Signal processing method and system for S-band narrow-band digital processor of phased array

Abstract

The invention discloses a signal processing method and system for an S-band narrowband digital processor of a phased array, and the method comprises the steps: receiving and preprocessing a service signal from a comprehensive processor, and converting the service signal into a zero intermediate frequency signal; channelizing and decomposing the signal into a plurality of sub-channels with different center frequencies and bandwidths through a multi-phase filter bank; a deep belief network is adopted, a plurality of restricted Boltzmann machines are stacked to form the deep belief network, and deep feature extraction is carried out on each sub-channel signal; searching an optimal beam forming weight in the feature space by using a particle swarm optimization algorithm; and finally, according to the obtained optimal weight, performing weighted summation on each sub-channel signal to complete beam forming, and converting a result back to an analog intermediate frequency signal for subsequent emission processing. According to the invention, the processing performance of the phased array system in S-band narrowband digital processing and the accuracy of beam forming can be improved.

Description

Technical Field

[0001] The present invention relates to the field of antenna technology, and in particular to a signal processing method and system of an S-band narrowband digital processor of a phased array. Background Art

[0002] In phased array antenna systems, especially in the application scenario of S-band narrowband digital processors, traditional signal processing methods usually rely on simple filtering and fixed algorithms to achieve channelization processing. These traditional methods generally down-convert and analog-to-digital convert (ADC) the received RF signal first, and then divide the broadband signal into multiple narrowband channels by means of fast Fourier transform, which is the so-called channelization process. However, in the feature extraction stage, the existing technology often adopts rule-based methods or relatively simple machine learning models, such as support vector machines (SVM) or artificial neural networks, which cannot fully capture the subtle changes and patterns in complex environments. For the beamforming part, the traditional approach is usually to use static or semi-static weight configuration to determine the weight coefficient of each array element based on pre-set parameters or a limited number of optimization iterations. Although this approach can meet the needs of beamforming to a certain extent, its flexibility and adaptability are obviously insufficient in the face of dynamically changing environments, and it is difficult to achieve the best beam pointing accuracy and anti-interference capability.

[0003] Therefore, the S-band narrowband digital processor of the existing phased array antenna system has limitations in processing performance, especially in beamforming accuracy, resulting in the inability to meet the application requirements of modern phased array antenna systems in complex environments in terms of processing performance and beamforming accuracy. Summary of the invention

[0004] The embodiments of the present invention provide a signal processing method and system for an S-band narrowband digital processor of a phased array, which can improve the processing performance and beamforming accuracy of the phased array system in S-band narrowband digital processing.

[0005] An embodiment of the present invention provides a signal processing method of an S-band narrowband digital processor of a phased array, comprising:

[0006] Acquire an S-band narrowband service signal output by an integrated processor of the phased array system, and modulate and preprocess the service signal to obtain a preprocessed zero intermediate frequency signal;

[0007] Inputting the preprocessed zero intermediate frequency signal into a preset polyphase filter bank for channelization processing, so that the zero intermediate frequency signal is decomposed into different sub-channels, each sub-channel having a different center frequency and bandwidth;

[0008] A deep belief network formed by stacking a plurality of restricted Boltzmann machines is used to perform deep extraction of signal features of each subchannel of the zero intermediate frequency signal to obtain signal features; wherein the number of nodes in the visible layer of the first restricted Boltzmann machine of the deep belief network is the same as the number of subchannels, and the number of nodes in the visible layers of subsequent restricted Boltzmann machines of the deep belief network decreases in sequence;

[0009] The signal features are input into a pre-trained particle swarm optimization algorithm, in which each particle represents a set of beamforming weights, and the optimal beamforming weights are searched in the feature space through particle position update and velocity update operations;

[0010] According to the optimal beamforming weight, weighted summing is performed on the zero intermediate frequency signals of the sub-channels to implement beam synthesis, so as to obtain a digital beam synthesis signal;

[0011] The digital beamforming signal is converted into an analog intermediate frequency signal, and the analog intermediate frequency signal is pre-processed and then output to the transmitting and receiving TR component of the phased array system for subsequent transmitting processing.

[0012] Another embodiment of the present invention provides a signal processing system of an S-band narrowband digital processor of a phased array, including:

[0013] A preprocessing module is used to obtain the S-band narrowband service signal output by the integrated processor of the phased array system, and modulate and preprocess the service signal to obtain a preprocessed zero intermediate frequency signal;

[0014] A channelization processing module, used for inputting the preprocessed zero intermediate frequency signal into a preset polyphase filter bank for channelization processing, so that the zero intermediate frequency signal is decomposed into different sub-channels, each sub-channel having a different center frequency and bandwidth;

[0015] A feature extraction module, used to perform deep extraction of signal features of each subchannel of the zero intermediate frequency signal using a deep belief network formed by stacking multiple restricted Boltzmann machines to obtain signal features; wherein the number of nodes in the visible layer of the first restricted Boltzmann machine of the deep belief network is the same as the number of subchannels, and the number of nodes in the visible layers of subsequent restricted Boltzmann machines of the deep belief network decreases in sequence;

[0016] An optimization module, used for inputting the signal features into a pre-trained particle swarm optimization algorithm, in which each particle represents a set of beamforming weights, and searching for the optimal beamforming weights in the feature space through particle position update and velocity update operations;

[0017] A beamforming module, configured to perform weighted summation on the zero intermediate frequency signals of the sub-channels according to an optimal beamforming weight to implement beamforming and obtain a digital beamforming signal;

[0018] The signal conversion module is used to convert the digital beam synthesis signal into an analog intermediate frequency signal, and output the analog intermediate frequency signal to the transmitting and receiving TR component of the phased array system after preprocessing for subsequent transmission processing.

[0019] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0020] First, the S-band narrowband service signal transmitted by the integrated processor of the phased array system is accurately captured, and converted into a high-quality zero-IF signal through modulation and preprocessing operations; then the zero-IF signal is fed into a preset adaptive multi-phase filter group to achieve the purpose of channelization, so that the signal can be clearly divided into sub-channels with different characteristics; then a deep belief network composed of multiple restricted Boltzmann machines stacked in a specific structure is used, and the signal characteristics of each sub-channel are deeply analyzed by relying on the visible layer nodes and subsequent decreasing settings that match the first restricted Boltzmann machine with the number of sub-channels; then these characteristics are input into the pre-trained particle swarm optimization algorithm, and the optimal beamforming weights are efficiently searched by using the dynamic update rules of the particle position and velocity; finally, the zero-IF signals of each sub-channel are accurately weighted and summed according to this weight, and the beam synthesis is efficiently completed to obtain a digital beam synthesis signal, which is then output to the transmitting and receiving TR components of the phased array system through digital-to-analog conversion and necessary preprocessing for subsequent transmission processing. From the above analysis, it can be seen that the embodiments of the present invention introduce a deep belief network to perform deep extraction of signal features in complex environments, and use a particle swarm optimization algorithm to dynamically adjust the beamforming weights, thereby effectively solving the problems of insufficient processing performance and inaccurate beamforming in the prior art, and significantly improving the processing performance of the phased array system in S-band narrowband digital processing and the accuracy of beamforming. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of a signal processing method of an S-band narrowband digital processor of a phased array provided by an embodiment of the present invention;

[0022] Figure 2 It is a structural schematic diagram of a signal processing system of an S-band narrowband digital processor of a phased array provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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 creative work are within the scope of protection of the present invention.

[0024] See also Figure 1 , is a flow chart of a signal processing method of a phased array S-band narrowband digital processor provided by an embodiment of the present invention. The signal processing method of the phased array S-band narrowband digital processor comprises steps S10 to S15:

[0025] S10, acquiring an S-band narrowband service signal output by an integrated processor of the phased array system, and modulating and preprocessing the service signal to obtain a preprocessed zero intermediate frequency signal;

[0026] S11, inputting the preprocessed zero intermediate frequency signal into a preset polyphase filter bank for channelization processing, so that the zero intermediate frequency signal is decomposed into different sub-channels, each sub-channel having a different center frequency and bandwidth;

[0027] S12, using a deep belief network formed by stacking multiple restricted Boltzmann machines to perform deep extraction of signal features of each subchannel of the zero intermediate frequency signal to obtain signal features; wherein the number of nodes in the visible layer of the first restricted Boltzmann machine of the deep belief network is the same as the number of subchannels, and the number of nodes in the visible layers of subsequent restricted Boltzmann machines of the deep belief network decreases in sequence;

[0028] S13, inputting the signal feature into a pre-trained particle swarm optimization algorithm, where each particle in the particle swarm optimization algorithm represents a set of beamforming weights, and searching for the optimal beamforming weights in the feature space through particle position update and velocity update operations;

[0029] S14, performing weighted summation on the zero intermediate frequency signals of the sub-channels according to the optimal beamforming weights to implement beamforming, and obtaining a digital beamforming signal;

[0030] S15, converting the digital beamforming signal into an analog intermediate frequency signal, and pre-processing the analog intermediate frequency signal and outputting it to the transmitting and receiving TR component of the phased array system for subsequent transmitting processing.

[0031] In this embodiment, specifically, step S10 includes: first, using a high-sensitivity RF receiving antenna adapted to the S band to accurately receive the service signal output by the phased array system integrated processor to ensure efficient capture of weak signals. After receiving, quickly enter the modulation link, select the modulation method according to the application requirements, if the focus is on power utilization efficiency, amplitude modulation may be used, and if the pursuit of anti-interference is pursued, frequency or phase modulation is preferred, and parameters such as the modulation index are finely adjusted to make the signal carry information more stable. Then start the pre-processing process, the low-noise amplifier enhances the signal with a gain of about 20-30dB, and strictly controls the introduction of noise with the low noise coefficient characteristics; the mixer works with the local oscillator to move the RF signal to the intermediate frequency according to the specific local oscillator frequency, and the intermediate frequency bandpass filter immediately comes into play, with the center frequency of 70MHz and the bandwidth of 10MHz setting to accurately filter out clutter and obtain a relatively pure intermediate frequency signal. The subsequent high-speed and high-precision ADC implements sampling at a sampling rate several times the highest frequency of the signal and a bit number of 12 or more according to the Nyquist theorem to complete the analog-to-digital conversion. Finally, after digital down-conversion, the signal spectrum is moved to near zero frequency according to the new local oscillator frequency and digital low-pass filtering, and the automatic gain control dynamic monitoring and adjustment are combined to make the signal amplitude meet the subsequent processing requirements and successfully produce a zero intermediate frequency signal.

