diving, turning, and U-turn motion modes are obtained respectively. , , and ;

[0110] The linear motion mode This indicates motion in which both the direction of motion and the track angle remain constant, with no horizontal or vertical trajectory deviation; the climb... The diving motion indicates a continuous increase in altitude; The movement indicates a continuous decrease in altitude; the turning... This refers to a movement that changes the direction of flight within a horizontal or inclined plane; the turning around... This indicates a movement that completes a 180° reversal in a short period of time;

[0111] Normalized and labeled training set , , , , The data is transferred to the ML-ELM-AE feature extraction module for feature extraction and model training.

[0112] Furthermore, in step three, the ML-ELM-AE feature extraction module has a multi-level structure, which is formed by stacking ELM-AE models layer by layer;

[0113] The ML-ELM-AE feature extraction module includes an input layer and a K-layer hidden layer (…). (Adjusted according to actual recognition needs) and output layer;

[0114] The input layer is used to receive preprocessed target motion time series data;

[0115] Each of the hidden layers is set with an activation function to extract the target motion abstract features layer by layer. The output of the previous hidden layer is used as the input of the next hidden layer, and the weights of adjacent hidden layers are connected by a weight matrix.

[0116] The output layer is the output of the Kth hidden layer, used to output the abstract feature matrix after multi-layer extraction.

[0117] Furthermore, the first ELM-AE model includes 'One input layer node,' Hidden layer nodes and ' Output layer node; the 2nd to the 'th' Each ELM-AE model contains Each input layer node Hidden layer nodes and Output layer nodes.

[0118] Furthermore, in step three, the training process of the ML-ELM-AE feature extraction module is as follows:

[0119] Step 3.1, train the first ELM-AE module:

[0120] First, randomly initialize the orthogonal weight matrix of the first input layer of the ELM-AE model. and orthogonal bias vector Combine the sigmoid activation function of the hidden layer to calculate the output matrix of the first hidden layer. :

[0121] ;

[0122] Where g(·) represents the Sigmoid activation function (which can also be replaced by the ReLU function);

[0123] Next, the weight matrix of the first output layer is calculated. :

[0124] The goal of the ELM-AE model is to minimize the reconstruction error of the input samples. The optimal output weight matrix is ​​obtained by estimating the cost function using regularized least squares. :

[0125] ;

[0126] in, This is a parameter that balances empirical risk and structural risk (C=0.01, adjusted based on training performance), obtained by taking... right If the partial derivatives are zero, then the output weight matrix can be obtained. :

[0127] ;

[0128] Finally, The weight matrices of the input and hidden layers of the first ELM-AE model are used to complete the training of a single-layer ELM-AE model; for transpose;

[0129] Step 3.2, train the second ELM-AE module:

[0130] By stacking ELM-AE models layer by layer, the abstract features of target motion are extracted layer by layer. The aim is to solve the problems of insufficient extraction and slow training of traditional feature extraction methods.

[0131] First, the output matrix of the first hidden layer. As input to the second ELM-AE module, repeat step 3.1 to initialize the second layer orthogonal bias vector. ,Will As the weight matrix between the input layer and the hidden layer of the second layer, calculate the output matrix of the second hidden layer. and weight matrix ;

[0132] Step 3.3, and so on, repeating steps 3.1 to 3.2, to the next step. Output matrix of hidden layer As the first The input to the hidden layer ELM-AE model, initializing the first... Layer orthogonal bias vector , As the first The weight matrices of the input layer and the hidden layer are used to calculate the weight matrix of the first layer. Output matrix of hidden layer The data is then used as an abstract feature matrix and transmitted to the CELM target motion pattern classification and optimization module.

[0133] Furthermore, in step five, the CELM target motion pattern classification optimization module is a CELM classifier, including... One input layer node, Hidden layer nodes and One output layer node;

[0134] The CELM classifier includes an input weight matrix. and bias vector ;

[0135] The CELM classifier described above is based on the CELM classifier. According to the difference vector set composed of samples between target motion pattern categories, the weights and bias terms from the input layer to the hidden layer of the CELM classifier are optimized. This results in the optimal data representation with class discrimination, solving the problems of traditional CELM being unable to utilize class information and having scattered sample mapping.

