Fuze falling angle identification method based on bispectrum analysis and deep belief network

Abstract

The invention discloses a fuze falling angle identification method based on bispectrum analysis and a deep belief network. The method comprises the following steps: S1, preprocessing a received fuze echo signal; s2, calculating a third-order cumulant of the preprocessed echo signal by applying an axial integral bispectrum method to obtain a bispectrum feature; s3, selecting and amplifying the integral bispectrum features by using a one-dimensional Fisher class separation degree to form feature vectors; s4, constructing a deep belief network model composed of two layers of restricted Boltzmann machines and a layer of back propagation neural network; and S5, inputting the bispectrum analysis features into a deep belief network, and completing accurate identification of the fuse falling angle. According to the method, the advantages of bispectrum analysis and the deep belief network are combined, so that efficient and accurate identification of the fuse falling angle is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of fuze signal processing, in particular to a fuze angle of fall recognition method based on bispectrum analysis and a deep belief network. BACKGROUND

[0002] Fuze accurate burst point control plays a crucial role in realizing efficient damage in the fuze-warhead matching process. The fuze burst point accurate control capability directly affects the comprehensive damage efficiency of the warhead on the ground target, and is an important index for the design of fuze of the ground killing and explosion type ammunition. In addition to the factors such as the type of striking ammunition, the type of striking target and the background of striking ground, the best burst height of the fuze is also affected by the angle of fall of the fuze. Therefore, the accurate recognition of the angle of fall of the fuze has an important influence on the damage efficiency of the ammunition.

[0003] The deep belief network (DBN) is a kind of probability generation model, which is a stack of multiple restricted Boltzmann machines (RBM), wherein each RBM layer is connected with the upper and lower layers, and the units in any layer are not connected with each other. Except for the first layer and the last layer, each layer of the DBN has two functions: as a hidden layer of the previous layer, or as an input (visible layer) of the next layer. A Softmax layer can be connected on the stacked RBM layer to use the whole DBN as a classifier, or it can be simply used for clustering of unlabeled data in unsupervised learning. The DBN is usually used for image, video sequence or motion capture recognition, clustering and generation.

[0004] High-order spectrum analysis is a frontier research direction in signal processing. In high-order spectrum, the bispectrum obtained from the third-order correlation function is most widely used. The bispectrum has the advantages of time shift invariance, scale change and phase preservation, and has good noise suppression effect. However, the two-dimensional bispectrum feature is difficult to further process, thus generating integral bispectrum feature extraction and its improved methods, such as integral bispectrum (RIB), axial integral bispectrum (AIB), circular integral bispectrum (CIB), and surrounding line integral bispectrum (SLIB). The integral bispectrum converts the two-dimensional bispectrum function into a one-dimensional function, which is beneficial to analysis and calculation.

[0005] In the increasingly complex electromagnetic environment of the battlefield, the echo signal of the fuze as the front end of the weapon system is often disturbed by noise.

[0006] Traditional fuze angle of fall recognition methods mostly rely on complex signal processing and feature extraction algorithms, and have problems such as large amount of calculation and limited recognition accuracy. With the development of deep learning, the deep belief network has made remarkable achievements in many fields due to its powerful feature extraction and classification ability, but its application in the field of fuze angle of fall recognition is still rare.

[0007] In view of the above problems, the application provides a fuze impact angle recognition method based on bispectrum analysis and deep belief network. SUMMARY

[0008] The application aims to overcome the existing defects and provide a fuze impact angle recognition method based on bispectrum analysis and deep belief network, which realizes efficient and accurate recognition of the fuze impact angle by combining the advantages of bispectrum analysis and deep belief network.

[0009] The technical scheme for realizing the above object is as follows: A fuze impact angle recognition method based on bispectrum analysis and deep belief network, comprising: Step S1, preprocessing the received fuze echo signal; Step S2, applying an axial integral bispectrum method to calculate the third-order cumulant of the preprocessed echo signal to obtain a bispectrum feature; Step S3, using a one-dimensional Fisher class separation degree to select and amplify the integral bispectrum feature to form a feature vector; Step S4, constructing a deep belief network model composed of two layers of restricted Boltzmann machines and one layer of back propagation neural networks; Step S5, inputting the bispectrum analysis feature into the deep belief network to complete accurate recognition of the fuze impact angle.

