Target characteristic real-time prediction and dynamic correction method based on deep learning

Abstract

The invention discloses a target characteristic real-time prediction and dynamic correction method based on deep learning, and belongs to the technical field of target recognition. The method mainly comprises the following steps: establishing a neural network model with a CNN-RNN structure; the radar reflection cross section and infrared radiation characteristic data of the target are cleaned, noise, missing values and abnormal values in the data are processed, and the data quality is ensured. Making a data set for training a Darknet53 convolutional neural network model, and obtaining shallow feature information through a feature extraction layer; performing feature fusion on the last layer of input of the network and the Soft-SPP module; extracting features of the fused information through a convolutional layer again; feature information obtained by the anchor frame obtained through preprocessing is used for a feature map output by the network; and inputting the data into an LSTM recurrent neural network in sequence, serializing feature information, and training an LSTM model to realize real-time prediction and analysis of a target radar reflection cross section and infrared radiation characteristics.

Description

Technical Field

[0001] The present invention belongs to the technical field of target recognition, and in particular relates to a method for real-time prediction and dynamic correction of target characteristics based on deep learning. Background Art

[0002] Object detection, also known as object recognition, is a classic problem in the field of computer vision. Its purpose is to obtain the two-dimensional bounding box and category of the object of interest from the input RGB image. Its accuracy and real-time performance are important features of the entire system.

[0003] In the modern military and security fields, accurate prediction and real-time correction of target characteristics are crucial. Existing simulation methods usually focus on a single physical phenomenon, while in practical applications, target characteristics are affected by multiple physical fields (such as electromagnetic fields, temperature fields, and radiation fields). Therefore, a method that comprehensively considers the interaction of multiple physical fields is needed to achieve more accurate target characteristic simulation. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention provides a method for real-time prediction and dynamic correction of target characteristics based on deep learning, so as to achieve efficient, accurate and real-time prediction of target characteristics. A comprehensive description of physical phenomena is provided by multi-physical field coupling processing technology, a deep learning prediction algorithm realizes rapid prediction and analysis, and a dynamic correction mechanism ensures the accuracy and reliability of its results. The present invention can be widely used in the fields of target detection, tracking and identification, stealth technology, anti-interference testing and evaluation of infrared imaging, etc. The present invention relates to the fields of electromagnetic simulation, infrared simulation, computational fluid dynamics (CFD) and artificial intelligence, and in particular refers to a method that can combine multi-physical field coupling processing technology and deep learning algorithm to perform real-time prediction and dynamic correction of electromagnetic scattering characteristics and infrared radiation characteristics of a target.

[0005] The first aspect of the present invention discloses a method for real-time prediction and dynamic correction of target characteristics based on deep learning, the method comprising:

[0006] Step 1: Collect historical simulation data and experimental data, clean the data, process the noise, missing values ​​and outliers, and generate a data set for network training;

[0007] Step 2: Preprocess the data set to obtain N anchor boxes, cluster the bounding box sizes in the anchor box labels using the K-means algorithm, and extract features of the radar reflection cross section and infrared radiation characteristics in the data set using the DarkNet53 network;

[0008] Step 3: Through the Soft-SPP module based on feature fusion, the features extracted by the DarkNet53 network are input into different pooling layers of the pooling core respectively. After the pooling layer uses the SoftPool method for pooling, the features of different feature channels are connected in series;

[0009] Step 4: Pass the concatenated feature information through a 1×1 and a 3×3 convolution layer to further extract the target features;

[0010] Step 5: The N / 3 anchor frames obtained by preprocessing the data are used as references for prediction of the final extracted feature image, and the construction of the convolutional neural network consisting of the DarkNet53 network, the Soft-SPP module, and the convolutional layer is completed;

[0011] Step 6: Input the characteristic information of the electromagnetic scattering characteristics and infrared radiation characteristics of the target predicted by the convolutional neural network into the LSTM recurrent neural network for further processing and network model training to obtain a neural network model with a CNN-RNN structure after training;

[0012] Step 7: Input the experimental data to be predicted and analyzed, including the electromagnetic scattering characteristics and infrared radiation characteristics of the target, into the trained network model for real-time prediction and analysis.

