An intelligent perception and recognition method for electromagnetic targets based on multi-feature heterogeneous fusion

Abstract

The present invention provides an electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion, which first obtains electromagnetic target signal source data information through a sensor, then pre-processes the data through heterogeneous data extraction and multi-feature fusion, constructs a target template through data samples to establish a target knowledge base, and then uses electromagnetic target perception and recognition based on machine learning strategies to complete radiation source recognition, and the recognition results obtained through the transfer learning strategy are stored in the target knowledge base, and the results obtained through the online learning strategy are used to update the knowledge base. The method disclosed by the present invention improves the recognition ability of electromagnetic targets by more accurately describing the characteristics of electromagnetic targets, and combines online learning, transfer learning and convolutional neural networks to improve the speed of electromagnetic target recognition and the robustness of the system to scene and perception device changes.

Description

Technical Field

[0001] The present invention belongs to the field of target recognition, and in particular relates to an electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion. Background Art

[0002] We are in an invisible electromagnetic environment all the time. By extracting and identifying the collected electromagnetic target signals, we can obtain the type, specific location and related attributes of the corresponding electromagnetic target, which is of great significance for studying the construction of complex electromagnetic environments and the identification of electromagnetic targets. It also helps researchers better understand and grasp the complex electromagnetic situation and make decisions more scientifically and quickly.

[0003] However, with the popularization of electrical and electronic equipment, the electromagnetic environment has become increasingly complex and changeable, and researchers need to understand more and more information. This has put forward higher requirements on the recognition ability and speed of electromagnetic signals. Traditional electromagnetic target recognition methods can no longer accurately identify electromagnetic targets and are time-consuming, making it difficult to adapt to the complex and changeable electromagnetic environment.

[0004] With the development and progress of machine learning technology, machine learning technology combined with feature extraction has been increasingly widely used in the field of intelligent perception and recognition of electromagnetic targets. However, the existing electromagnetic target perception and recognition models mostly use a single feature for target recognition, which often makes it difficult to achieve accurate recognition of electromagnetic targets. In addition, given that electromagnetic targets often switch frequently in complex environments, this puts forward new requirements for the recognition of electromagnetic targets. The perception and recognition model should not only be able to identify electromagnetic targets, but also be adaptable and able to cope with different scenarios and changes in perception devices. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides an electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion. In order to make full use of the heterogeneous features of electromagnetic targets, the multi-attribute features of electromagnetic targets are fused to improve the recognition rate of electromagnetic targets. In addition, online learning and convolutional neural networks based on transfer learning are combined and applied to the field of electromagnetic target intelligent perception and recognition in complex environments, which improves the speed of electromagnetic target recognition and the robustness of the system to changes in scenes and perception devices.

[0006] The object of the present invention is achieved in that the steps include:

[0007] Step 1: electromagnetic target signal source data acquisition, using acquisition equipment to obtain electromagnetic target signal source data information;

[0008] Step 2: Data preprocessing: heterogeneous data extraction and multi-feature fusion are used to preprocess the data;

[0009] Step 3: Establish a target knowledge base, and construct a target template through data samples to establish the target knowledge base;

[0010] Step 4: The convolutional neural network electromagnetic target perception and recognition method based on transfer learning stores the recognition results obtained by the transfer learning strategy into the target knowledge base established in step 3;

[0011] Step 5: The electromagnetic target perception and recognition method based on online learning updates the target knowledge base established in step 3 with the recognition results obtained by the online learning strategy.

[0012] Furthermore, the acquisition device in step 1 is an omnidirectional antenna.

[0013] Furthermore, the heterogeneous data in step 2 specifically include eight characteristic parameters, namely, maximum value, mean, standard deviation, standard mean deviation, coefficient of variation, standard skewness coefficient, standard kurtosis coefficient and linear correlation coefficient.

[0014] Furthermore, the method of multi-feature fusion in step 2 is as follows: adaptive linear weighting is performed on the above eight heterogeneous features, and the fused features are where y n and ω n They represent the feature vector and weight corresponding to the nth feature respectively. The weight represents the contribution of the feature in the recognition process. The sum of the weights of the eight features is 1. The weights corresponding to the eight features automatically obtain the optimal value through iteration during the training process.

[0015] Furthermore, the target knowledge base in step 3 is established by constructing a target template using eight features of the electromagnetic target sample and the category of the electromagnetic target to which it belongs.