[0032] In this embodiment, specifically, step S15 includes: selecting a high-performance digital-to-analog conversion (DAC) chip to undertake the conversion of the digital beam synthesis signal to the analog intermediate frequency signal, for example, selecting a DAC chip with a resolution of 14 bits or even higher, which has ultra-high conversion accuracy, can accurately restore the signal details, and the sampling rate accurately matches the S-band signal characteristics to ensure signal integrity. During the conversion process, the stability and accuracy of the clock signal are controlled to avoid the introduction of noise or distortion due to clock jitter. After obtaining the analog intermediate frequency signal, pre-processing is carried out, and a low-pass filter is used to "purify" it. With its steep attenuation characteristics, high-frequency quantization noise and sampling spurious signals are effectively filtered out to ensure the purity of the signal spectrum. After layers of screening and optimization, the analog intermediate frequency signal is finally accurately output to the transmit and receive (TR) component of the phased array system, laying a solid foundation for the subsequent transmit processing link, so that it can participate in task processes such as signal radiation, target detection or long-distance communication with a high-quality signal state.

[0033] As one example of the above embodiment, the preprocessed zero intermediate frequency signal is input into a preset polyphase filter bank for channelization processing, so that the zero intermediate frequency signal is decomposed into different sub-channels, each sub-channel having a different center frequency and bandwidth, including:

[0034] The prototype filter is designed using the Kaiser window function and the maximum passband attenuation A p and stopband minimum attenuation A s Determine the β parameter of the Kaiser window function, and set the length of the prototype filter to N and the cutoff frequency to fc ; The β parameter is calculated by the following formula: Assume the length of the prototype filter is Where Δf is the normalized transition bandwidth, Calculate, f s2 is the stopband edge frequency, f p2 is the passband edge frequency, f s is the sampling frequency; the cutoff frequency of the prototype filter is

[0035] The prototype filter is decomposed into M sub-filters by polyphase decomposition. The coefficient of the sub-filter after polyphase decomposition is h k [n] = h[nM+k] (k = 0, 1, ..., M-1); The multiphase decomposition process is represented by matrix operations, assuming that the prototype filter coefficient matrix is ​​H = [h[0], h[1], ..., h[N-1]] T , then the sub-filter coefficient matrix H k Available through H k =P k H(k=0,1,…,M-1) is obtained, where P k is the multiphase decomposition matrix, whose elements are

[0036] The preprocessed zero intermediate frequency signal s zIF [n] is input to the polyphase filter bank for channelization processing and decomposed into various subchannels; the output of the kth subchannel is Assume that the input data sequence of the polyphase filter bank is X = [x[0], x[1], ..., x[L-1]] T ; L is the data length of the input data of the polyphase filter bank, and the subchannel output sequence is Y k =[y k [0],y k [1], …, y k [L-1]] T , then Y k =H k X, where H k is the coefficient matrix of the kth sub-filter, obtained by the above polyphase decomposition.

[0037] In this embodiment, first, according to the S-band narrowband signal characteristics and the established passband and stopband attenuation requirements, the Kaiser window function is selected to design the prototype filter. Carefully consider the signal spectrum specifications, accurately determine the passband and stopband edge frequencies, calculate the normalized transition bandwidth in combination with the sampling frequency, and then determine the key parameters of the Kaiser window function, and at the same time clarify the appropriate length and cutoff frequency of the prototype filter, so as to shape a prototype filter that can accurately screen frequency components and effectively suppress clutter. Then, matrix operations are used to implement the multiphase decomposition of the prototype filter, construct a specific multiphase decomposition matrix, follow the rigorous matrix multiplication rules, and disassemble multiple sub-filter coefficient matrices from the prototype filter coefficient matrix, successfully differentiate the number of sub-filters that match the system channelization requirements, optimize the processing structure, and create conditions for parallel and efficient signal processing. Finally, the zero intermediate frequency signal data sequence is input into the preset multi-phase filter group, and each sub-channel strictly follows the established operation rules and processes the input signal in an orderly manner based on the sub-filter coefficient matrix obtained by decomposition, so as to achieve the goal of accurately decomposing the zero intermediate frequency signal into different sub-channels, and give each sub-channel a differentiated center frequency and bandwidth, so that it can carry independent and clear signal components. In summary, the present embodiment can accurately divide the broadband signal into multiple narrowband sub-channels, ensuring good isolation between the sub-channels, thereby improving the accuracy of subsequent feature extraction and beamforming. In addition, the prototype filter design optimized by the Kaiser window function and its multi-phase decomposition strategy not only enhances the flexibility and adaptability of the system, but also ensures stable performance under different working conditions, providing solid technical support for efficient signal processing in complex electromagnetic environments.

[0038] As one example of the above embodiment, the deep belief network formed by stacking multiple restricted Boltzmann machines is used to perform deep extraction of signal features of each subchannel of the zero intermediate frequency signal to obtain signal features, including:

[0039] Construction and initialization of deep belief network, including:

[0040] The number of visible layer nodes M of the first restricted Boltzmann machine of the deep belief network is determined according to the number of subchannels; the number of subchannels is the same as the number of visible layer nodes; the number of hidden layer nodes of the first restricted Boltzmann machine is determined at the same time; the number of visible layer nodes of subsequent restricted Boltzmann machines of the deep belief network is successively equal to the number of hidden layer nodes of the previous restricted Boltzmann machine, and the number of hidden layer nodes decreases;

[0041] Randomly initialize the connection weights w in all restricted Boltzmann machines ij and the visible layer bias a i , hidden layer bias b j , connection weight w ijThe initialization range is set between [-0.1, 0.1], and the visible layer bias a i and the hidden layer bias b j Initialized to a value of [-0.05, 0.05];

[0042] Training of deep belief networks, including:

[0043] For the first restricted Boltzmann machine:

[0044] Input data preparation: The zero intermediate frequency signal y of each subchannel k [n] (k = 1, 2, ..., M) is organized into training data batches according to time series, each batch contains B time sample points, and the input matrix dimension is M × B;

[0045] Forward propagation calculates the hidden layer state probability: using the formula Calculate the activation probability of each node in the hidden layer, where is a dynamic adjustment factor related to the signal frequency. f k,j It represents the frequency component of the k-th subchannel associated with the j-th hidden layer node, reflecting the frequency correlation between a specific subchannel and the hidden layer node; represents the sum of the frequency components associated with all subchannels and all hidden layer nodes, which is used to calculate a single f k，j Normalize it so that The value of can be within a reasonable relative range. By combining the signal frequency factor, the role of the connection weight in calculating the activation probability of the hidden layer node is dynamically changed, allowing the deep belief network to learn and capture features based on the different frequency characteristics of the signal; ij is the connection weight between the visible layer and the hidden layer, v i Enter a value for the visible layer node, b j is the hidden layer node bias; this dynamic adjustment factor is used to dynamically adjust the effect of the connection weight according to the signal frequency characteristics, so that the network has different sensitivities to signals with different frequency components to mine signal features; where P(h j =1|v) represents the probability that the state of the jth node in the hidden layer takes the value of 1 under the condition of a given visible layer state vector v. It is an indicator to measure the possibility of the node being activated, and its value range is between 0 and 1. j represents the state variable of the jth node in the hidden layer, and its final value can only be 0 or 1; v is the visible layer state vector, and its dimension is the same as the number of visible layer nodes. The element v in the vector i Corresponding to the input value of the i-th node in the visible layer, this input value comes from the specific sample data of the zero intermediate frequency signal of each sub-channel, and is the basic data passed to the hidden layer for calculating the activation probability;

[0046] Hidden layer state sampling: According to the calculated probability P(h j =1|v) samples the hidden layer state h to generate a hidden layer state value of 0 or 1;

[0047] Back propagation reconstructs the visible layer: According to the hidden layer state h, through the formula Calculate the probability of reconstructing the visible layer state, and then sample to obtain the reconstructed visible layer state v′, where a i is the visible layer node bias, This factor comprehensively considers the influence of hidden layer node states and signal frequency on reconstruction so that the reconstruction process reflects signal characteristics;

[0048] Calculate the reconstruction error: Use the formula to calculate the reconstruction error E:

[0049]

[0050] , where v i，b and v′ i，b are the original and reconstructed values ​​of the visible layer node i at the bth sample point, h j，b is the actual state value of the jth node in the hidden layer at the bth sample point, h′ j，b is the state value of the hidden layer at the bth sample point calculated again according to the reconstructed visible layer state, and λ is the regularization parameter; the reconstruction error is weighted and adjusted from different angles through the exponential term and the cosine term. The exponential term dynamically scales the visible layer reconstruction error according to the hidden layer state difference, and the cosine term adjusts the regularization of the connection weight according to the average state of the hidden layer, so that the deep belief network can balance the fitting ability and generalization ability;

[0051] Connection weight and bias update: Use the gradient descent algorithm to update the connection weight w according to the following formula ij , visible layer bias a i and the hidden layer bias b j ;

[0052]

[0053] , Δa i =η( <v i >- <v′ i >), Δb j =η( <h j >- <h′ j 〉), where Δw ij Represents the connection weight w ij The update amount, that is, each time the training is iterated, the connection weight needs to be adjusted to change the value, and the current connection weight is updated by this update amount; <vi h j > represents the expectation of the product of the input value of the ith node in the visible layer and the state value of the ith node in the hidden layer over the entire training batch data. It is obtained by performing corresponding calculations and averaging the data of all sample points, reflecting the average performance of the association between the two nodes on the training data under the current network parameters; <v′ i h′ j > is the expectation of the product of the i-th node value of the reconstructed visible layer and the j-th node state value of the hidden layer calculated again based on the reconstructed visible layer, which is also calculated based on the average of all sample point data and used for <v i h j >Compare and reflect the differences before and after reconstruction to determine the direction and magnitude of the connection weight adjustment; It is an exponential adjustment term based on the difference between the hidden layer state before and after reconstruction, which dynamically adjusts the connection weight update amount; It is the cosine function term calculated by combining the original state and the reconstructed state of the hidden layer, and works together with the regularization parameter and the connection weight to further adjust the update of the connection weight according to the state of the hidden layer; η is the learning rate; Δa i Indicates the update amount of the bias of the i-th node in the visible layer, which determines the value of the bias that needs to be adjusted in each iteration and is used to gradually optimize the bias parameters; <v i > is the expectation of the input value of the i-th node in the visible layer over the entire training batch data, reflecting the average input data of the node; <v′ i >To reconstruct the expectation of the value of the i-th node in the visible layer over the entire training batch data, <v i >Comparison, reflecting the average difference between the visible layer nodes before and after reconstruction, in order to determine the direction and magnitude of the bias adjustment; Δb j Represents the update amount of the bias of the jth node in the hidden layer. Its function is to determine the specific value of the bias adjustment of the node in the hidden layer at each iteration; <h j > is the expectation of the state value of the jth node in the hidden layer on the entire training batch data, reflecting the average state of the node in the hidden layer on the training data; <h′ j > represents the expectation of the j-th node state value of the hidden layer calculated again based on the reconstructed visible layer over the entire training batch data, and <h j >Comparison reflects the average difference in the state of the node in the hidden layer before and after reconstruction, based on which the direction and magnitude of the adjustment of the hidden layer bias are determined, so that the network can continuously optimize the activation behavior of the hidden layer nodes during the training process to learn and mine signal features;