[0013] According to the method of the first aspect of the present invention, the method further includes constructing a dynamic correction mechanism; specifically including:

[0014] Set the initial parameters of the constructed neural network model with CNN-RNN structure;

[0015] The constructed neural network model and initial parameters are used for simulation, and the error between the predicted result and the real data of the target characteristic actually collected is calculated using the mean-split error (MSE);

[0016] Analyze the distribution and characteristics of errors, use the gradient descent method to adjust the neural network model parameters, and update the neural network model;

[0017] Repeat the above steps until the stopping condition is met, and use an independent validation set to validate the modified neural network model;

[0018] During the application of the neural network model, new experimental data are collected and fed back into the correction mechanism to achieve continuous optimization of the neural network model.

[0019] According to the method of the first aspect of the present invention, in step 1, the data is cleaned; specifically, the data is smoothed using a sliding average and a median filter to reduce noise; the missing values ​​are filled in based on similarity using a K-nearest neighbor algorithm; and outliers are detected and filtered based on the statistical 3σ principle.

[0020] According to the method of the first aspect of the present invention, in step 3, based on the Soft-SPP module of feature fusion, SoftPool is used instead of MaxPool for pooling; wherein:

[0021]

[0022] Among them, a i is the pixel point in the neighborhood R, w i is the weight of the i-th pixel, and j is the pixel number in the neighborhood;

[0023] The output value of the pooling is calculated using the weight w i and the corresponding pixel point a i We weight and sum them to get:

[0024]

[0025] in, Represents the output value after pooling, and ﹡ represents the dot product.

[0026] According to the method of the first aspect of the present invention, in step 6: the output features of the convolutional neural network are serialized so that the input data of the LSTM network is consistent, and through sequence modeling and long-term dependencies of memory units, the time series information of electromagnetic scattering characteristics and infrared radiation characteristics is further analyzed and extracted to capture the changes and behavior patterns of the target in a dynamic environment.

[0027] According to the method of the first aspect of the present invention, an attention mechanism is added to the neural network, and the back propagation algorithm and adaptive learning rate adjustment strategy are used to improve the recognition and prediction accuracy of target characteristics to adapt to different scenarios and dynamic changes in target characteristics.

[0028] A second aspect of the present invention discloses a device for real-time prediction and dynamic correction of target characteristics based on deep learning, the device comprising a processing unit, the processing unit being configured to execute:

[0029] Step 1: Collect historical simulation data and experimental data, clean the data, process the noise, missing values ​​and outliers, and generate a data set for network training;

[0030] Step 2: Preprocess the data set to obtain N anchor boxes, cluster the bounding box sizes in the anchor box labels using the K-means algorithm, and extract features of the radar reflection cross section and infrared radiation characteristics in the data set using the DarkNet53 network;

[0031] Step 3: Through the Soft-SPP module based on feature fusion, the features extracted by the DarkNet53 network are input into different pooling layers of the pooling core respectively. After the pooling layer uses the SoftPool method for pooling, the features of different feature channels are connected in series;

[0032] Step 4: Pass the concatenated feature information through a 1×1 and a 3×3 convolution layer to further extract the target features;

[0033] Step 5: The N / 3 anchor frames obtained by preprocessing the data are used as references for prediction of the final extracted feature image, and the construction of the convolutional neural network consisting of the DarkNet53 network, the Soft-SPP module, and the convolutional layer is completed;

[0034] Step 6: Input the characteristic information of the electromagnetic scattering characteristics and infrared radiation characteristics of the target predicted by the convolutional neural network into the LSTM recurrent neural network for further processing and network model training to obtain a neural network model with a CNN-RNN structure after training;

[0035] Step 7: Input the experimental data to be predicted and analyzed, including the electromagnetic scattering characteristics and infrared radiation characteristics of the target, into the trained network model for real-time prediction and analysis.

[0036] According to the apparatus of the second aspect of the present invention, the processing unit is further configured to execute:

[0037] Set the initial parameters of the constructed neural network model with CNN-RNN structure;

[0038] The constructed neural network model and initial parameters are used for simulation, and the error between the predicted result and the real data of the target characteristic actually collected is calculated using the mean-split error (MSE);

[0039] Analyze the distribution and characteristics of errors, use the gradient descent method to adjust the neural network model parameters, and update the neural network model;

[0040] Repeat the above steps until the stopping condition is met, and use an independent validation set to validate the modified neural network model;

[0041] During the application of the neural network model, new experimental data are collected and fed back into the correction mechanism to achieve continuous optimization of the neural network model.