[0016] Furthermore, the transfer learning strategy in step 4 is as follows: considering the large amount of training sample data in electromagnetic target feature extraction and the poor adaptability to different scenarios, transfer learning is introduced to adjust the pre-trained model. First, the structure of the convolutional layer in the pre-trained model is not changed, and the locked weights and parameters are loaded into the convolutional layer. A new fully connected layer is designed for the new training task to replace the original fully connected layer. Finally, new electromagnetic data is input, the convolutional layer is frozen, the new fully connected layer is trained, and the new electromagnetic data is classified.

[0017] Furthermore, the online learning strategy in step 5 is as follows:

[0018] First, the training sample data of each type of electromagnetic target is preprocessed by alignment and normalization, and then divided into frames. Then, the template corresponding to each frame is constructed through the update formula of the mean and variance. When new electromagnetic training data is added, the template is updated or a new template is constructed by judging whether the corresponding frame template exists. The update formula of the mean and variance is:

[0019]

[0020]

[0021] in, represents the mean value of the template corresponding to the nth frame of the gth type electromagnetic target at time t-1, represents the variance of the template corresponding to the nth frame of the gth type electromagnetic target at time t-1, represents the total number of training data of the nth frame of the gth type electromagnetic target at time t-1, represents the total number of samples of the nth frame of the gth type electromagnetic target at time t, Σ Yt represents the variance of the newly added electromagnetic training data, Y i The nth frame of newly added data for the gth type electromagnetic target at time t after alignment and normalization.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows: the patent of the present invention provides an electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion, which combines multi-feature heterogeneous extraction technology with intelligent decision-making separation method, avoiding the subjectivity of weighting the heterogeneous features of electromagnetic targets according to their importance during the classification process. The application of multi-feature heterogeneous extraction technology maximizes the accuracy of electromagnetic target intelligent perception. In addition, the electromagnetic target intelligent perception and recognition method based on convolutional neural network based on transfer learning and online learning is adaptive and robust, can be used according to different scenarios, and requires fewer samples. Compared with traditional recognition algorithms, the training speed and recognition accuracy can better meet the needs of electromagnetic target perception and recognition in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is an overall structural diagram of an electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion of the present invention;

[0024] Figure 2 This is a flow chart of intelligent perception and recognition of electromagnetic targets in complex environments of the present invention;

[0025] Figure 3 Online update and expansion of flow chart for the template of the present invention;

[0026] Figure 1This is an overall structural diagram of an electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion according to the present invention. Figure 2 against Figure 1 The specific process of intelligent perception and identification of electromagnetic targets in complex environments is explained in detail. Figure 3 against Figure 1 The specific process of online updating and expansion of templates in the electromagnetic target perception and recognition method based on online learning is described in detail. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0028] The present invention provides an electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion, such as Figure 2 As shown, the specific steps are as follows:

[0029] Step 1: electromagnetic target signal source data acquisition: using acquisition equipment to obtain electromagnetic target signal source data information;

[0030] The electromagnetic target data is collected using omnidirectional antennas.

[0031] Step 2: Data preprocessing: Heterogeneous feature extraction and multi-feature fusion are used to preprocess the data;

[0032] Before identifying the electromagnetic target, the heterogeneous features of the electromagnetic target are first extracted. At the same time, multi-feature information fusion technology is introduced to fuse different features of the same electromagnetic target, which can obtain more comprehensive feature information and help improve the effect of subsequent identification.

[0033] 1) Heterogeneous feature extraction

[0034] By comprehensively utilizing target attributes of different dimensions such as electromagnetic target position characteristic parameters, dispersion characteristic parameters, distribution characteristic parameters and correlation characteristic parameters, heterogeneous features are established and extracted, including maximum value, mean, standard deviation, standard mean deviation, coefficient of variation, standard skewness coefficient, standard kurtosis coefficient and linear correlation coefficient. The eight characteristic quantities are defined as follows:

[0035] Maximum value MAX=max(x n )

[0036] Mean

[0037] Standard Deviation

[0038] Standard mean deviation

[0039] Coefficient of variation

[0040] Standard skewness coefficient

[0041] Standardized Kurtosis

[0042] Linear correlation coefficient

[0043] 2) Multi-feature fusion

[0044] The heterogeneous features of the samples represent different attributes of the samples. Therefore, when fusion of multiple features is performed, the multiple features of the samples cannot be directly spliced ​​together. In order to make full use of the above eight features, the eight features are adaptively linearly weighted to achieve feature fusion while maintaining the dimension. The fused features are where y n and ω n They represent the feature vector and weight corresponding to the nth feature respectively. The weight represents the contribution of the feature in the recognition process. The sum of the weights of the eight features is 1. The weights corresponding to the eight features automatically obtain the optimal value through iteration during the training process.