[0054] Subsequent restricted Boltzmann machine training: The hidden layer output obtained from the previous restricted Boltzmann machine training is used as the visible layer input of the next restricted Boltzmann machine, and the above training process is repeated until all restricted Boltzmann machines are trained. During the training process, the multipath effect characteristics are reflected through the frequency-related responses of the hidden layer nodes to different sub-channel signals at different time delays, and the interference signal characteristics are mined from the hidden layer's response to abnormal frequency components and special signal patterns.

[0055] The signal features of each sub-channel of the zero intermediate frequency signal are deeply extracted by using the trained deep belief network formed by stacking a plurality of restricted Boltzmann machines to obtain signal features.

[0056] In this embodiment, the problem of insufficient signal feature extraction and poor beamforming accuracy in the prior art is solved. The core lies in the use of deep belief networks (DBNs) to achieve efficient feature extraction. In the DBN construction stage, the number of visible layer nodes of the first restricted Boltzmann machine (RBM) is accurately set according to the actual number of sub-channels to ensure one-to-one correspondence with the sub-channels. The number of hidden layer nodes of the first RBM is carefully determined in combination with signal complexity and computing resources. Subsequent RBMs follow the visible layer node number connection rules and the hidden layer node reduction strategy to be stacked in order. At the same time, all connection weights are randomly initialized in the range of [-0.1, 0.1], and the bias is initialized to 0, laying the foundation for training. In the training process, for the first RBM, the zero intermediate frequency signals of each sub-channel are strictly organized into batches containing a specific number of time sample points in time sequence, and the input matrix dimension is determined accordingly. The forward propagation uses a dynamic adjustment factor formula that is deeply coupled with the signal frequency to calculate the activation probability of the hidden layer nodes. The factor considers the quantum channel-hidden layer frequency correlation and the sum of the normalized frequency components, so that the network can accurately capture the signal characteristics according to the frequency. After obtaining the hidden layer state based on probability sampling, back propagation reconstructs the visible layer according to the formula of comprehensive hidden layer state and frequency, and calculates the error using the reconstruction error formula that takes into account hidden layer state difference scaling and average state regularization. Gradient descent is used to update the connection weights, visible layer and hidden layer biases according to the complex formula of associated data expectation, dynamic adjustment terms and cosine function to ensure steady network optimization. Subsequent RBM iterative training is driven by the output of the previous hidden layer. The fully trained DBN can deeply analyze the signal and accurately extract the deep-level features covering the complex delay-frequency response of the multipath effect, abnormal frequency of the interference signal and special modes, and input key signal features for the subsequent particle swarm algorithm, which greatly helps to search for the optimal beamforming weight, significantly improve the S-band narrowband signal processing performance of the phased array system, enhance the beam pointing accuracy and anti-interference performance, and comprehensively optimize the system's comprehensive performance to adapt to complex electromagnetic environments.

[0057] As one example of the above embodiment, the signal feature is input into a pre-trained particle swarm optimization algorithm, each particle in the particle swarm optimization algorithm represents a set of beamforming weights, and the optimal beamforming weights are searched in the feature space through the position update and velocity update operations of the particles, including:

[0058] The particle swarm optimization algorithm is pre-trained in the following way:

[0059] Particle initialization process: First, determine the particle encoding method, that is, make it clear that each particle represents a set of beamforming weights, and the number of weights corresponds to the number of phased array antenna elements. Then, within a given reasonable value range, randomly initialize the position and speed of the particle, set the initial value of the individual optimal position of each particle to its current position, and randomly select the position of a particle as the initial value of the global optimal position, in preparation for subsequent iterative optimization based on the fitness function, including:

[0060] Determine particle encoding: Each particle in the particle swarm optimization algorithm represents a set of beamforming weights w p =[w p1 , w p2 ,…,w pN ] T , where N is the number of elements of the phased array antenna;

[0061] Randomly initialize particle positions and velocities: used as beamforming weights w p The particle position x p In [-w min , w max ] Random value within the range; speed v p In [-v min , v max ] is randomly set within the range; at the same time, the individual optimal position p of each particle is initialized p =x p , and randomly select the position of a particle as the initial value of the global optimal position g;

[0062] Fitness function design and calculation process: Construct a fitness function that integrates multiple indicators. This function is a weighted sum of related indicators such as beam pointing accuracy, sidelobe level suppression, signal-to-interference ratio, beam shape consistency, and beam energy concentration. By analyzing the beam signal synthesized based on the current beamforming weight, the corresponding data of beam pointing angle, main lobe power, sidelobe power, signal power, and interference signal power are obtained, and the specific values ​​are substituted into the calculation formula of each indicator to obtain the specific values, which are then multiplied by the corresponding weight coefficients and summed to obtain the fitness function value of each particle. This value is compared with the individual optimal fitness value of the particle and the global optimal fitness value. If it is better than the former, the corresponding optimal position is updated, thereby guiding the particle to search for a better solution in subsequent iterations, including:

[0063] Construct fitness function: F(w p )=αF 1 +βF 2 +γF 3 +δF 4 +∈F 5 , where α, β, γ, δ, ∈ are weight coefficients;

[0064] Calculate the beam pointing accuracy F 1 :The actual beam pointing angle θ is obtained through the spatial spectrum estimation algorithm a , and the desired beam pointing angle θ d In contrast, the formula Calculate the beam pointing accuracy, where K is the preset number of measurements and κ is the preset focusing factor; θ a，k is the actual beam pointing angle obtained by the kth measurement, which is calculated by analyzing the beam signal synthesized based on the current beamforming weights through the spatial spectrum estimation algorithm; θ d It is the preset expected beam pointing angle, which is the angle value of the ideal beam pointing direction pre-set according to system design or application requirements. It is used as a benchmark for comparison with the actual beam pointing angle to determine the error size of the beam pointing and guide the particles to adjust the beam forming weight during the iteration process.

[0065] Calculate the sidelobe level suppression F 2 :Based on the current beamforming weight w p The synthesized beam signal is subjected to power spectrum analysis to obtain the main lobe power P main and the sidelobe average power P sidelobe ,calculate ∈ is a preset minimum value to prevent the denominator from being 0, μ is the weight factor, and f sidelobe represents the sidelobe frequency range; P(f) is the power spectrum function of the beam signal, which represents the power value at frequency f;

[0066] Calculate the signal-to-interference ratio F3 :The signal power P is obtained by combining wavelet transform with energy detection, an interference detection and signal power estimation method based on time-frequency analysis. s and interference signal power P i ,calculate v is the enhancement coefficient, which is a preset minimum value to prevent the denominator from being 0;

[0067] Calculate beam shape consistency F 4 :Define the desired beam shape function G(f) by calculating the current beam forming weights w p The synthesized beam signal is subjected to power spectrum analysis to obtain the actual beam shape function P(f), and then calculated f total represents the total frequency range, ω is the shape adjustment frequency; G(f) is the power distribution function that the ideal beam should have at different frequency points, which is pre-set according to system design or application requirements and serves as a standard template for comparison with the actual beam shape;

[0068] Calculate the beam energy concentration F 5 : Calculate the total beam energy Main lobe energy f main represents the main lobe frequency range, then ρ is the preset enhancement factor;

[0069] Calculate the fitness function value F(w p ), and the individual optimal fitness value of the particle F(p p ) comparison, if F(w p )<F(p p ), then update the individual optimal position p p =x p ; At the same time, F(w p ) is compared with the global optimal fitness value F(g). If F(w p )<F(g), then update the global optimal position g=x p ;

[0070] Particle update process: According to the particle speed update formula, the new speed of the particle is calculated by combining the inertia weight, learning factor, random number and dynamic adjustment matrix related to the particle's historical position and speed; the inertia weight decreases with the number of iterations, so that the particle has a larger exploration range in the early stage and tends to search locally in the later stage; the learning factor controls the degree to which the particle approaches the individual optimal and global optimal positions; the dynamic adjustment matrix makes fine adjustments to the particle update direction according to the historical changes in the particle position and speed; then, according to the particle position update formula, the new speed is added to the current position to obtain the new position of the particle; the fitness function calculation and evaluation process and the particle update process are repeated after multiple iterations, which specifically include:

[0071] Particle velocity update: According to the formula Update the particle velocity, where ω is the inertia weight, initially set to 0.95, and Regularly decreasing, t is the current iteration number, T PSO is the total number of iterations; c 1 and c 2 is the learning factor; r 1 and r 2 is a random number between [0, 1]; ⊙ represents element-by-element multiplication, T 1 and T 2 are two matrices related to the particle position and velocity history, whose elements and Represent the jth element of the velocity and position of particle p at the tth iteration respectively; represents the velocity vector of particle p at the t+1th iteration, each dimension of which corresponds to the velocity of the particle in the direction of that dimension when the beamforming weights of the phased array antenna element are updated;

[0072] Particle position update: according to the formula Update particle positions;

[0073] Iterative optimization: Repeat the fitness function calculation and evaluation, particle update steps, and after a preset T PSO Iterations;

[0074] The signal features are input into a pre-trained particle swarm optimization algorithm. Through the particle position update and velocity update operations, as the iteration proceeds, the particles continuously search in the feature space and gradually move to the area with the optimal fitness function value. Finally, the optimal beamforming weight is searched in the feature space, so that the beam can achieve better performance in related performance indicators; the performance indicators include: pointing accuracy, sidelobe level suppression, anti-interference ability, beam shape and energy concentration.