[0010] Preferably, in step S1, the preprocessing includes but is not limited to denoising, filtering and signal enhancement.

[0011] Preferably, in step S2, the axial integral bispectrum method is applied to calculate the third-order cumulant of the preprocessed echo signal to obtain a bispectrum feature, which includes: In the bispectrum double-frequency plane, along the path parallel to the axis or axis, the integral bispectrum is obtained, for a signal and its bispectrum has: its axial integral bispectrum is calculated by the following formula: .

[0012] Preferably, in step S3, the one-dimensional Fisher class separation degree is used to select and amplify the integral bispectrum feature to form a feature vector, which includes: The one-dimensional Fisher equation is expressed as: ; In the formula, and respectively represent the first The mean and variance of the bispectral values ​​of all samples of the signal class. This represents the overall center of the sample bispectrum for all types of signals at frequencies. Both indicate signal type Indicates the first The bispectrum was calculated from a set of observational data; The larger, the better and The stronger the classability between signal classes, the better the selection of signals with... The set of frequencies of the strongest Fisher class separability is used as the feature frequencies; Signals for different fuse drop angles Features, calculate the separation degree between different angles And standardize the process: ; In the formula, The feature points representing different fuse drop angles used for AIB calculations represent the number of feature points. and take the previous Each ... ; Re-establish the feature vectors of the four types of fuse drop angle signals, the first and the Class The feature vectors are as follows: ; .

[0013] Preferably, in step S4, a deep belief network model is constructed, consisting of two layers of Restricted Boltzmann Machines (RBMs) and one layer of Backpropagation Neural Network (BNN). The two RBMs are used for feature extraction, and the BNN is used for classification. This includes: The first input layer uses the signal feature vector as input information to transmit data information and perform normalization processing on the input signal. The second hidden layer uses a restricted Boltzmann machine as its activation function. When the feature vector is mapped to different feature spaces, the feature information is preserved, and a joint distribution between the observed data and the label is established. The number of hidden units is used as a parameter for optimization. The third layer is the output layer, which is a fully connected layer. It linearly weights the information output by the hidden layer neurons and outputs it as the output result of the entire neural network.

[0014] Preferably, in step S4, training the deep belief network model includes: The visible layer state, This is the visible layer offset. The state of the hidden layer. For the bias of the hidden layer, Let be the weight matrix between the visible layer and the hidden layer, where the visible layer has real values. The energy function can be expressed as: ; In the formula, This represents the weight matrix between the visible layer and the hidden layer. This indicates the number of neurons in the visible layer. express transpose, express transpose, This represents the standard deviation of neurons in the visible layer; By defining the energy function, the probability distribution function can be defined as follows: ; In the formula, Represents all possible hidden layer states Summation, Represents all possible visible layer states and hidden layer state Perform double summation; A two-layer restricted Boltzmann machine arbitrarily approximates the probability distribution function. The maximum likelihood estimation method is used to describe the probability distribution function of the input data. To simplify the calculation, the negative minimum likelihood function is adopted: ; ; In the formula, These represent the parameters that need to be adjusted during training; the network training process is the parameter... The optimization process; This refers to the ability of data features extracted from the network to represent the original data, and it determines whether the extracted hidden features can be used for classification. It forms the basis for network optimization and adjustment, provides direction for adjustment, and ensures the smooth progress of the entire network training; Hidden layer training process: ; In the formula, This represents the number of neurons in the hidden layer; Visible layer training process: ; The training process of a network is essentially a search for the optimal parameters, which are: ; Update parameters using gradient descent: ; In the formula, Indicates the learning rate; For the Bernoulli-restricted Boltzmann machine, the parameters are obtained by using the contrast divergence approximation: ; ; ; In the formula, Represents the loss function with respect to the weight matrix. The partial derivatives are used for weight updates. Represents the loss function with respect to the visible layer offset. The partial derivative is used for offset updates. This indicates that the loss function is biased towards the hidden layer. The partial derivative is used for bias updates. Indicates the first The state of the visible layer in the nth iteration represents the state of the th... The state of the hidden layer in the next iteration. This indicates the initial state of the visible layer. This represents the inferred state of the hidden layer in the initial state.