[0042] The third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the method for real-time prediction and dynamic correction of target characteristics based on deep learning described in the first aspect of the present disclosure is implemented.

[0043] The fourth aspect of the present invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for real-time prediction and dynamic correction of target characteristics based on deep learning described in the first aspect of the present disclosure is implemented.

[0044] In summary, the present invention establishes a neural network model with a CNN-RNN structure; the radar cross section (RCS) and infrared radiation characteristic data of the target are cleaned, and the noise, missing values ​​and outliers in the data are processed to ensure data quality. The present invention makes a data set for Darknet53 convolutional neural network model training, obtains shallow feature information through the feature extraction layer; the last layer of the network input is feature fused with the Soft-SPP module; the fused information is again extracted through the convolution layer; the feature map output by the network uses the feature information obtained by the anchor frame obtained by preprocessing; it is sequentially input into the LSTM recurrent neural network, the feature information is serialized, and the LSTM model is trained to realize real-time prediction and analysis of the target radar cross section and infrared radiation characteristics; a dynamic correction mechanism is designed, and the target characteristic model is corrected online in combination with real-time prediction data and experimental data to improve the prediction accuracy.

[0045] The present invention comprehensively considers the interaction among electromagnetic fields, temperature fields and infrared radiation fields, provides more comprehensive prediction results, and adopts deep learning algorithms to realize real-time prediction and precise correction of target characteristics. The system can continuously learn and optimize according to new test data, thus ensuring the efficiency and accuracy of prediction, and is suitable for complex real-time application requirements.

[0046] The beneficial technical effects brought about by the present invention include: the present invention comprehensively considers the interaction between electromagnetic fields, temperature fields and infrared radiation fields, provides more comprehensive prediction results, adopts deep learning algorithms, realizes real-time prediction and precise correction of target characteristics, and the system can continuously learn and optimize according to new test data, ensuring the efficiency and accuracy of the prediction, and is suitable for complex real-time application requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0048] Figure 1 A flowchart of a method for real-time prediction of target characteristics based on deep learning;

[0049] Figure 2A flowchart of a method for dynamic correction of target characteristics based on deep learning;

[0050] Figure 3 This is the DarkNet53 backbone network structure diagram;

[0051] Figure 4 This is the structural diagram of the Soft-SPP module;

[0052] Figure 5 This is the LSTM network structure diagram. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0054] The first aspect of the present invention discloses a method for real-time prediction and dynamic correction of target characteristics based on deep learning, the method comprising (such as Figure 1 shown):

[0055] Step 1: Collect historical simulation data and experimental data, clean the data, process the noise, missing values ​​and outliers, and generate a data set for network training;

[0056] Step 2: Preprocess the data set to obtain N anchor boxes, cluster the bounding box sizes in the anchor box labels using the K-means algorithm, and extract features of the radar reflection cross section and infrared radiation characteristics in the data set using the DarkNet53 network;

[0057] Step 3: Through the Soft-SPP module based on feature fusion, the features extracted by the DarkNet53 network are input into different pooling layers of the pooling core respectively. After the pooling layer uses the SoftPool method for pooling, the features of different feature channels are connected in series;

[0058] Step 4: Pass the concatenated feature information through a 1×1 and a 3×3 convolution layer to further extract the target features;

[0059] Step 5: The N / 3 anchor frames obtained by preprocessing the data are used as references for prediction of the final extracted feature image, and the construction of the convolutional neural network consisting of the DarkNet53 network, the Soft-SPP module, and the convolutional layer is completed;

[0060] Step 6: Input the characteristic information of the electromagnetic scattering characteristics and infrared radiation characteristics of the target predicted by the convolutional neural network into the LSTM recurrent neural network for further processing and network model training to obtain a neural network model with a CNN-RNN structure after training;

[0061] Step 7: Input the experimental data to be predicted and analyzed, including the electromagnetic scattering characteristics and infrared radiation characteristics of the target, into the trained network model for real-time prediction and analysis.