[0075] In this embodiment, we focus on the optimization problem of beamforming weights in the phased array S-band narrowband digital processor, and propose an innovative strategy based on the particle swarm optimization (PSO) algorithm. In the particle initialization stage, the particle encoding is clearly defined in accordance with the number of phased array antenna elements, so that each particle accurately corresponds to a set of beamforming weights, and then the particle position is randomly assigned in the range of [-0.4, 0.4] to adapt to the S-band signal characteristics, and the speed is randomly set in the range of [-0.08, 0.08]. At the same time, the individual and global optimal position initial values ​​are reasonably set to lay the foundation for iteration. The fitness function construction comprehensively considers beam pointing accuracy, sidelobe level suppression, signal-to-interference ratio, beam shape consistency and energy concentration. With the help of multiple technical means such as spatial spectrum estimation, power spectrum analysis, wavelet transform combined with energy detection, the beam signal synthesized based on the current beam weight is deeply analyzed, key data such as beam pointing angle and power in each frequency band are accurately extracted, and the customized indicator calculation module is substituted to weigh the importance of each indicator and assign weight coefficients, and the fitness value is strictly synthesized to accurately guide the particle search direction, compare and update the individual and global optimal in real time, and drive the particles to strive for a better solution. In the particle update stage, a rigorous physical model is followed. The speed update formula integrates the inertia weight (initial 0.95 and linearly decreases with iteration, cleverly balancing global exploration and local deep cultivation), learning factors (finely controlling the rhythm of particles approaching the optimal position), random numbers (injecting randomness to break the local optimal trap) and a dynamic adjustment matrix based on the particle's historical trajectory (intelligently calibrating the update direction according to position and speed changes) to calculate the new speed; combined with the position update formula, the particles are prompted to move steadily. After multiple iterations (preset number of times), the fitness is continuously optimized. After the signal features extracted by the deep belief network are imported into the pre-trained PSO algorithm, the particles shuttle in the complex feature space to find the best. The technical effects of this embodiment are: breaking through the limitations of traditional beamforming, improving beam pointing accuracy, and accurately focusing on the target; being able to suppress sidelobe levels and reduce interference introduction; significantly enhancing anti-interference capabilities and improving signal stability; being able to form an ideal beam shape to suit a variety of application scenarios; being able to condense beam energy, improve transmission energy efficiency, and comprehensively optimize the signal processing and beamforming comprehensive performance of the phased array system in the S-band narrowband scenario.

[0076] As one example of the above embodiment, the step of performing weighted summation on the zero intermediate frequency signals of the sub-channels to achieve beam synthesis according to the optimal beam forming weights to obtain a digital beam synthesis signal includes:

[0077] (I) Beamforming step, including:

[0078] Beam signal calculation: Let the optimized beamforming weight be w opt =[w opt，1 , w opt，2 ,…,w opt，M ] T, the zero intermediate frequency signal of each subchannel is y k [n](k=1, 2, ..., M); synthesized beam signal Where L is the number of signal delay samples considered, f c is the center frequency, T s is the sampling period;

[0079] Distributed computing architecture application: Distributed computing architecture is used to divide the sub-channels into P groups, each group contains sub-channels, Indicates rounding up; suppose the pth processing unit is responsible for processing the weighted summation task of subchannel (p-1)Q+1 to min(pQ, M), p = 1, 2, ..., P; the intermediate result calculated by processing unit p Finally, the results of all processing units are added to obtain the final beam signal

[0080] (II) Dynamic beam adjustment steps include:

[0081] Signal strength monitoring: calculate the average power of the signal Where N s To calculate the number of samples for the average power, we compare P at different times. avg To monitor signal strength changes;

[0082] Interference signal spectrum analysis: Perform fast Fourier transform on the beam signal B[n] to obtain the spectrum Assume that the spectrum range of the normal signal is [f min , f max ], by detecting the obvious peaks outside this range in the spectrum, the existence of new interference sources and their frequency characteristics can be determined; if there is a frequency f i Satisfy S(f i )>T h and T h is the spectrum peak detection threshold, it is determined that a new interference source appears;

[0083] Beam optimization model updates, including:

[0084] Trigger condition judgment: When |P avg -P avg，prev |>3dB or when a new interference source is detected, restart the beam optimization model; P avg，prev is the average power at the previous moment;

[0085] Deep belief network feature extraction: preprocess and organize the current signal features as input data for the deep belief network; among them, the signal strength and spectrum energy distribution information in different frequency bands are combined into a feature vector in is the energy of frequency band k, obtained by integrating the spectrum in the corresponding frequency band; then the deep belief network is trained according to the training method, and the connection weights and biases are updated using the contrast divergence algorithm to extract the deep features of the signal; the signal features include: signal strength and interference spectrum;

[0086] Particle swarm optimization algorithm optimizes beamforming weights: Particle encoding is constructed in the training mode of particle swarm optimization algorithm, and each particle represents a new set of beamforming weights; when calculating the fitness function value, the current signal feature data and the following beamforming formula are used; where w p，k is the beamforming weight corresponding to particle p; after multiple iterations, the new optimal beamforming weight w is obtained new，opt ;

[0087] Beamforming update: According to the new optimal beamforming weights w new，opt Perform beam synthesis, that is, according to the beam synthesis formula Calculate new beam signals to achieve dynamic tracking and adaptive adjustment of the beam.

[0088] In this embodiment, the purpose is to achieve beam synthesis by weighted summing of the zero intermediate frequency signals of each subchannel through optimized beamforming weights, and then obtain a digital beam synthesis signal, and on this basis, introduce a dynamic beam adjustment mechanism to ensure that the beam maintains optimal performance in a complex electromagnetic environment. The specific working process is as follows: First, in the beam synthesis stage, according to the optimal beamforming weights obtained by the particle swarm optimization algorithm, the system weighted sums the zero intermediate frequency signals of each subchannel to generate a synthesized beam signal. In order to improve the computing efficiency, a distributed computing architecture is adopted, and the subchannel tasks are assigned to multiple processing units for parallel processing. Each processing unit is responsible for the weighted summation of a specific subchannel group, and finally the results of all processing units are summarized to complete the synthesis of the beam signal. Then, in the dynamic beam adjustment stage, the system continuously monitors the average power change of the beam signal, and analyzes the beam signal spectrum through fast Fourier transform (FFT) to identify possible new interference sources. Once a significant change in signal strength or a new interference source is detected, the update of the beam optimization model is triggered. During this process, the system uses a deep belief network (DBN) to deeply extract the current signal characteristics (such as signal strength and spectral energy distribution), and re-optimizes the beamforming weights through a particle swarm optimization algorithm to adapt to environmental changes. Finally, beamforming is performed again based on the newly optimized beamforming weights to ensure that the beam can track the target in real time and adjust adaptively to maintain the best signal reception quality. This technical solution not only achieves efficient and accurate beamforming, but also enhances the system's dynamic response capability and anti-interference performance, thereby greatly improving the application effect and reliability of the phased array S-band narrowband digital processor in complex electromagnetic environments.

[0089] See also Figure 2 , is a schematic diagram of the structure of a signal system of a phased array S-band narrowband digital processor provided by an embodiment of the present invention. The signal processing system of the phased array S-band narrowband digital processor includes:

[0090] The preprocessing module 10 is used to obtain the S-band narrowband service signal output by the integrated processor of the phased array system, and modulate and preprocess the service signal to obtain a preprocessed zero intermediate frequency signal;

[0091] A channelization processing module 11 is used to input the pre-processed zero intermediate frequency signal into a preset polyphase filter bank for channelization processing, so that the zero intermediate frequency signal is decomposed into different sub-channels, each sub-channel having a different center frequency and bandwidth;

[0092] The feature extraction module 12 is used to perform deep extraction of signal features of each subchannel of the zero intermediate frequency signal using a deep belief network formed by stacking multiple restricted Boltzmann machines to obtain signal features; wherein the number of nodes in the visible layer of the first restricted Boltzmann machine of the deep belief network is the same as the number of subchannels, and the number of nodes in the visible layer of subsequent restricted Boltzmann machines of the deep belief network decreases in sequence;

[0093] The optimization module 13 is used to input the signal features into a pre-trained particle swarm optimization algorithm, in which each particle represents a set of beamforming weights, and searches for the optimal beamforming weights in the feature space through particle position update and velocity update operations;

[0094] A beamforming module 14 is used to perform weighted summation on the zero intermediate frequency signals of the sub-channels according to the optimal beamforming weights to implement beamforming and obtain a digital beamforming signal;

[0095] The signal conversion module 15 is used to convert the digital beam synthesis signal into an analog intermediate frequency signal, and output the analog intermediate frequency signal to the transmitting and receiving TR component of the phased array system after preprocessing for subsequent transmission processing.

[0096] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0097] First, the S-band narrowband service signal transmitted by the integrated processor of the phased array system is accurately captured, and converted into a high-quality zero-IF signal through modulation and preprocessing operations; then the zero-IF signal is fed into a preset adaptive multi-phase filter group to achieve the purpose of channelization, so that the signal can be clearly divided into sub-channels with different characteristics; then a deep belief network composed of multiple restricted Boltzmann machines stacked in a specific structure is used, and the signal characteristics of each sub-channel are deeply analyzed by relying on the visible layer nodes and subsequent decreasing settings that match the first restricted Boltzmann machine with the number of sub-channels; then these characteristics are input into the pre-trained particle swarm optimization algorithm, and the dynamic update rules of the particle position and velocity are used to efficiently search for the optimal beamforming weights; finally, the zero-IF signals of each sub-channel are accurately weighted and summed according to this weight, and the beam synthesis is efficiently completed to obtain a digital beam synthesis signal, which is then output to the transmitting and receiving TR components of the phased array system through digital-to-analog conversion and necessary preprocessing for subsequent transmission processing. From the above analysis, it can be seen that the embodiments of the present invention introduce a deep belief network to perform deep extraction of signal features in complex environments, and use a particle swarm optimization algorithm to dynamically adjust the beamforming weights, thereby effectively solving the problems of insufficient processing performance and inaccurate beamforming in the prior art, and significantly improving the processing performance of the phased array system in S-band narrowband digital processing and the accuracy of beamforming.