[0015] Preferably, in step S5, the bispectral analysis features are input into a deep belief network to accurately identify the fuse drop angle, including: The amplitude probability density distribution of the signal was analyzed using 200 periodic signals. Each period contains 512 points, and the range between the maximum and minimum amplitudes is divided into 512 parts. Calculate the number of points in each part, know the amplitude probability density curve of the measured signal, and obtain the theoretical distribution curve based on the theoretical distribution formula; Subtract the corresponding points, and then use the sum of the differences as the fitting error. Use the fitting error to determine the degree of fit between the measured signal amplitude probability density curve and the theoretical distribution curve. The fitting error is expressed as: ; In the formula, This represents the probability density curve of the measured amplitude data. Represents the theoretical distribution curve; A goodness-of-fit is introduced to describe the degree of shape fit between the theoretical distribution curve and the probability density curve of the ground echo amplitude. The statistical parameter of the goodness-of-fit is expressed as the coefficient of determination. : ; In the formula, Represents the actual observed value. Indicates the theoretical prediction value. This represents the mean of the theoretically predicted values; The measured echo amplitude probability density curves for different fuze drop angles are compared with the theoretical distribution curves to achieve accurate identification of the fuze drop angle.

[0016] The beneficial effects of this invention are: This invention uses the echo signal from bispectral analysis as the input of a deep belief network. Based on the powerful feature extraction and classification capabilities of the deep belief network, and the advantages of bispectral analysis in preserving signal features, suppressing Gaussian noise, and improving recognition efficiency, it fully leverages the advantages of both to improve the accuracy and efficiency of fuse drop angle recognition. Attached Figure Description

[0017] Figure 1 This is a flowchart of a fuse landing angle identification method based on bispectral analysis and deep belief network according to the present invention; Figure 2 This is a schematic diagram of AIB on the bispectral dual-frequency plane in this invention; Figure 3 This is a schematic diagram of the deep belief network structure in this invention; Figure 4 This is a grassland type distribution map provided in the embodiments of the present invention; Figure 5 This is an AIB analysis diagram of different fuse drop angles in an embodiment of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will now be clearly and completely described in conjunction with the accompanying drawings. In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] like Figure 1 As shown, a fuse landing angle identification method based on bispectral analysis and deep belief network includes: Step S1: Preprocess the received fuze echo signal.

[0021] In the embodiments, preprocessing includes, but is not limited to, denoising, filtering, and signal enhancement.

[0022] Step S2: Apply the axial integral bispectral method to calculate the third-order cumulative quantity of the preprocessed echo signal and obtain the bispectral characteristics.

[0023] In this embodiment, the axial integral bispectral method is applied to calculate the third-order cumulant of the preprocessed echo signal to obtain bispectral features, including: like Figure 2 As shown, on the bispectral dual-frequency plane, along with... shaft or Path integration parallel to the axis yields an axially integrated bispectrum, for the signal and its bispectral have: Its axial integral bispectral Calculated using the following formula: .

[0024] Step S3: Use one-dimensional Fisher class separation to select and amplify the integral bispectral features to form a feature vector.

[0025] In this embodiment, a one-dimensional Fisher class is introduced to calculate the discriminative power of AIB (axial integral bispectral) feature points in the signal. Unimportant AIB feature points are removed from the signal, and highly discriminative AIB feature points are selected as feature vectors. This reduces the dimensionality of the feature vectors, shortens the network training time, and improves the network's applicability.

[0026] In this embodiment, a one-dimensional Fisher-class separation is used to select and amplify the integral bispectral features to form a feature vector, including: The one-dimensional Fisher equation is expressed as: ; In the formula, and They represent the first The mean and variance of the bispectral values ​​of all samples of the signal class. This represents the overall center of the sample bispectrum for all types of signals at frequencies. Both indicate signal type Indicates the first The bispectrum was calculated from a set of observational data; The larger, the better and The stronger the classability between signal classes, the better the selection of signals with... The set of frequencies of the strongest Fisher class separability is used as the feature frequencies; Signals for different fuse drop angles Features, calculate the separation degree between different angles And standardize the process: ; In the formula, Feature points representing different fuse drop angles used in AIB calculations. Indicates the number of feature points; and take the previous Each ... ; Re-establish the feature vectors of the four types of fuse drop angle signals, the first and the Class The feature vectors are as follows: ; .

[0027] Step S4: Construct a deep belief network model consisting of two layers of restricted Boltzmann machines and one layer of backpropagation neural network.