[0062] According to the method of the first aspect of the present invention, the method further includes constructing a dynamic correction mechanism; specifically including (such as Figure 2 shown):

[0063] Set the initial parameters of the constructed neural network model with CNN-RNN structure;

[0064] The constructed neural network model and initial parameters are used for simulation, and the error between the predicted result and the real data of the target characteristic actually collected is calculated using the mean-split error (MSE);

[0065] Analyze the distribution and characteristics of errors, use the gradient descent method to adjust the neural network model parameters, and update the neural network model;

[0066] Repeat the above steps until the stopping condition is met, and use an independent validation set to validate the modified neural network model;

[0067] During the application of the neural network model, new experimental data are collected and fed back into the correction mechanism to achieve continuous optimization of the neural network model.

[0068] According to the method of the first aspect of the present invention, in step 1, the data is cleaned; specifically, the data is smoothed using a sliding average and a median filter to reduce noise; the missing values ​​are filled in based on similarity using a K-nearest neighbor algorithm; and outliers are detected and filtered based on the statistical 3σ principle.

[0069] According to the method of the first aspect of the present invention, in step 3, based on the Soft-SPP module of feature fusion, SoftPool is used instead of MaxPool for pooling; wherein:

[0070]

[0071] Among them, a i is the pixel point in the neighborhood R, w i is the weight of the i-th pixel, and j is the pixel number in the neighborhood;

[0072] The output value of the pooling is calculated using the weight w i and the corresponding pixel point ai We weight and sum them to get:

[0073]

[0074] in, Represents the output value after pooling, and ﹡ represents the dot product.

[0075] According to the method of the first aspect of the present invention, in step 6: the output features of the convolutional neural network are serialized so that the input data of the LSTM network is consistent, and through sequence modeling and long-term dependencies of memory units, the time series information of electromagnetic scattering characteristics and infrared radiation characteristics is further analyzed and extracted to capture the changes and behavior patterns of the target in a dynamic environment.

[0076] According to the method of the first aspect of the present invention, an attention mechanism is added to the neural network, and the back propagation algorithm and adaptive learning rate adjustment strategy are used to improve the recognition and prediction accuracy of target characteristics to adapt to different scenarios and dynamic changes in target characteristics.

[0077] First embodiment

[0078] A method for real-time prediction and dynamic correction of target characteristics based on deep learning includes: establishing a neural network model with a CNN-RNN structure; cleaning the radar cross section (RCS) and infrared radiation characteristic data of the target, processing the noise, missing values ​​and outliers in the data, and ensuring the data quality. A data set is made for Darknet53 convolutional neural network model training, and shallow feature information is obtained through the feature extraction layer; the last layer of the network input is subjected to feature fusion with the Soft-SPP module; the fused information is subjected to the convolution layer again to extract features; the feature map output by the network uses the feature information obtained by the anchor frame obtained by preprocessing; the feature map is sequentially input into the LSTM recurrent neural network, and its feature information is serialized, and the LSTM model is trained to realize real-time prediction and analysis of the target radar cross section and infrared radiation characteristics; a dynamic correction mechanism is designed, and the target characteristic model is corrected online in combination with real-time prediction data and experimental data to improve the prediction accuracy. The present invention comprehensively considers the interaction among electromagnetic fields, temperature fields and infrared radiation fields, provides more comprehensive prediction results, and adopts deep learning algorithms to realize real-time prediction and precise correction of target characteristics. The system can continuously learn and optimize according to new test data, thus ensuring the efficiency and accuracy of prediction, and is suitable for complex real-time application requirements.

[0079] Second embodiment

[0080] Multi-physics coupling is an advanced numerical simulation technology that can simultaneously consider and simulate multiple physical phenomena and their interactions. These physical phenomena usually include but are not limited to electromagnetic fields, thermal fields, mechanical fields, fluid fields, acoustic fields, etc. In multi-physics problems, different physical fields influence and restrict each other through various physical mechanisms. Therefore, the simulation of a single physical field often cannot accurately reflect the actual physical process.