[0098] As an improvement of the above solution, the channelization processing module is specifically used for:

[0099] The prototype filter is designed using the Kaiser window function and the maximum passband attenuation A p and stopband minimum attenuation A s Determine the β parameter of the Kaiser window function, and set the length of the prototype filter to N and the cutoff frequency to f c ; The β parameter is calculated by the following formula: Assume the length of the prototype filter is Where Δf is the normalized transition bandwidth, Calculate, f s2 is the stopband edge frequency, f p2 is the passband edge frequency, f s is the sampling frequency; the cutoff frequency of the prototype filter is

[0100] The prototype filter is decomposed into M sub-filters by polyphase decomposition. The coefficient of the sub-filter after polyphase decomposition is h k [n] = h[nM+k] (k = 0, 1, ..., M-1); The multiphase decomposition process is represented by matrix operations, assuming that the prototype filter coefficient matrix is ​​H = [h[0], h[1], ..., h[N-1]] T , then the sub-filter coefficient matrix H k Available through H k =P k H(k=0,1,…,M-1) is obtained, where P k is the multiphase decomposition matrix, whose elements are

[0101] The preprocessed zero intermediate frequency signal s zIF [n] is input to the polyphase filter bank for channelization processing and decomposed into various subchannels; the output of the kth subchannel is Assume that the input data sequence of the polyphase filter bank is X = [x[0], x[1], ..., x[L-1]] T ; L is the data length of the input data of the polyphase filter bank, and the subchannel output sequence is Y k =[y k [0],y k [1], …, y k [L-1]] T , then Y k =H k X, where H k is the coefficient matrix of the kth sub-filter, obtained by the above polyphase decomposition.

[0102] As an improvement of the above solution, the feature extraction module is specifically used for:

[0103] Construction and initialization of deep belief network, including:

[0104] The number of visible layer nodes M of the first restricted Boltzmann machine of the deep belief network is determined according to the number of subchannels; the number of subchannels is the same as the number of visible layer nodes; the number of hidden layer nodes of the first restricted Boltzmann machine is determined at the same time; the number of visible layer nodes of subsequent restricted Boltzmann machines of the deep belief network is successively equal to the number of hidden layer nodes of the previous restricted Boltzmann machine, and the number of hidden layer nodes decreases;

[0105] Randomly initialize the connection weights w in all restricted Boltzmann machines ij and the visible layer bias a i , hidden layer bias b j , connection weight w ij The initialization range is set between [-0.1, 0.1], and the visible layer bias a i and the hidden layer bias b j Initialized to a value of [-0.05, 0.05];

[0106] Training of deep belief networks, including:

[0107] For the first restricted Boltzmann machine:

[0108] Input data preparation: The zero intermediate frequency signal y of each subchannel k [n] (k = 1, 2, ..., M) is organized into training data batches according to time series, each batch contains B time sample points, and the input matrix dimension is M × B;

[0109] Forward propagation calculates the hidden layer state probability: using the formula Calculate the activation probability of each node in the hidden layer, where is a dynamic adjustment factor related to the signal frequency. f k,j It represents the frequency component of the k-th subchannel associated with the j-th hidden layer node, reflecting the frequency correlation between a specific subchannel and the hidden layer node; represents the sum of the frequency components associated with all subchannels and all hidden layer nodes, which is used to calculate a single f k，j Normalize it so that The value of can be within a reasonable relative range. By combining the signal frequency factor, the role of the connection weight in calculating the activation probability of the hidden layer node is dynamically changed, allowing the deep belief network to learn and capture features based on the different frequency characteristics of the signal; ij is the connection weight between the visible layer and the hidden layer, vi Enter a value for the visible layer node, b j is the hidden layer node bias; this dynamic adjustment factor is used to dynamically adjust the effect of the connection weight according to the signal frequency characteristics, so that the network has different sensitivities to signals with different frequency components to mine signal features; where P(h j =1|v) represents the probability that the state of the jth node in the hidden layer takes the value of 1 under the condition of a given visible layer state vector v. It is an indicator to measure the possibility of the node being activated, and its value range is between 0 and 1. j Represents the hidden layer j The state variable of a node can only take a final value of 0 or 1; V is the visible layer state vector, whose dimension is the same as the number of visible layer nodes. The element v in the vector i Corresponding to the input value of the i-th node in the visible layer, this input value comes from the specific sample data of the zero intermediate frequency signal of each sub-channel, and is the basic data passed to the hidden layer for calculating the activation probability;

[0110] Hidden layer state sampling: According to the calculated probability P(h j =1|v) samples the hidden layer state h to generate a hidden layer state value of 0 or 1;

[0111] Back propagation reconstructs the visible layer: According to the hidden layer state h, through the formula Calculate the probability of reconstructing the visible layer state, and then sample to obtain the reconstructed visible layer state v′, where a i is the visible layer node bias, This factor comprehensively considers the influence of hidden layer node states and signal frequency on reconstruction so that the reconstruction process reflects signal characteristics;

[0112] Calculate the reconstruction error: Use the formula to calculate the reconstruction error E:

[0113]

[0114] , where v i，b and v′ i，b are the original and reconstructed values ​​of the visible layer node i at the bth sample point, h j，b is the actual state value of the jth node in the hidden layer at the bth sample point, h′ j，b is the state value of the hidden layer at the bth sample point calculated again according to the reconstructed visible layer state, and λ is the regularization parameter; the reconstruction error is weighted and adjusted from different angles through the exponential term and the cosine term. The exponential term dynamically scales the visible layer reconstruction error according to the hidden layer state difference, and the cosine term adjusts the regularization of the connection weight according to the average state of the hidden layer, so that the deep belief network can balance the fitting ability and generalization ability;

[0115] Connection weight and bias update: Use the gradient descent algorithm to update the connection weight w according to the following formula ij , visible layer bias a i and the hidden layer bias b j ;

[0116]

[0117] , Δa i =η( <v i >- <v′ i >), Δb j =η( <h j >- <h′ j 〉), where Δw ij Represents the connection weight w ij The update amount, that is, each time the training is iterated, the connection weight needs to be adjusted to change the value, and the current connection weight is updated by this update amount; <v i h j > represents the expectation of the product of the input value of the ith node in the visible layer and the state value of the ith node in the hidden layer over the entire training batch data. It is obtained by performing corresponding calculations and averaging the data of all sample points, reflecting the average performance of the association between the two nodes on the training data under the current network parameters; <v′ i h′ j > is the expectation of the product of the i-th node value of the reconstructed visible layer and the j-th node state value of the hidden layer calculated again based on the reconstructed visible layer, which is also calculated based on the average of all sample point data and used for <v i h j >Compare and reflect the differences before and after reconstruction to determine the direction and magnitude of the connection weight adjustment; It is an exponential adjustment term based on the difference between the hidden layer state before and after reconstruction, which dynamically adjusts the connection weight update amount; It is the cosine function term calculated by combining the original state and the reconstructed state of the hidden layer, and works together with the regularization parameter and the connection weight to further adjust the update of the connection weight according to the state of the hidden layer; η is the learning rate; Δa i Indicates the update amount of the bias of the i-th node in the visible layer, which determines the value of the bias that needs to be adjusted in each iteration and is used to gradually optimize the bias parameters; <v i > is the expectation of the input value of the i-th node in the visible layer over the entire training batch data, reflecting the average input data of the node; <v i >To reconstruct the expectation of the value of the i-th node in the visible layer over the entire training batch data, <v i>Comparison reflects the average difference before and after reconstruction of the visible layer nodes, so as to determine the direction and magnitude of the bias adjustment; Δbj represents the update amount of the jth node bias in the hidden layer, which is used to determine the specific value of the bias adjustment of the node in the hidden layer at each iteration; <h j > is the expectation of the state value of the jth node in the hidden layer on the entire training batch data, reflecting the average state of the node in the hidden layer on the training data; <h′ j > represents the expectation of the j-th node state value of the hidden layer calculated again based on the reconstructed visible layer over the entire training batch data, and <h j >Comparison reflects the average difference in the state of the node in the hidden layer before and after reconstruction, based on which the direction and magnitude of the adjustment of the hidden layer bias are determined, so that the network can continuously optimize the activation behavior of the hidden layer nodes during the training process to learn and mine signal features;

[0118] Subsequent restricted Boltzmann machine training: The hidden layer output obtained from the previous restricted Boltzmann machine training is used as the visible layer input of the next restricted Boltzmann machine, and the above training process is repeated until all restricted Boltzmann machines are trained. During the training process, the multipath effect characteristics are reflected through the frequency-related responses of the hidden layer nodes to different sub-channel signals at different time delays, and the interference signal characteristics are mined from the hidden layer's response to abnormal frequency components and special signal patterns.

[0119] The signal features of each sub-channel of the zero intermediate frequency signal are deeply extracted by using the trained deep belief network formed by stacking a plurality of restricted Boltzmann machines to obtain signal features.