[0028] In this embodiment, a deep belief network (DBN) model consisting of two layers of restricted Boltzmann machines (RBMs) and one layer of backpropagation neural networks (BPs) is constructed, such as... Figure 3 As shown, a two-layer Restricted Boltzmann Machine is used for feature extraction, and a one-layer Backpropagation Neural Network is used for classification, including: The first input layer uses the signal feature vector as input information to transmit data information and perform normalization processing on the input signal. The second hidden layer uses a restricted Boltzmann machine as its activation function. When the feature vector is mapped to different feature spaces, the feature information is preserved, and a joint distribution between the observed data and the label is established. The number of hidden units is used as a parameter for optimization. The third layer is the output layer, which is a fully connected layer. It linearly weights the information output by the hidden layer neurons and outputs it as the output result of the entire neural network.

[0029] In this embodiment, training the deep belief network model includes: The visible layer state, This is the visible layer offset. The state of the hidden layer. For the bias of the hidden layer, Let be the weight matrix between the visible layer and the hidden layer, where the visible layer has real values. The energy function can be expressed as: ; In the formula, This represents the weight matrix between the visible layer and the hidden layer. This indicates the number of neurons in the visible layer. express transpose, express transpose, This represents the standard deviation of neurons in the visible layer; By defining the energy function, the probability distribution function can be defined as follows: ; In the formula, Represents all possible hidden layer states Summation, Represents all possible visible layer states and hidden layer state Perform double summation; A two-layer restricted Boltzmann machine arbitrarily approximates the probability distribution function. The maximum likelihood estimation method is used to describe the probability distribution function of the input data. To simplify the calculation, the negative minimum likelihood function is adopted: ; ; In the formula, These represent the parameters that need to be adjusted during training; the network training process is the parameter... The optimization process; This refers to the ability of data features extracted from the network to represent the original data, and it determines whether the extracted hidden features can be used for classification. It forms the basis for network optimization and adjustment, providing adjustment direction and ensuring the smooth progress of the entire network training; based on the visible layer, the state of all neurons is directly calculated. This parallel mechanism can greatly improve the speed of training and inference. The reason why parallel computation is possible is that neurons between RBM layers do not form connections, eliminating the dependency between neurons between layers. Hidden layer training process: ; In the formula, This represents the number of neurons in the hidden layer; Visible layer training process: ; The training process of a network is essentially a search for the optimal parameters, which are: ; Update parameters using gradient descent: ; In the formula, Indicates the learning rate; For the Bernoulli-restricted Boltzmann machine, the parameters are obtained by using the contrast divergence approximation: ; ; ; In the formula, Represents the loss function with respect to the weight matrix. The partial derivatives are used for weight updates. Represents the loss function with respect to the visible layer offset. The partial derivative is used for offset updates. This indicates that the loss function is biased towards the hidden layer. The partial derivative is used for bias updates. Indicates the first The state of the visible layer in the next iteration. Indicates the first The state of the hidden layer in the next iteration. This indicates the initial state of the visible layer. This represents the inferred state of the hidden layer in the initial state.

[0030] The training process of DBN can be viewed as the initialization of weight parameters, overcoming the shortcomings of BP networks. It only requires a partial search of the weight parameter space. Compared to feedforward neural networks, DBN converges faster.

[0031] Step S5: Input the bispectral analysis features into the deep belief network to accurately identify the fuse drop angle.

[0032] In this embodiment, bispectral analysis features are input into a deep belief network to accurately identify the fuse drop angle, including: The amplitude probability density distribution (PDF) of the signal was analyzed using 200-cycle signals. Each cycle contains 512 points, and the range between the maximum and minimum amplitudes is divided into 512 parts. Calculate the number of points in each part, know the PDF curve of the measured signal, and obtain the theoretical distribution curve according to the theoretical distribution formula; Subtract the corresponding points, and then use the sum of the differences as the fitting error. Use the fitting error to determine the degree of fit between the measured signal PDF curve and the theoretical distribution curve. The fitting error is expressed as: ; In the formula, This represents the measured amplitude PDF curve. Represents the theoretical distribution curve; A goodness-of-fit is introduced to describe the degree of shape fit between the theoretical distribution curve and the ground echo amplitude PDF curve. The statistical parameter of the goodness-of-fit is expressed as the coefficient of determination. : ; In the formula, Represents the actual observed value. Indicates the theoretical prediction value. This represents the mean of the theoretically predicted values; The measured echo amplitude PDF curves for different fuze drop angles are compared with the theoretical distribution curves to achieve accurate identification of the fuze drop angle.