[0081] A neural network with a CNN-RNN structure is a classic deep learning technology that automatically learns and extracts features from large amounts of data by simulating the way the human brain processes information, and then predicts specific attributes of the target object. This algorithm uses the structure of a convolutional neural network, which is composed of multiple layers of convolutional layers and activation layers stacked together, each of which can capture different levels of features of the input data. During the training process, the network generates predictions through forward propagation and adjusts network parameters through backpropagation to minimize prediction errors. Ultimately, the algorithm can efficiently and accurately predict the electromagnetic scattering characteristics or infrared radiation characteristics of the target, and is widely used in military reconnaissance, medical imaging analysis, and industrial detection. Its construction specifically includes the following steps:

[0082] Step 1: Collect historical simulation data and experimental data, clean the data, process the noise, missing values ​​and outliers in the data, and ensure data quality. Use methods such as sliding average and median filtering to smooth the data to reduce the impact of noise; use the K nearest neighbor algorithm (KNN) to fill in missing values ​​based on similarity; and use the 3σ principle (triple standard deviation method) based on statistical methods to detect and filter outliers.

[0083] In statistics, the 68–95–99.7 principle is the percentage of values ​​within one standard deviation, two standard deviations, and three standard deviations from the mean in a normal distribution. More precise numbers are 68.27%, 95.45%, and 99.73%. This is the 3σ principle:

[0084] The probability that the value is distributed in (μ-σ,μ+σ) is 0.6827;

[0085] The probability that the value is distributed in (μ-2σ,μ+2σ) is 0.9545;

[0086] The probability that the value is distributed in (μ-3σ,μ+3σ) is 0.9973.

[0087] It can be considered that the values ​​of a normally distributed data set are almost all concentrated in the interval (μ-3σ,μ+3σ), and the possibility of exceeding this range is less than 0.3%. Therefore, the data that is not in this interval will be eliminated as outliers.

[0088] Step 2: Preprocess the dataset to obtain N anchor boxes, use the K-means algorithm to cluster the bounding box sizes in the anchor box labels, and use the DarkNet53 network to extract features of the radar reflection cross section and infrared radiation characteristics in the dataset.

[0089] The idea of ​​the K-means algorithm is to classify given samples based on the Euclidean distance. If two samples are close to each other, they are considered to belong to the same category. Finally, the sample set is divided into k clusters, and the points in the clusters are connected as closely as possible, while the distance between clusters is as large as possible. The average overlap (AvgIoU) is used as a measure of similarity between data:

[0090]

[0091] Among them, B is the cluster sample, that is, the width and height of the real box of the data set; C is the center of the cluster; n k The number of samples belonging to the kth cluster; n is the total number of samples; k represents the number of clusters; I IOU (B, C) represents the intersection-over-union ratio of the cluster center box and the cluster box.

[0092] Step 3: Add the Soft-SPP module to the last layer of the DarkNet50 network.

[0093] The DarkNet53 network uses a lot of 1x1 convolution and 3x3 convolution, among which 1x1 is mainly used for channel expansion and reduction. The overall convolution network adopts the structure of convolution layer + batch normalization layer + activation function layer. Figure 3 shown.

[0094] The Soft-SPP module consists of four branches. The features of the previous level are soft-pooled by three sizes of pooling kernels, and then connected in series with the unpooled features of the previous level, i.e., superimposed. Figure 4 In the softpool part, different degrees of padding operations are performed to ensure that the output size of each branch is consistent. The specific operations are as follows:

[0095] For the local area R in the feature map a, each pixel a is calculated by the following formula i Calculate a weight:

[0096]

[0097] Among them, w i is the corresponding pixel weight, and j represents the element in the neighborhood R of the pixel.

[0098] The final pooled output value is calculated using the weight w i and the corresponding pixel point a iWeighted and summed we get:

[0099]

[0100] in, Represents the output value after pooling, and ﹡ represents the dot product.

[0101] Step 4: The concatenated feature information is passed through a 1×1 and a 3×3 convolutional layer again to further extract the target features;

[0102] Step 5: Input the characteristic information of the electromagnetic scattering characteristics and infrared radiation characteristics of the target obtained by the convolutional neural network into the LSTM recurrent neural network for further processing and network model training, such as Figure 5 As shown in the figure. LSTM is a special type of recurrent neural network that solves the gradient vanishing and gradient exploding problems in traditional RNNs. It is particularly suitable for processing time series data and can effectively capture and learn long-term dependencies. Its core lies in the unique unit structure. Each LSTM unit includes a memory unit and three gates: input gate, forget gate, and output gate. These gates control the flow of information in the memory unit, thereby achieving modeling and memory of long-term dependencies of sequence data.