[0136] Furthermore, the weight matrix and bias vector The calculation process is as follows:

[0137] Step 5.1, calculate the weight vector of the hidden layer nodes. and bias ;

[0138] Step 5.1.1: Calculate the weight vector of the first hidden layer node. and bias ;

[0139] From any two motion pattern categories (such as linear motion) and climbing Select the first sample from ) and , Represents a linear motion sample set The first sample in; Represents the sample set of climbing motion The first sample in;

[0140] The features of the first sample in each of the two motion pattern categories were calculated using the ML-ELM-AE feature extraction model. and (Features of the first sample in the linear motion sample set) Features of the first sample in the climbing motion sample set ); and calculate the difference characteristics. ;

[0141] Standardize the differential features and calculate the corresponding biases; then we have:

[0142] ;

[0143] ;

[0144] Step 5.1.2, repeat step 5.1.1, from any two motion pattern categories (such as linear motion) and climbing Select the first from ) The 2nd to 3rd samples were obtained. The weight vector of each hidden layer node and bias ,

[0145] Step 5.2, based on the weight vector of the hidden layer nodes obtained in Step 5.1 and bias Construct the input layer weight matrix of the CELM classifier and bias vector .

[0146] Ensure that the weight matrix and bias vector can reflect the feature differences of different categories of samples, so that the mapping of samples from the original space to the new feature space is regular and avoids the sample mapping being scattered.

[0147] Furthermore, in step five, the process of classifying the target motion pattern from the target motion abstract feature matrix is ​​as follows:

[0148] Step 5-1, calculate the output matrix H of the hidden layer of the CELM classifier:

[0149] Abstract feature matrix The input is fed into the CELM classifier, and combined with the weight matrix W and the bias vector B, the hidden layer output matrix H of the CELM classifier is calculated:

[0150] ;

[0151] Step 5-2: In order to obtain the weight matrix β of the output layer of the CELM classifier that performs well on the training sample set, it is necessary to ensure that its training error is minimized. The objective function is to minimize the squared difference between Hβ and the motion pattern label matrix T corresponding to the training samples. The training error is evaluated through the objective function. The solution with the minimum objective function is the optimal solution.

[0152] The objective function Q is as follows:

[0153] ;

[0154] The weight β connecting the hidden layer and the output layer is solved by minimizing the approximate squared difference, resulting in β:

[0155] ;

[0156] in, yes The Moore-Penrose generalized inverse matrix, Obtained through singular value decomposition; T represents the motion pattern label matrix;

[0157] Step 5-3, motion pattern determination:

[0158] Based on the weight matrix of the output layer of the CELM classifier Calculate the target motion pattern category score for each sample, with the score range being [0,1]. The semantic label with the highest category score is the target motion pattern corresponding to the sample, thus completing the inference of the target motion pattern.

[0159] Furthermore, in step six, the process of outputting the recognition result is as follows:

[0160] The pattern output module outputs the recognition results of the target motion pattern; the output results include the semantic label of the target motion pattern of the sample, the corresponding category score, and the recognition accuracy.

[0161] If the recognition accuracy is less than the preset threshold (the preset threshold is 90%), return to step three;

[0162] If the recognition accuracy is greater than or equal to the preset threshold (the preset threshold is 90%), the task ends.

[0163] The output results facilitate subsequent model optimization and reuse; if the recognition accuracy is lower than the preset threshold (the preset threshold is 90%), return to step 3 (ML-ELM-AE multi-layer feature extraction), readjust the model parameters and repeat the training until the recognition accuracy meets the preset requirements.