[0120] As an improvement of the above solution, the optimization module is specifically used for:

[0121] The particle swarm optimization algorithm is pre-trained in the following way:

[0122] Particle initialization process: First, determine the particle encoding method, that is, make it clear that each particle represents a set of beamforming weights, and the number of weights corresponds to the number of phased array antenna elements. Then, within a given reasonable value range, randomly initialize the position and speed of the particle, set the initial value of the individual optimal position of each particle to its current position, and randomly select the position of a particle as the initial value of the global optimal position, in preparation for subsequent iterative optimization based on the fitness function, including:

[0123] Determine particle encoding: Each particle in the particle swarm optimization algorithm represents a set of beamforming weights w p =[w p1 , w p2 ,…,w pN ] T, where N is the number of elements of the phased array antenna;

[0124] Randomly initialize particle positions and velocities: used as beamforming weights w p The particle position x p In [-w min , w max ] Random value within the range; speed v p In [-v min , v max ] is randomly set within the range; at the same time, the individual optimal position p of each particle is initialized p =x p , and randomly select the position of a particle as the initial value of the global optimal position g;

[0125] Fitness function design and calculation process: Construct a fitness function that integrates multiple indicators. This function is a weighted sum of related indicators such as beam pointing accuracy, sidelobe level suppression, signal-to-interference ratio, beam shape consistency, and beam energy concentration. By analyzing the beam signal synthesized based on the current beamforming weight, the corresponding data of beam pointing angle, main lobe power, sidelobe power, signal power, and interference signal power are obtained, and the specific values ​​are substituted into the calculation formula of each indicator to obtain the specific values, which are then multiplied by the corresponding weight coefficients and summed to obtain the fitness function value of each particle. This value is compared with the individual optimal fitness value of the particle and the global optimal fitness value. If it is better than the former, the corresponding optimal position is updated, thereby guiding the particle to search for a better solution in subsequent iterations, including:

[0126] Construct fitness function: F(w p )=αF 1 +βF 2 +γF 3 +δF 4 +∈F 5 , where α, β, γ, δ, ∈ are weight coefficients;

[0127] Calculate the beam pointing accuracy F 1 :The actual beam pointing angle θ is obtained through the spatial spectrum estimation algorithm a , and the desired beam pointing angle θ d In contrast, the formula Calculate the beam pointing accuracy, where K is the preset number of measurements and κ is the preset focusing factor; θ a，k is the actual beam pointing angle obtained by the kth measurement, which is calculated by analyzing the beam signal synthesized based on the current beamforming weights through the spatial spectrum estimation algorithm; θ dIt is the preset expected beam pointing angle, which is the angle value of the ideal beam pointing direction pre-set according to system design or application requirements. It is used as a benchmark for comparison with the actual beam pointing angle to determine the error size of the beam pointing and guide the particles to adjust the beam forming weight during the iteration process.

[0128] Calculate the sidelobe level suppression F 2 :Based on the current beamforming weight w p The synthesized beam signal is subjected to power spectrum analysis to obtain the main lobe power P main and the sidelobe average power P sidelobe ,calculate ∈ is a preset minimum value to prevent the denominator from being 0, μ is the weight factor, and f sidelobe represents the sidelobe frequency range; P(f) is the power spectrum function of the beam signal, which represents the power value at frequency f;

[0129] Calculate the signal-to-interference ratio F 3 :The signal power P is obtained by combining wavelet transform with energy detection, an interference detection and signal power estimation method based on time-frequency analysis. s and interference signal power P i ,calculate ν is the enhancement coefficient, which is a preset minimum value to prevent the denominator from being 0;

[0130] Calculate beam shape consistency F 4 :Define the desired beam shape function G(f) by calculating the current beam forming weights w p The synthesized beam signal is subjected to power spectrum analysis to obtain the actual beam shape function P(f), and then calculated f total represents the total frequency range, ω is the shape adjustment frequency; G(f) is the power distribution function that the ideal beam should have at different frequency points, which is pre-set according to system design or application requirements and serves as a standard template for comparison with the actual beam shape;

[0131] Calculate the beam energy concentration F 5 : Calculate the total beam energy Main lobe energy f main represents the main lobe frequency range, then ρ is the preset enhancement factor;

[0132] Calculate the fitness function value F(w p ), and the individual optimal fitness value of the particle F(p p ) comparison, if F(w p )<F(p p ), then update the individual optimal position p p =xp ; At the same time, F(w p ) is compared with the global optimal fitness value F(g). If F(w p )<F(g), then update the global optimal position g=x p ;

[0133] Particle update process: According to the particle speed update formula, the new speed of the particle is calculated by combining the inertia weight, learning factor, random number and dynamic adjustment matrix related to the particle's historical position and speed; the inertia weight decreases with the number of iterations, so that the particle has a larger exploration range in the early stage and tends to search locally in the later stage; the learning factor controls the degree to which the particle approaches the individual optimal and global optimal positions; the dynamic adjustment matrix makes fine adjustments to the particle update direction according to the historical changes in the particle position and speed; then, according to the particle position update formula, the new speed is added to the current position to obtain the new position of the particle; the fitness function calculation and evaluation process and the particle update process are repeated after multiple iterations, which specifically include:

[0134] Particle velocity update: According to the formula Update the particle velocity, where ω is the inertia weight, initially set to 0.95, and Regularly decreasing, t is the current iteration number, T PSO is the total number of iterations; c 1 and c 2 is the learning factor; r 1 and r 2 is a random number between [0, 1]; ⊙ represents element-by-element multiplication, T 1 and T 2 are two matrices related to the particle position and velocity history, whose elements and Represent the jth element of the velocity and position of particle p at the tth iteration respectively; represents the velocity vector of particle p at the t+1th iteration, each dimension of which corresponds to the velocity of the particle in the direction of that dimension when the beamforming weights of the phased array antenna element are updated;

[0135] Particle position update: according to the formula Update particle positions;

[0136] Iterative optimization: Repeat the fitness function calculation and evaluation, particle update steps, and after a preset T PSO Iterations;

[0137] The signal features are input into a pre-trained particle swarm optimization algorithm. Through the particle position update and velocity update operations, as the iteration proceeds, the particles continuously search in the feature space and gradually move to the area with the optimal fitness function value. Finally, the optimal beamforming weight is searched in the feature space, so that the beam can achieve better performance in related performance indicators; the performance indicators include: pointing accuracy, sidelobe level suppression, anti-interference ability, beam shape and energy concentration.

[0138] As an improvement of the above solution, the beamforming module is specifically used for:

[0139] 1. Beamforming

[0140] Beam signal calculation: Let the optimized beamforming weight be w opt =[w opt，1 , w opt，2 ,…,w opt，M ] T , the zero intermediate frequency signal of each sub-channel is y k [n](k=1, 2, ..., M); synthesized beam signal Where L is the number of signal delay samples considered, f c is the center frequency, T s is the sampling period;

[0141] Distributed computing architecture application: Distributed computing architecture is used to divide the sub-channels into P groups, each group contains sub-channels, Indicates rounding up; suppose the pth processing unit is responsible for processing the weighted summation task of subchannel (p-1)Q+1 to min(pQ, M), p = 1, 2, ..., P; the intermediate result calculated by processing unit p Finally, the results of all processing units are added to obtain the final beam signal

[0142] (II) Dynamic beam adjustment:

[0143] Signal strength monitoring: calculate the average power of the signal Where N s To calculate the number of samples for the average power, we compare P at different times. avg To monitor signal strength changes;

[0144] Interference signal spectrum analysis: Perform fast Fourier transform on the beam signal B[n] to obtain the spectrum Assume that the spectrum range of the normal signal is [f min , f max ], by detecting the obvious peaks outside this range in the spectrum, the existence of new interference sources and their frequency characteristics can be determined; if there is a frequency f iSatisfy S(f i )>T h and T h is the spectrum peak detection threshold, it is determined that a new interference source appears;

[0145] Beam optimization model updates, including:

[0146] Trigger condition judgment: When |P avg -P avg，prev |>3dB or when a new interference source is detected, restart the beam optimization model; P avg，prev is the average power at the previous moment;

[0147] Deep belief network feature extraction: preprocess and organize the current signal features as input data for the deep belief network; among them, the signal strength and spectrum energy distribution information in different frequency bands are combined into a feature vector in is the energy of frequency band k, obtained by integrating the spectrum in the corresponding frequency band; then the deep belief network is trained according to the training method, and the connection weights and biases are updated using the contrast divergence algorithm to extract the deep features of the signal; the signal features include: signal strength and interference spectrum;

[0148] Particle swarm optimization algorithm optimizes beamforming weights: Particle encoding is constructed in the training mode of particle swarm optimization algorithm, and each particle represents a new set of beamforming weights; when calculating the fitness function value, the current signal feature data and the following beamforming formula are used; where w p，k is the beamforming weight corresponding to particle p; after multiple iterations, the new optimal beamforming weight w is obtained new，opt ;

[0149] Beamforming update: According to the new optimal beamforming weights w new，opt Perform beam synthesis, that is, according to the beam synthesis formula Calculate new beam signals to achieve dynamic tracking and adaptive adjustment of the beam.

[0150] It should be noted that the system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art may understand and implement it without creative work.

[0151] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A signal processing method for an S-band narrowband digital processor of a phased array, characterized in that: include: Acquire an S-band narrowband service signal output by an integrated processor of the phased array system, and modulate and preprocess the service signal to obtain a preprocessed zero intermediate frequency signal; Inputting the preprocessed zero intermediate frequency signal into a preset polyphase filter bank for channelization processing, so that the zero intermediate frequency signal is decomposed into different sub-channels, each sub-channel having a different center frequency and bandwidth; A deep belief network formed by stacking a plurality of restricted Boltzmann machines is used to perform deep extraction of signal features of each subchannel of the zero intermediate frequency signal to obtain signal features; wherein the number of nodes in the visible layer of the first restricted Boltzmann machine of the deep belief network is the same as the number of subchannels, and the number of nodes in the visible layers of subsequent restricted Boltzmann machines of the deep belief network decreases in sequence; The signal features are input into a pre-trained particle swarm optimization algorithm, in which each particle represents a set of beamforming weights, and the optimal beamforming weights are searched in the feature space through particle position update and velocity update operations; According to the optimal beamforming weight, weighted summing is performed on the zero intermediate frequency signals of the sub-channels to implement beam synthesis, so as to obtain a digital beam synthesis signal; The digital beamforming signal is converted into an analog intermediate frequency signal, and the analog intermediate frequency signal is pre-processed and then output to the transmitting and receiving TR component of the phased array system for subsequent transmitting processing.

2. The signal processing method of the phased array S-band narrowband digital processor as claimed in claim 1, characterized in that: The pre-processed zero intermediate frequency signal is input into a preset polyphase filter bank for channelization processing, so that the zero intermediate frequency signal is decomposed into different sub-channels, each sub-channel having a different center frequency and bandwidth, including: The prototype filter is designed using the Kaiser window function and the maximum passband attenuation A p and stopband minimum attenuation A s Determine the β parameter of the Kaiser window function, and set the length of the prototype filter to N and the cutoff frequency to f c ; The β parameter is calculated by the following formula: Assume the length of the prototype filter is Where Δf is the normalized transition bandwidth, Calculate, f s2 is the stopband edge frequency, f p2 is the passband edge frequency, f s is the sampling frequency; the cutoff frequency of the prototype filter is The prototype filter is decomposed into M sub-filters by polyphase decomposition. The coefficient of the sub-filter after polyphase decomposition is h k [n] = h[nM+k] (k = 0, 1, ..., M-1); The multiphase decomposition process is represented by matrix operations, assuming that the prototype filter coefficient matrix is ​​H = [h[0], h[1], ..., h[N-1]] T , then the sub-filter coefficient matrix H k Available through H k =P k H(k=0,1,…,M-1) is obtained, where P k is the multiphase decomposition matrix, whose elements are The preprocessed zero intermediate frequency signal s zIF [n] is input to the polyphase filter bank for channelization processing and decomposed into various subchannels; the output of the kth subchannel is Assume that the input data sequence of the polyphase filter bank is X = [x[0], x[1], ..., x[L-1]] T ; L is the data length of the input data of the polyphase filter bank, and the subchannel output sequence is Y k =[y k [0],y k [1], …, y k [L-1]] T , then Y k =H k X, where H k is the coefficient matrix of the kth sub-filter, obtained by the above polyphase decomposition.