[0103] The feature information of the target obtained by the convolutional neural network is serialized and then passed through the input gate to determine which information will be added to the memory unit. A sigmoid activation function is used to determine the update amount of each state, and a tanh activation function is used to create a new candidate value vector, i.e., the update vector.

[0104] The cell state is the memory transferred between time series. The new cell state is composed of the previous state and the current candidate value. The forget gate determines which information in the previous memory cell state will be forgotten or retained. Through the sigmoid activation function, it determines whether the information in the past state will be transferred to the current state. The formula is:

[0105]

[0106] where c <t>< / t> is the cell state, c <t-1>< / t-1> is the unit state at the previous moment, c' <t>< / t> is the current candidate value, is the forget gate output, is the updated forget gate output, σ represents the activation function, W f and b f are learnable weights and biases, X <t>< / t>is the current input, a <t-1>< / t-1> is the input at the previous moment.

[0107] The sigmoid function is used to ensure that the output value is between [0,1], and the forget gate Same dimensions as previous cell state, can be element-wise multiplied.

[0108] The final output of LSTM is determined by the output gate, which controls the output result based on the current input and memory cell state and the sigmoid activation function, while using a tanh activation function to scale the memory cell state to the desired range.

[0109] Step 6: Establish a dynamic correction mechanism, set the initial network model parameters according to the experimental data, use the current network model and parameters for simulation, and calculate the average error between the predicted results and the real data of the actual collected target characteristics; by analyzing the distribution and characteristics of the error, use the gradient descent method to adjust the network model parameters, update the network model, and ensure the generalization ability of the model; in the process of model application, collect new experimental data and feed these data back to the correction mechanism to achieve continuous optimization of the model.

[0110] A second aspect of the present invention discloses a device for real-time prediction and dynamic correction of target characteristics based on deep learning, the device comprising a processing unit, the processing unit being configured to execute:

[0111] Step 1: Collect historical simulation data and experimental data, clean the data, process the noise, missing values ​​and outliers, and generate a data set for network training;

[0112] Step 2: Preprocess the data set to obtain N anchor boxes, cluster the bounding box sizes in the anchor box labels using the K-means algorithm, and extract features of the radar reflection cross section and infrared radiation characteristics in the data set using the DarkNet53 network;

[0113] Step 3: Through the Soft-SPP module based on feature fusion, the features extracted by the DarkNet53 network are input into different pooling layers of the pooling core respectively. After the pooling layer uses the SoftPool method for pooling, the features of different feature channels are connected in series;

[0114] Step 4: Pass the concatenated feature information through a 1×1 and a 3×3 convolution layer to further extract the target features;

[0115] Step 5: The N / 3 anchor frames obtained by preprocessing the data are used as references for prediction of the final extracted feature image, and the construction of the convolutional neural network consisting of the DarkNet53 network, the Soft-SPP module, and the convolutional layer is completed;

[0116] Step 6: Input the characteristic information of the electromagnetic scattering characteristics and infrared radiation characteristics of the target predicted by the convolutional neural network into the LSTM recurrent neural network for further processing and network model training to obtain a neural network model with a CNN-RNN structure after training;

[0117] Step 7: Input the experimental data to be predicted and analyzed, including the electromagnetic scattering characteristics and infrared radiation characteristics of the target, into the trained network model for real-time prediction and analysis.

[0118] According to the apparatus of the second aspect of the present invention, the processing unit is further configured to execute:

[0119] Set the initial parameters of the constructed neural network model with CNN-RNN structure;

[0120] The constructed neural network model and initial parameters are used for simulation, and the error between the predicted result and the real data of the target characteristic actually collected is calculated using the mean-split error (MSE);

[0121] Analyze the distribution and characteristics of errors, use the gradient descent method to adjust the neural network model parameters, and update the neural network model;

[0122] Repeat the above steps until the stopping condition is met, and use an independent validation set to validate the modified neural network model;

[0123] During the application of the neural network model, new experimental data are collected and fed back into the correction mechanism to achieve continuous optimization of the neural network model.