Claims

1. A method for recognizing target motion patterns in ML-ELM-AE containing class information, characterized in that, The specific steps are as follows: Step 1: Collect multivariate time series data of the target's motion process through the data acquisition module to obtain the target motion time series data; Step 2: The target motion time series data is preprocessed using the data preprocessing module to obtain a normalized multivariate time series dataset; Step 3: Construct and train the ML-ELM-AE feature extraction module; Step 4: Extract features from the normalized multivariate time series data using the trained ML-ELM-AE feature extraction module to obtain the target motion abstract feature matrix; Step 5: Classify the target motion pattern using the CELM target motion pattern classification optimization module on the target motion abstract feature matrix; Step Six: Output the recognition results through the target motion pattern output module; The output results include the semantic label of the target motion pattern of the sample, the corresponding category score, and the recognition accuracy. If the recognition accuracy is less than the preset threshold, return to step three; If the recognition accuracy is greater than or equal to the preset threshold, the task ends.

2. The ML-ELM-AE target motion pattern recognition method containing class information according to claim 1, characterized in that, In step one, the data acquisition process is as follows: The data acquisition module collects multivariate time series data during the target's motion process. ,in, The number of multivariate time series data. Representing the A multivariate time series variable; the multivariate time series variable It contains M dimensions, M=7; the 7 dimensions are the target's longitude, latitude, altitude, velocity, acceleration, Doppler velocity, and the distance between the target and the observation point; After data collection is completed, the multivariate time series dataset will be... Transmitted to the data preprocessing module; multivariate time series dataset The window is dynamically updated using a sliding window, with the sliding window length set to 'a'.

3. The ML-ELM-AE target motion pattern recognition method containing class information according to claim 1, characterized in that, In step two, the data preprocessing process is as follows: Multivariate time series datasets transmitted by the data acquisition module Z-score normalization is performed to avoid the impact of data fluctuations on subsequent feature extraction and classification accuracy. The Z-score normalization formula is as follows: ； In the formula, Represents the normalized ; Representing the Victor One multivariate time series variable; Representing the All variables The mean; Representing the All variables The variance; It is a local minimum; After normalizing the m-dimensional multivariate time series variables, To obtain multivariate time series data .

4. The ML-ELM-AE target motion pattern recognition method containing class information according to claim 3, characterized in that, The multivariate time series dataset Generated using simulation tools, containing 100 multivariate time series data points, represented as follows: ; After normalization and labeling, the training set for linear motion mode is obtained. , Represents linear motion mode; Similarly, after normalization and labeling, training sets for climbing, diving, turning, and U-turn motion modes are obtained respectively. , , and ; The linear motion mode This indicates motion in which both the direction of motion and the track angle remain constant, with no horizontal or vertical trajectory deviation; the climb... The diving motion indicates a continuous increase in altitude; This indicates a motion with a continuously decreasing altitude; the turning... This refers to a movement that changes the direction of flight within a horizontal or inclined plane; the turning around... This indicates a movement that completes a 180° reversal in a short period of time; Normalized and labeled training set , , , , The data is then transferred to the ML-ELM-AE feature extraction module.

5. The ML-ELM-AE target motion pattern recognition method containing class information according to claim 1, characterized in that, In step three, the ML-ELM-AE feature extraction module has a multi-level structure, which is formed by stacking ELM-AE models layer by layer; The ML-ELM-AE feature extraction module includes an input layer, a K-layer hidden layer, and an output layer. ; The input layer is used to receive preprocessed target motion time series data; Each of the hidden layers is set with an activation function to extract the abstract features of target motion layer by layer. The output of the previous hidden layer is used as the input of the next hidden layer, and the weights of adjacent hidden layers are connected by a weight matrix. The output layer is the output of the Kth hidden layer, used to output the abstract feature matrix after multi-layer extraction.

6. The ML-ELM-AE target motion pattern recognition method containing class information according to claim 5, characterized in that, The first ELM-AE model of the ML-ELM-AE feature extraction module includes 'One input layer node,' Hidden layer nodes and 'N output layer nodes, L>N'; the 2nd to kth ELM-AE models each contain L input layer nodes, L hidden layer nodes, and L output layer nodes.