[0033] The present invention will be described in detail through the following specific embodiments. The test used a 122GHz signal with a 1GHz bandwidth and a sawtooth linear frequency modulation (LFM) signal. The test equipment and gimbal were mounted on the top of a fixed device. The measuring equipment was positioned five meters vertically above the ground. The angle of the signal's fall to the ground was controlled by the gimbal. Data measurements were taken for three different terrain types. The amplitude probability density distribution of the data was analyzed. The test data were compared with normal, Rayleigh, K-distribution, Weibull, normal-logarithmic, gamma, G0, and K-root distributions. The main focus was on three terrain types: alpine marsh meadow, alpine meadow, and alpine semi-desert grassland. The grassland type distribution maps are shown below. Figure 4 As shown, using an integrated transceiver radar system, taking alpine meadows as an example, the echo signals of four different fuse drop angles of 20°, 45°, 65°, and 85° were tested.

[0034] like Figure 5 The figure shows the Alpine Meadow Echoes (AIBs) of the alpine meadow at elevation angles of 20°, 45°, 65°, and 85°. It is clear from the figure that the signal characteristics differ significantly for different elevation angles. The AIBs are used as signal features and as input information for the DBN (Deep Learning Network). The horizontal axis represents the number of points, and the vertical axis represents the normalized value. The normalized features are used as network input. Normalization is beneficial for network training and prediction, avoiding overfitting and underfitting.

[0035] Ground echoes from four different landing angles were classified using DBN. There were 60 data points for each landing angle, and a total of 240 data points for the four different landing angles.

[0036] To verify the effectiveness and universality of the method, two different terrains, alpine dry grassland and alpine swamp, were selected for angle of attack identification. The results showed that the selected AIB features can effectively identify the glancing angle of targets in typical alpine regions.

[0037] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying the landing angle of a fuse based on bispectral analysis and deep belief networks, characterized in that, include: Step S1: Preprocess the received fuze echo signal; Step S2: Apply the axial integral bispectral method to calculate the third-order cumulant of the preprocessed echo signal and obtain the bispectral characteristics. Step S3: Use one-dimensional Fisher class separation to select and amplify the integral bispectral features to form a feature vector; Step S4: Construct a deep belief network model consisting of two layers of restricted Boltzmann machines and one layer of backpropagation neural network; Step S5: Input the bispectral analysis features into the deep belief network to accurately identify the fuse drop angle.

2. The fuse landing angle identification method based on bispectral analysis and deep belief network according to claim 1, characterized in that, In step S1, preprocessing includes, but is not limited to, denoising, filtering, and signal enhancement.

3. The fuse landing angle identification method based on bispectral analysis and deep belief network according to claim 1, characterized in that, In step S2, the axial integral bispectral method is applied to calculate the third-order cumulative quantity of the preprocessed echo signal, obtaining bispectral characteristics, including: On the bispectral dual-frequency plane, along with shaft or Path integration parallel to the axis yields an axially integrated bispectrum, for the signal and its bispectral have: Its axial integral bispectral Calculated using the following formula: 。 4. The fuse landing angle identification method based on bispectral analysis and deep belief network according to claim 3, characterized in that, In step S3, a one-dimensional Fisher class separation degree is used to select and amplify the integral bispectral features to form a feature vector, including: The one-dimensional Fisher equation is expressed as: ； In the formula, and They represent the first The mean and variance of the bispectral values ​​of all samples of the signal class. This represents the overall center of the sample bispectrum for all types of signals at frequencies. and Both indicate the signal type. Indicates the first The bispectrum was calculated from a set of observational data; The larger, the better and The stronger the classability between signal classes, the better the selection of signals with... The set of frequencies of the strongest Fisher class separability is used as the feature frequencies; Signals for different fuse drop angles Features, calculate the separation degree between different angles And standardize it: ； In the formula, Feature points representing different fuse drop angles used in AIB calculations. Indicates the number of feature points; and take the previous Each ... ； Re-establish the feature vectors of the four types of fuse drop angle signals, the first and the Class The feature vectors are as follows: ； 。 5. The fuse landing angle identification method based on bispectral analysis and deep belief network according to claim 4, characterized in that, In step S4, a deep belief network model is constructed, consisting of two layers of Restricted Boltzmann Machines (RBMs) and one layer of Backpropagation Neural Network (BNN). The two RBMs are used for feature extraction, and the BNN is used for classification. The first input layer uses the signal feature vector as input information to transmit data information and perform normalization processing on the input signal; The second hidden layer uses a restricted Boltzmann machine as its activation function. When the feature vector is mapped to different feature spaces, the feature information is preserved, and a joint distribution between the observed data and the label is established. The number of hidden units is used as a parameter for optimization. The third layer is the output layer, which is a fully connected layer. It linearly weights the information output by the hidden layer neurons and outputs it as the output result of the entire neural network.