[0124] The third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the method for real-time prediction and dynamic correction of target characteristics based on deep learning described in the first aspect of the present disclosure is implemented.

[0125] The fourth aspect of the present invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for real-time prediction and dynamic correction of target characteristics based on deep learning described in the first aspect of the present disclosure is implemented.

[0126] In summary, the present invention establishes a neural network model with a CNN-RNN structure; the radar cross section (RCS) and infrared radiation characteristic data of the target are cleaned, and the noise, missing values ​​and outliers in the data are processed to ensure data quality. The present invention makes a data set for Darknet53 convolutional neural network model training, obtains shallow feature information through the feature extraction layer; the last layer of the network input is feature fused with the Soft-SPP module; the fused information is again extracted through the convolution layer; the feature map output by the network uses the feature information obtained by the anchor frame obtained by preprocessing; it is sequentially input into the LSTM recurrent neural network, the feature information is serialized, and the LSTM model is trained to realize real-time prediction and analysis of the target radar cross section and infrared radiation characteristics; a dynamic correction mechanism is designed, and the target characteristic model is corrected online in combination with real-time prediction data and experimental data to improve the prediction accuracy.

[0127] The present invention comprehensively considers the interaction among electromagnetic fields, temperature fields and infrared radiation fields, provides more comprehensive prediction results, and adopts deep learning algorithms to realize real-time prediction and precise correction of target characteristics. The system can continuously learn and optimize according to new test data, thus ensuring the efficiency and accuracy of prediction, and is suitable for complex real-time application requirements.

[0128] The beneficial technical effects brought about by the present invention include: the present invention comprehensively considers the interaction between electromagnetic fields, temperature fields and infrared radiation fields, provides more comprehensive prediction results, adopts deep learning algorithms, realizes real-time prediction and precise correction of target characteristics, and the system can continuously learn and optimize according to new test data, ensuring the efficiency and accuracy of the prediction, and is suitable for complex real-time application requirements.

[0129] Please note that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above-mentioned embodiments only express several implementation methods of the present application, and their descriptions are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, without departing from the concept of the present application, several variations and improvements can be made, which all belong to the scope of protection of the present application. Therefore, the scope of protection of the patent in this application shall be based on the attached claims.

Claims

1. A method for real-time prediction and dynamic correction of target characteristics based on deep learning, characterized in that: The method comprises: Step 1: Collect historical simulation data and experimental data, clean the data, process the noise, missing values ​​and outliers, and generate a data set for network training; Step 2: Preprocess the data set to obtain N anchor boxes, cluster the bounding box sizes in the anchor box labels using the K-means algorithm, and extract features of the radar reflection cross section and infrared radiation characteristics in the data set using the DarkNet53 network; Step 3: Through the Soft-SPP module based on feature fusion, the features extracted by the DarkNet53 network are input into different pooling layers of the pooling core respectively. After the pooling layer uses the SoftPool method for pooling, the features of different feature channels are connected in series; Step 4: Pass the concatenated feature information through a 1×1 and a 3×3 convolution layer to further extract the target features; Step 5: The N / 3 anchor frames obtained by preprocessing the data are used as references for prediction of the final extracted feature image, and the construction of the convolutional neural network consisting of the DarkNet53 network, the Soft-SPP module, and the convolutional layer is completed; Step 6: Input the characteristic information of the electromagnetic scattering characteristics and infrared radiation characteristics of the target predicted by the convolutional neural network into the LSTM recurrent neural network for further processing and network model training to obtain a neural network model with a CNN-RNN structure after training; Step 7: Input the experimental data to be predicted and analyzed, including the electromagnetic scattering characteristics and infrared radiation characteristics of the target, into the trained network model for real-time prediction and analysis.

2. The method for real-time prediction and dynamic correction of target characteristics based on deep learning according to claim 1, characterized in that: The method further includes constructing a dynamic correction mechanism; specifically including: Set the initial parameters of the constructed neural network model with CNN-RNN structure; The constructed neural network model and initial parameters are used for simulation, and the error between the predicted result and the real data of the target characteristic actually collected is calculated using the mean-split error (MSE); Analyze the distribution and characteristics of errors, use the gradient descent method to adjust the neural network model parameters, and update the neural network model; Repeat the above steps until the stopping condition is met, and use an independent validation set to validate the modified neural network model; During the application of the neural network model, new experimental data are collected and fed back into the correction mechanism to achieve continuous optimization of the neural network model.