7. The ML-ELM-AE target motion pattern recognition method containing class information according to claim 1, characterized in that, In step three, the training process of the ML-ELM-AE feature extraction module is as follows: Step 3.1, train the first ELM-AE module: First, randomly initialize the orthogonal weight matrix of the first input layer of the ELM-AE model. and orthogonal bias vector Combine the sigmoid activation function of the hidden layer to calculate the output matrix of the first hidden layer. : ； Where g(·) represents the Sigmoid activation function; Next, the weight matrix of the first output layer is calculated. : The optimal output weight matrix is ​​obtained by estimating the cost function using regularized least squares. : ； in, It is a parameter that balances empirical risk and structural risk, by taking... right If the partial derivatives are zero, then the output weight matrix can be obtained. : ； Finally, The weight matrices of the input and hidden layers of the first ELM-AE model are used to complete the training of a single-layer ELM-AE model; for Transpose of; Step 3.2, train the second ELM-AE module: First, the output matrix of the first hidden layer. As input to the second ELM-AE module, repeat step 3.1 to initialize the second layer orthogonal bias vector. ,Will As the weight matrix between the input layer and the hidden layer of the second layer, calculate the output matrix of the second hidden layer. and weight matrix ; Step 3.3, and so on, repeating steps 3.1 to 3.2, to the next step. Output matrix of hidden layer As the first The input to the hidden layer ELM-AE model, initializing the first... Layer orthogonal bias vector , As the first The weight matrices of the input layer and the hidden layer are used to calculate the weight matrix of the first layer. Output matrix of hidden layer The data is then used as an abstract feature matrix and transmitted to the CELM target motion pattern classification and optimization module.

8. The ML-ELM-AE target motion pattern recognition method containing class information according to claim 1, characterized in that, In step five, the CELM target motion pattern classification optimization module is a CELM classifier, including... One input layer node, Hidden layer nodes and One output layer node; The CELM classifier includes an input weight matrix. and bias vector .

9. The ML-ELM-AE target motion pattern recognition method containing class information according to claim 8, characterized in that, The weight matrix and bias vector The calculation process is as follows: Step 5.1, calculate the weight vector of the hidden layer nodes. and bias ; Step 5.1.1, calculate the weight vector of the first hidden layer node. and bias ; Select the first sample from any two motion pattern categories; The features of the first sample in each of the two motion pattern categories were calculated using the ML-ELM-AE feature extraction model. and And calculate the difference characteristics. ; Standardize the differential features and calculate the corresponding biases; then we have: ； ； Step 5.1.2, repeat step 5.1.1, select the first from any two motion pattern categories. The sample number was used to obtain the 2nd to the 3rd samples. The weight vector of each hidden layer node and bias , Step 5.2, based on the weight vector of the hidden layer nodes obtained in Step 5.1 and bias Construct the input weight matrix of the CELM classifier and bias vector .

10. The ML-ELM-AE target motion pattern recognition method containing class information according to claim 1, characterized in that, In step five, the process of classifying the target motion pattern from the target motion abstract feature matrix is ​​as follows: Step 5-1, calculate the output matrix H of the hidden layer of the CELM classifier: Abstract feature matrix The input is fed into the CELM classifier, and combined with the weight matrix W and the bias vector B, the hidden layer output matrix H of the CELM classifier is calculated: ； Step 5-2: Evaluate the training error using the objective function; the solution with the minimum objective function is the optimal solution. The objective function Q is as follows: ； The weight β connecting the hidden layer and the output layer is solved by minimizing the approximate squared difference, resulting in β: ； in, yes The Moore-Penrose generalized inverse matrix, Obtained through singular value decomposition; T represents the motion pattern label matrix; Step 5-3, motion pattern determination: Based on the weight matrix of the output layer of the CELM classifier Calculate the target motion pattern category score for each sample, with the score range being [0,1]. The semantic label with the highest category score is the target motion pattern corresponding to the sample, thus completing the inference of the target motion pattern.

Images