[0045] Step 3: Establish a target knowledge base: construct a target template through data samples (eight features of electromagnetic targets and the categories of electromagnetic targets to which they belong) to establish a target knowledge base. The recognition results obtained by the subsequent transfer learning strategy are stored in the target knowledge base, and the results obtained by the online learning strategy are used to update the knowledge base.

[0046] Step 4: Convolutional neural network electromagnetic target perception and recognition method based on transfer learning

[0047] 1) Pre-trained Model

[0048] The pre-trained model contains 6 convolutional layers, 6 pooling layers and 2 fully connected layers, and the activation function is ReLU. The back propagation based on the gradient descent algorithm is used in the training process, and the loss function uses the weighted cross entropy loss function.

[0049] The optimal weights were obtained during the training process. The fused feature F can be calculated, and F is input into the Softmax function classifier through two fully connected layers. When there are G categories of electromagnetic targets, a G-dimensional vector is output, which is the probability of each category to which the electromagnetic target sample belongs:

[0050]

[0051] Among them, p i =P(y=i|x),i=1,2,…G,O i (l) is the output value of the i-th node, and the category with the highest probability value is selected as the classification result of the electromagnetic target.

[0052] 2) Transfer learning strategy

[0053] Taking into account the large amount of training sample data in electromagnetic target feature extraction and the poor adaptability to different scenarios, transfer learning is introduced to adjust the pre-trained model. First, the structure of the convolutional layer in the pre-trained model is not changed, and the locked weights and parameters are loaded into the convolutional layer. A new fully connected layer is designed for the new training task to replace the original fully connected layer. Finally, new sample data is input, the convolutional layer is frozen, the new fully connected layer is trained, and the new sample data is classified.

[0054] The new pre-training model only needs a small number of training samples and a little training to obtain new electromagnetic target classification and recognition results, and the recognition results obtained through the transfer learning strategy are stored in the target knowledge base.

[0055] Step 5: Electromagnetic target perception and recognition method based on online learning;

[0056] In fact, new electromagnetic training data will be continuously added when identifying electromagnetic targets, which requires real-time classification of electromagnetic targets. Through online updating and expansion, it is only necessary to use the newly added electromagnetic target data to update the previously learned model parameters, without having to re-learn all the training data.

[0057] The online update and expansion method of the online intelligent recognition algorithm is described below.

[0058] Assume that the mean value of the template corresponding to the nth frame of the gth type electromagnetic target at time t-1 is With variance They are:

[0059]

[0060]

[0061] in, Represents the total number of training data of the nth frame of the gth type electromagnetic target at time t-1.

[0062] Y i t is the newly added electromagnetic data of the nth frame of the gth type electromagnetic target at time t. First, the newly added electromagnetic data is preprocessed to obtain Y i , then Y i Bring it into the template's update formula, the update formula is as follows:

[0063]

[0064]

[0065] in, represents the total number of samples of the nth frame of the gth type electromagnetic target at time t, Represents the variance of the newly added electromagnetic training data.

[0066] If the template of the nth frame of the gth electromagnetic target does not exist, the template of this type is expanded, that is, the newly added data Y t Build the template for that frame by updating the formula.

[0067] The specific steps of the online electromagnetic target recognition method are as follows:

[0068] (1) Perform preprocessing of alignment and normalization on the training sample data of each type of electromagnetic target, and perform frame division;

[0069] (2) Construct the template corresponding to each frame through the update formula of mean and variance;

[0070] (3) When new electromagnetic training data appears, the new electromagnetic data is first divided into frames, and then the template is updated or a new template is constructed by judging whether the corresponding frame template exists. The specific process of this step is as follows: Figure 3 shown.

[0071] (4) After alignment and normalization preprocessing, the test samples are input into the adaptive Gaussian classifier. The discriminant formula is as follows:

[0072]

[0073] Among them, x is the test sample, D g,n is the Mahalanobis distance between the test sample and each frame template.

[0074] Determination D g,n When the minimum is taken, the category of the corresponding frame template is the category to which the test sample x belongs, and then the results obtained through the online learning strategy are used to update the knowledge base.