3. The method for real-time prediction and dynamic correction of target characteristics based on deep learning according to claim 2, characterized in that: In step 1, the data is cleaned, including: using sliding average and median filtering to smooth the data to reduce noise; using the K nearest neighbor algorithm to fill in missing values ​​based on similarity; and detecting and filtering outliers based on the statistical 3σ principle.

4. The method for real-time prediction and dynamic correction of target characteristics based on deep learning according to claim 3, characterized in that: In step 3, based on the Soft-SPP module of feature fusion, SoftPool is used instead of MaxPool for pooling; where: Among them, a i is the pixel point in the neighborhood R, w i is the weight of the i-th pixel, and j is the pixel number in the neighborhood; The output value of the pooling is calculated using the weight w i and the corresponding pixel point a i We weight and sum them to get: in, Represents the output value after pooling, and ﹡ represents the dot product.

5. The method for real-time prediction and dynamic correction of target characteristics based on deep learning according to claim 4, characterized in that: In step 6: the output features of the convolutional neural network are serialized to make the input data of the LSTM network consistent. Through sequence modeling and long-term dependencies of memory units, the time series information of electromagnetic scattering characteristics and infrared radiation characteristics is further analyzed and extracted to capture the changes and behavior patterns of the target in a dynamic environment.

6. The method for real-time prediction and dynamic correction of target characteristics based on deep learning according to claim 5, characterized in that: An attention mechanism is added to the neural network, and the back-propagation algorithm and adaptive learning rate adjustment strategy are used to improve the recognition and prediction accuracy of target characteristics to adapt to different scenarios and dynamic changes in target characteristics.

7. A device for real-time prediction and dynamic correction of target characteristics based on deep learning, characterized in that: The apparatus comprises a processing unit configured to perform: Step 1: Collect historical simulation data and experimental data, clean the data, process the noise, missing values ​​and outliers, and generate a data set for network training; Step 2: Preprocess the data set to obtain N anchor boxes, cluster the bounding box sizes in the anchor box labels using the K-means algorithm, and extract features of the radar reflection cross section and infrared radiation characteristics in the data set using the DarkNet53 network; Step 3: Through the Soft-SPP module based on feature fusion, the features extracted by the DarkNet53 network are input into different pooling layers of the pooling core respectively. After the pooling layer uses the SoftPool method for pooling, the features of different feature channels are connected in series; Step 4: Pass the concatenated feature information through a 1×1 and a 3×3 convolution layer to further extract the target features; Step 5: The N / 3 anchor frames obtained by preprocessing the data are used as references for prediction of the final extracted feature image, and the construction of the convolutional neural network consisting of the DarkNet53 network, the Soft-SPP module, and the convolutional layer is completed; Step 6: Input the characteristic information of the electromagnetic scattering characteristics and infrared radiation characteristics of the target predicted by the convolutional neural network into the LSTM recurrent neural network for further processing and network model training to obtain a neural network model with a CNN-RNN structure after training; Step 7: Input the experimental data to be predicted and analyzed, including the electromagnetic scattering characteristics and infrared radiation characteristics of the target, into the trained network model for real-time prediction and analysis.

8. The device for real-time prediction and dynamic correction of target characteristics based on deep learning according to claim 7, characterized in that: The processing unit is further configured to perform: Set the initial parameters of the constructed neural network model with CNN-RNN structure; The constructed neural network model and initial parameters are used for simulation, and the error between the predicted result and the real data of the target characteristic actually collected is calculated using the mean-split error (MSE); Analyze the distribution and characteristics of errors, use the gradient descent method to adjust the neural network model parameters, and update the neural network model; Repeat the above steps until the stopping condition is met, and use an independent validation set to validate the modified neural network model; During the application of the neural network model, new experimental data are collected and fed back into the correction mechanism to achieve continuous optimization of the neural network model.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it implements the method for real-time prediction and dynamic correction of target characteristics based on deep learning as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements the method for real-time prediction and dynamic correction of target characteristics based on deep learning as described in any one of claims 1 to 6.

Images