6. The fuse landing angle identification method based on bispectral analysis and deep belief network according to claim 5, characterized in that, Step S4, training the deep belief network model, includes: The visible layer state, This is the visible layer offset. The state of the hidden layer. For the bias of the hidden layer, Let be the weight matrix between the visible layer and the hidden layer, where the visible layer has real values. The energy function can be expressed as: ； In the formula, This represents the weight matrix between the visible layer and the hidden layer. This indicates the number of neurons in the visible layer. express transpose, express transpose, This represents the standard deviation of neurons in the visible layer; By defining the energy function, the probability distribution function can be defined as follows: ； In the formula, Represents all possible hidden layer states Summation, Represents all possible visible layer states and hidden layer state Perform double summation; A two-layer restricted Boltzmann machine arbitrarily approximates the probability distribution function. The maximum likelihood estimation method is used to describe the probability distribution function of the input data. To simplify the calculation, the negative minimum likelihood function is adopted: ； ； In the formula, These represent the parameters that need to be adjusted during training; the network training process is the parameter... The optimization process; This refers to the ability of data features extracted from the network to represent the original data, and it determines whether the extracted hidden features can be used for classification. It forms the basis for network optimization and adjustment, provides direction for adjustment, and ensures the smooth progress of the entire network training; Hidden layer training process: ； In the formula, This represents the number of neurons in the hidden layer; Visible layer training process: ； The training process of a network is essentially a search for the optimal parameters, which are: ； Update parameters using gradient descent: ； In the formula, Indicates the learning rate; For the Bernoulli-restricted Boltzmann machine, the parameters are obtained by using the contrast divergence approximation: ； ； ； In the formula, Represents the loss function with respect to the weight matrix. The partial derivatives are used for weight updates. Represents the loss function with respect to the visible layer offset. The partial derivative is used for offset updates. This indicates that the loss function is biased towards the hidden layer. The partial derivative is used for bias updates. Indicates the first The state of the visible layer in the next iteration. Indicates the first The state of the hidden layer in the next iteration. This indicates the initial state of the visible layer. This represents the inferred state of the hidden layer in the initial state.

7. The fuse landing angle identification method based on bispectral analysis and deep belief network according to claim 6, characterized in that, In step S5, the bispectral analysis features are input into a deep belief network to accurately identify the fuse landing angle, including: The amplitude probability density distribution of the signal was analyzed using 200 periodic signals. Each period contains 512 points, and the range between the maximum and minimum amplitudes is divided into 512 parts. Calculate the number of points in each part, know the amplitude probability density curve of the measured signal, and obtain the theoretical distribution curve based on the theoretical distribution formula; Subtract the corresponding points, and then use the sum of the differences as the fitting error. Use the fitting error to determine the degree of fit between the measured signal amplitude probability density curve and the theoretical distribution curve. The fitting error is expressed as: ； In the formula, This represents the probability density curve of the measured amplitude data. Represents the theoretical distribution curve; A goodness-of-fit is introduced to describe the degree of shape fit between the theoretical distribution curve and the probability density curve of the ground echo amplitude. The statistical parameter of the goodness-of-fit is expressed as the coefficient of determination. : ； In the formula, Represents the actual observed value. Indicates the theoretical prediction value. This represents the mean of the theoretically predicted values; The measured echo amplitude probability density curves for different fuze drop angles are compared with the theoretical distribution curves to achieve accurate identification of the fuze drop angle.

Images