3. The signal processing method of the S-band narrowband digital processor of the phased array as claimed in claim 1, characterized in that: The deep belief network formed by stacking multiple restricted Boltzmann machines is used to perform deep extraction of signal features of each sub-channel of the zero intermediate frequency signal to obtain signal features, including: Construction and initialization of deep belief network, including: The number of visible layer nodes M of the first restricted Boltzmann machine of the deep belief network is determined according to the number of subchannels; the number of subchannels is the same as the number of visible layer nodes; the number of hidden layer nodes of the first restricted Boltzmann machine is determined at the same time; the number of visible layer nodes of subsequent restricted Boltzmann machines of the deep belief network is successively equal to the number of hidden layer nodes of the previous restricted Boltzmann machine, and the number of hidden layer nodes decreases; Randomly initialize the connection weights w in all restricted Boltzmann machines ij and the visible layer bias a i , hidden layer bias b j , connection weight w ij The initialization range is set between [-0.1, 0.1], and the visible layer bias a i and the hidden layer bias b j Initialized to a value of [-0.05, 0.05]; Training of deep belief networks, including: For the first restricted Boltzmann machine: Input data preparation: The zero intermediate frequency signal y of each subchannel k [n] (k = 1, 2, ..., M) is organized into training data batches according to time series, each batch contains B time sample points, and the input matrix dimension is M × B; Forward propagation calculates the hidden layer state probability: using the formula Calculate the activation probability of each node in the hidden layer, where is a dynamic adjustment factor related to the signal frequency. f k，j It represents the frequency component of the k-th subchannel associated with the j-th hidden layer node, reflecting the frequency correlation between a specific subchannel and the hidden layer node; represents the sum of the frequency components associated with all subchannels and all hidden layer nodes, which is used to calculate a single f k，j Normalize it so that The value of can be within a reasonable relative range. By combining the signal frequency factor, the role of the connection weight in calculating the activation probability of the hidden layer node is dynamically changed, allowing the deep belief network to learn and capture features based on the different frequency characteristics of the signal; ij is the connection weight between the visible layer and the hidden layer, υ i Enter a value for the visible layer node, b j is the hidden layer node bias; this dynamic adjustment factor is used to dynamically adjust the effect of the connection weight according to the signal frequency characteristics, so that the network has different sensitivities to signals with different frequency components to mine signal features; where P(h j =1|v) represents the probability that the state of the jth node in the hidden layer takes the value of 1 under the condition of a given visible layer state vector v. It is an indicator to measure the possibility of the node being activated, and its value range is between 0 and 1. j represents the state variable of the jth node in the hidden layer, and its final value can only be 0 or 1; v is the visible layer state vector, and its dimension is the same as the number of visible layer nodes. The element v in the vector i Corresponding to the input value of the i-th node in the visible layer, this input value comes from the specific sample data of the zero intermediate frequency signal of each sub-channel, and is the basic data passed to the hidden layer for calculating the activation probability; Hidden layer state sampling: According to the calculated probability P(h j =1|v) samples the hidden layer state h to generate a hidden layer state value of 0 or 1; Back propagation reconstructs the visible layer: According to the hidden layer state h, through the formula Calculate the probability of reconstructing the visible layer state, and then sample to obtain the reconstructed visible layer state v′, where a i is the visible layer node bias, This factor comprehensively considers the influence of hidden layer node states and signal frequency on reconstruction so that the reconstruction process reflects signal characteristics; Calculate the reconstruction error: Use the formula to calculate the reconstruction error E: , where v i，b and v′ i，b are the original and reconstructed values ​​of the visible layer node i at the bth sample point, h j，b is the actual state value of the jth node in the hidden layer at the bth sample point, h′ j，b is the state value of the hidden layer at the bth sample point calculated again according to the reconstructed visible layer state, and λ is the regularization parameter; the reconstruction error is weighted and adjusted from different angles through the exponential term and the cosine term. The exponential term dynamically scales the visible layer reconstruction error according to the hidden layer state difference, and the cosine term adjusts the regularization of the connection weight according to the average state of the hidden layer, so that the deep belief network can balance the fitting ability and generalization ability; Connection weight and bias update: Use the gradient descent algorithm to update the connection weight w according to the following formula ij , visible layer bias a i and the hidden layer bias b j ; , Δa i =η( <v i >- <v′ i >), Δb j =η( <h j >- <h′ j 〉), where Δw ij Represents the connection weight w ij The update amount, that is, each time the training is iterated, the connection weight needs to be adjusted to change the value, and the current connection weight is updated by this update amount; <v i h j > represents the expectation of the product of the input value of the ith node in the visible layer and the state value of the ith node in the hidden layer over the entire training batch data. It is obtained by performing corresponding calculations and averaging the data of all sample points, reflecting the average performance of the association between the two nodes on the training data under the current network parameters; <v′ i h j > is the expectation of the product of the i-th node value of the reconstructed visible layer and the j-th node state value of the hidden layer calculated again based on the reconstructed visible layer, which is also calculated based on the average of all sample point data and used for <v i h j >Compare and reflect the differences before and after reconstruction to determine the direction and magnitude of the connection weight adjustment; It is an exponential adjustment term based on the difference between the hidden layer state before and after reconstruction, which dynamically adjusts the connection weight update amount; It is the cosine function term calculated by combining the original state and the reconstructed state of the hidden layer, and works together with the regularization parameter and the connection weight to further adjust the update of the connection weight according to the state of the hidden layer; η is the learning rate; Δa i Indicates the update amount of the bias of the i-th node in the visible layer, which determines the value of the bias that needs to be adjusted in each iteration and is used to gradually optimize the bias parameters; <v i > is the expectation of the input value of the i-th node in the visible layer over the entire training batch data, reflecting the average input data of the node; <v′ i >To reconstruct the expectation of the value of the i-th node in the visible layer over the entire training batch data, <v i >Comparison, reflecting the average difference between the visible layer nodes before and after reconstruction, in order to determine the direction and magnitude of the bias adjustment; Δb j Represents the update amount of the bias of the jth node in the hidden layer. Its function is to determine the specific value of the bias adjustment of the node in the hidden layer at each iteration; <h j > is the expectation of the state value of the jth node in the hidden layer on the entire training batch data, reflecting the average state of the node in the hidden layer on the training data; <h′ j > represents the expectation of the j-th node state value of the hidden layer calculated again based on the reconstructed visible layer over the entire training batch data, and <h j >Comparison reflects the average difference in the state of the node in the hidden layer before and after reconstruction, based on which the direction and magnitude of the adjustment of the hidden layer bias are determined, so that the network can continuously optimize the activation behavior of the hidden layer nodes during the training process to learn and mine signal features; Subsequent restricted Boltzmann machine training: The hidden layer output obtained from the previous restricted Boltzmann machine training is used as the visible layer input of the next restricted Boltzmann machine, and the above training process is repeated until all restricted Boltzmann machines are trained. During the training process, the multipath effect characteristics are reflected through the frequency-related responses of the hidden layer nodes to different sub-channel signals at different time delays, and the interference signal characteristics are mined from the hidden layer's response to abnormal frequency components and special signal patterns. The signal features of each sub-channel of the zero intermediate frequency signal are deeply extracted by using the trained deep belief network formed by stacking a plurality of restricted Boltzmann machines to obtain signal features.