[0075] The above has completely and clearly described the specific implementation method of the electromagnetic target intelligent perception and identification method based on multi-feature heterogeneous fusion proposed by the present invention. It should be pointed out that any replacement and improvement made by a general technician in this technical field under the enlightenment of the present invention should be covered within the scope of protection of the present invention.

[0076] In summary, the present invention discloses an intelligent perception and recognition method for electromagnetic targets based on multi-feature heterogeneous fusion. The method first obtains electromagnetic target signal source data information through sensors, then pre-processes the data through heterogeneous data extraction and multi-feature fusion, constructs target templates through data samples to establish a target knowledge base, and then uses electromagnetic target perception and recognition based on machine learning strategies to complete radiation source recognition. The recognition results obtained through the transfer learning strategy are stored in the target knowledge base, and the results obtained through the online learning strategy are used to update the knowledge base. The method disclosed in the present invention improves the recognition ability of electromagnetic targets by more accurately describing the characteristics of electromagnetic targets, and combines online learning, transfer learning and convolutional neural networks to improve the speed of electromagnetic target recognition and the robustness of the system to changes in scenes and perception devices.

Claims

1. An electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion, characterized in that: The following steps are involved: Step 1: Collect electromagnetic target signal source data information; Step 2: preprocessing the collected electromagnetic target signal source data, extracting the heterogeneous features of the electromagnetic target, and adaptively linearly weighting different heterogeneous features of the same electromagnetic target to achieve multi-feature fusion while maintaining the dimension; the heterogeneous features are different attributes of the electromagnetic target; Step 3: Taking the heterogeneous features of the electromagnetic target and the category of the electromagnetic target to which it belongs as sample data, constructing a target template through the data samples to establish a target knowledge base; Step 4: Pre-train the convolutional neural network model, adjust the pre-trained convolutional neural network model based on the transfer learning strategy, and store the recognition results obtained by the transfer learning strategy into the target knowledge base established in step 3; Step 5: The electromagnetic target perception and recognition method based on online learning updates the target knowledge base with the recognition results obtained by the online learning strategy.

2. According to claim 1, the electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion is characterized in that: The acquisition device in step 1 is an omnidirectional antenna.

3. According to claim 1, the electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion is characterized by: The heterogeneous features in step 2 are specifically eight heterogeneous feature parameters: maximum value, mean, standard deviation, standard mean deviation, coefficient of variation, standard skewness coefficient, standard kurtosis coefficient and linear correlation coefficient.

4. The electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion according to claim 3 is characterized by: In step 2, eight heterogeneous features are adaptively linearly weighted, and the fused features are: where y n and ω n They represent the feature vector and weight corresponding to the nth feature respectively. The weight represents the contribution of the feature in the recognition process. The sum of the weights of the eight features is 1. The weights corresponding to the eight features automatically obtain the optimal value through iteration during the training process.

5. The electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion according to claim 3 is characterized by: In the step 3, a corresponding target knowledge base is constructed by constructing a target template based on eight features of the electromagnetic target sample and the category of the electromagnetic target to which it belongs.

6. The electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion according to claim 1 is characterized by: The transfer learning strategy in step 4 is: first, the structure of the convolutional layer in the pre-trained model is not changed, the locked weights and parameters are loaded into the convolutional layer, and a new fully connected layer is designed for the new training task to replace the original fully connected layer; finally, new electromagnetic data is input, the convolutional layer is frozen, the new fully connected layer is trained, and the new electromagnetic data is classified.

7. The electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion according to claim 1 is characterized by: The online learning strategy in step 5 is as follows: first, the training sample data of each type of electromagnetic target is aligned and normalized for preprocessing, and then divided into frames. Then, the template corresponding to each frame is constructed through the update formula of the mean and variance. When new electromagnetic training data is added, the template is updated or a new template is constructed by judging whether the corresponding frame template exists.

8. The electromagnetic target intelligent perception and recognition method based on multi-feature heterogeneous fusion according to claim 7 is characterized by: The update formulas for the mean and variance are: in: represents the mean value of the template corresponding to the nth frame of the gth type electromagnetic target at time t-1, represents the variance of the template corresponding to the nth frame of the gth type electromagnetic target at time t-1, represents the total number of training data of the nth frame of the gth type electromagnetic target at time t-1, represents the total number of samples of the nth frame of the gth type electromagnetic target at time t, represents the variance of the newly added electromagnetic training data, Y i The nth frame of newly added data for the gth type electromagnetic target at time t after alignment and normalization.

Images