4. The signal processing method of the phased array S-band narrowband digital processor as claimed in claim 1, characterized in that: The signal feature is input into a pre-trained particle swarm optimization algorithm, in which each particle represents a set of beamforming weights, and the optimal beamforming weights are searched in the feature space through the position update and velocity update operations of the particles, including: The particle swarm optimization algorithm is pre-trained in the following way: Particle initialization process: First, determine the particle encoding method, that is, make it clear that each particle represents a set of beamforming weights, and the number of weights corresponds to the number of phased array antenna elements; then, within a given reasonable value range, randomly initialize the position and speed of the particle, set the initial value of the individual optimal position of each particle to its current position, and randomly select the position of a particle as the initial value of the global optimal position, in preparation for the subsequent iterative optimization based on the fitness function, specifically including: Determine particle encoding: Each particle in the particle swarm optimization algorithm represents a set of beamforming weights w p =[w p1 , w p2 ,…,w pN ] T , where N is the number of elements of the phased array antenna; Randomly initialize particle positions and velocities: used as beamforming weights w p The particle position x p In [-w min , w max ] Random value within the range; speed v p In [-v min , v max ] is randomly set within the range; at the same time, the individual optimal position p of each particle is initialized p =x p , and randomly select the position of a particle as the initial value of the global optimal position g; Fitness function design and calculation process: Construct a fitness function that integrates multiple indicators. This function is a weighted sum of related indicators such as beam pointing accuracy, sidelobe level suppression, signal-to-interference ratio, beam shape consistency, and beam energy concentration. By analyzing the beam signal synthesized based on the current beamforming weight, the corresponding data of beam pointing angle, main lobe power, sidelobe power, signal power, and interference signal power are obtained, and the specific values ​​are substituted into the calculation formula of each indicator to obtain the specific values, which are then multiplied by the corresponding weight coefficients and summed to obtain the fitness function value of each particle. This value is compared with the individual optimal fitness value of the particle and the global optimal fitness value. If it is better than the former, the corresponding optimal position is updated, thereby guiding the particle to search for a better solution in subsequent iterations, including: Construct fitness function: F(w p )=αF1+βF2+γF3+δF4+∈F5, where α, β, γ, δ, ∈ are weight coefficients; Calculate the beam pointing accuracy F1: Get the actual beam pointing angle θ through the spatial spectrum estimation algorithm a , and the desired beam pointing angle θ d In contrast, the formula Calculate the beam pointing accuracy, where K is the preset number of measurements and κ is the preset focusing factor; θ a，k is the actual beam pointing angle obtained by the kth measurement, which is calculated by analyzing the beam signal synthesized based on the current beamforming weights through the spatial spectrum estimation algorithm; θ d It is the preset expected beam pointing angle, which is the angle value of the ideal beam pointing direction pre-set according to system design or application requirements. It is used as a benchmark for comparison with the actual beam pointing angle to determine the error size of the beam pointing and guide the particles to adjust the beam forming weight during the iteration process. Calculate the sidelobe level suppression F2: Based on the current beamforming weight w p The synthesized beam signal is subjected to power spectrum analysis to obtain the main lobe power P main and the sidelobe average power P sidelone ,calculate ∈ is a preset minimum value to prevent the denominator from being 0, μ is the weight factor, and f sidelobe represents the sidelobe frequency range; P(f) is the power spectrum function of the beam signal, which represents the power value at frequency f; Calculate the signal-to-interference ratio F3: Use wavelet transform combined with energy detection, an interference detection and signal power estimation method based on time-frequency analysis, to obtain the signal power P s and interference signal power P i ,calculate ν is the enhancement coefficient, which is a preset minimum value to prevent the denominator from being 0; Calculate beam shape consistency F4: Define the desired beam shape function G(f) by calculating the beam shape consistency based on the current beamforming weights w p The synthesized beam signal is subjected to power spectrum analysis to obtain the actual beam shape function P(f), and then calculated f total represents the total frequency range, ω is the shape adjustment frequency; G(f) is the power distribution function that the ideal beam should have at different frequency points, which is pre-set according to system design or application requirements and serves as a standard template for comparison with the actual beam shape; Calculate beam energy concentration F5: Calculate total beam energy Main lobe energy f main represents the main lobe frequency range, then ρ is the preset enhancement factor; Calculate the fitness function value F(w p ), and compare it with the individual optimal fitness value F(p p ). If F(w p ) < F(p p ), then update the individual optimal position p p = x p ; At the same time, compare F(w p ) with the global optimal fitness value F(g). If F(w p ) < F(g), then update the global optimal position g = x p ; Particle update process: According to the particle speed update formula, the new speed of the particle is calculated by combining the inertia weight, learning factor, random number and dynamic adjustment matrix related to the particle's historical position and speed; the inertia weight decreases with the number of iterations, so that the particle has a larger exploration range in the early stage and tends to search locally in the later stage; the learning factor controls the degree to which the particle approaches the individual optimal and global optimal positions; the dynamic adjustment matrix makes fine adjustments to the particle update direction according to the historical changes in the particle position and speed; then, according to the particle position update formula, the new speed is added to the current position to obtain the new position of the particle; the fitness function calculation and evaluation process and the particle update process are repeated after multiple iterations, which specifically include: Particle velocity update: According to the formula Update the particle velocity, where ω is the inertia weight, initially set to 0.95, and Regularly decreasing, t is the current iteration number, T PSO is the total number of iterations; c1 and c2 are learning factors; r1 and r2 are random numbers between [0, 1]; ⊙ represents element-by-element multiplication, T1 and T2 are two matrices related to the particle position and velocity history, whose elements and Represent the jth element of the velocity and position of particle p at the tth iteration respectively; represents the velocity vector of particle p at the t+1th iteration, each dimension of which corresponds to the velocity of the particle in the direction of that dimension when the beamforming weights of the phased array antenna element are updated; Particle position update: According to the formula Update particle positions; Iterative optimization: Repeat the fitness function calculation and evaluation, particle update steps, and after a preset T PSO Iterations; The signal features are input into a pre-trained particle swarm optimization algorithm. Through the particle position update and velocity update operations, as the iteration proceeds, the particles continuously search in the feature space and gradually move to the area with the optimal fitness function value. Finally, the optimal beamforming weight is searched in the feature space, so that the beam can achieve better performance in related performance indicators; the performance indicators include: pointing accuracy, sidelobe level suppression, anti-interference ability, beam shape and energy concentration.

5. The signal processing method of the phased array S-band narrowband digital processor as claimed in claim 4, characterized in that: The step of performing weighted summation on the zero intermediate frequency signals of the sub-channels according to the optimal beamforming weights to achieve beam synthesis and obtain a digital beam synthesis signal includes: (I) Beamforming step, including: Beam signal calculation: Let the optimized beamforming weight be w opt =[w opt，1 , w opt，2 ,…,w opt，M ] T , the zero intermediate frequency signal of each subchannel is y k [n](k=1, 2, ..., M); synthesized beam signal Where L is the number of signal delay samples considered, f c is the center frequency, T s is the sampling period; Distributed computing architecture application: Using distributed computing architecture, the sub-channels are divided into P groups, each group contains sub-channels, Indicates rounding up; suppose the pth processing unit is responsible for processing the weighted summation task of subchannel (p-1)Q+1 to min(pQ, M), p = 1, 2, ..., P; the intermediate result calculated by processing unit p Finally, the results of all processing units are added to obtain the final beam signal (II) Dynamic beam adjustment steps include: Signal strength monitoring: calculate the average power of the signal Where N s To calculate the number of samples for the average power, we compare P at different times. avg To monitor signal strength changes; Interference signal spectrum analysis: Perform fast Fourier transform on the beam signal B[n] to obtain the spectrum Assume that the spectrum range of the normal signal is [f min , f max ], by detecting the obvious peaks outside this range in the spectrum, the existence of new interference sources and their frequency characteristics can be determined; if there is a frequency f i Satisfy S(f i )>T h and T h is the spectrum peak detection threshold, it is determined that a new interference source appears; Beam optimization model updates, including: Trigger condition judgment: When |P avg -P avg,prev |>3dB or when a new interference source is detected, restart the beam optimization model; P avg，prev is the average power at the previous moment; Deep belief network feature extraction: preprocess and organize the current signal features as input data for the deep belief network; among them, the signal strength and spectrum energy distribution information in different frequency bands are combined into a feature vector in is the energy of frequency band k, obtained by integrating the spectrum in the corresponding frequency band; then the deep belief network is trained according to the training method, and the connection weights and biases are updated using the contrast divergence algorithm to extract the deep features of the signal; the signal features include: signal strength and interference spectrum; Particle swarm optimization algorithm optimizes beamforming weights: Particle encoding is constructed in the training mode of particle swarm optimization algorithm, and each particle represents a new set of beamforming weights; when calculating the fitness function value, the current signal feature data and the following beamforming formula are used; where w p，k is the beamforming weight corresponding to particle p; after multiple iterations, the new optimal beamforming weight w is obtained new，opt ; Beamforming update: According to the new optimal beamforming weights w new，opt Perform beam synthesis, that is, according to the beam synthesis formula Calculate new beam signals to achieve dynamic tracking and adaptive adjustment of the beam.

6. A signal processing system of a phased array S-band narrowband digital processor, characterized in that: include: A preprocessing module is used to obtain the S-band narrowband service signal output by the integrated processor of the phased array system, and modulate and preprocess the service signal to obtain a preprocessed zero intermediate frequency signal; A channelization processing module, used for inputting the preprocessed zero intermediate frequency signal into a preset polyphase filter bank for channelization processing, so that the zero intermediate frequency signal is decomposed into different sub-channels, each sub-channel having a different center frequency and bandwidth; A feature extraction module, used to perform deep extraction of signal features of each subchannel of the zero intermediate frequency signal using a deep belief network formed by stacking multiple restricted Boltzmann machines to obtain signal features; wherein the number of nodes in the visible layer of the first restricted Boltzmann machine of the deep belief network is the same as the number of subchannels, and the number of nodes in the visible layers of subsequent restricted Boltzmann machines of the deep belief network decreases in sequence; An optimization module, used for inputting the signal features into a pre-trained particle swarm optimization algorithm, in which each particle represents a set of beamforming weights, and searching for the optimal beamforming weights in the feature space through particle position update and velocity update operations; A beamforming module, configured to perform weighted summation on the zero intermediate frequency signals of the sub-channels according to an optimal beamforming weight to implement beamforming and obtain a digital beamforming signal; The signal conversion module is used to convert the digital beam synthesis signal into an analog intermediate frequency signal, and output the analog intermediate frequency signal to the transmitting and receiving TR component of the phased array system after preprocessing for subsequent transmission processing.

7. The signal processing system of the phased array S-band narrowband digital processor as claimed in claim 6, characterized in that: The channelization processing module is specifically used for: The prototype filter is designed using the Kaiser window function and the maximum passband attenuation A p and stopband minimum attenuation A s Determine the β parameter of the Kaiser window function, and set the length of the prototype filter to N and the cutoff frequency to f c ; The β parameter is calculated by the following formula: Assume the length of the prototype filter is Where Δf is the normalized transition bandwidth, Calculate, f s2 is the stopband edge frequency, f p2 is the passband edge frequency, f s is the sampling frequency; the cutoff frequency of the prototype filter is The prototype filter is decomposed into M sub-filters by polyphase decomposition. The coefficient of the sub-filter after polyphase decomposition is h k [n] = h[nM+k] (k = 0, 1, ..., M-1); The multiphase decomposition process is represented by matrix operations, assuming that the prototype filter coefficient matrix is ​​H = [h[0], h[1], ..., h[N-1]] T , then the sub-filter coefficient matrix H k Available through H k =P k H(k=0,1,…,M-1) is obtained, where P k is the multiphase decomposition matrix, whose elements are The preprocessed zero intermediate frequency signal s zIF [n] is input to the polyphase filter bank for channelization processing and decomposed into various subchannels; the output of the kth subchannel is Assume that the input data sequence of the polyphase filter bank is X = [x[0], x[1], ..., x[L-1]] T ; L is the data length of the input data of the polyphase filter bank, and the subchannel output sequence is Y k =[y k [0],y k [1], …, y k [L-1]] T , then Y k =H k X, where H k is the coefficient matrix of the kth sub-filter, obtained by the above polyphase decomposition.

8. The signal processing system of the phased array S-band narrowband digital processor as claimed in claim 6, characterized in that: The feature extraction module is specifically used for: Construction and initialization of deep belief network, including: The number of visible layer nodes M of the first restricted Boltzmann machine of the deep belief network is determined according to the number of subchannels; the number of subchannels is the same as the number of visible layer nodes; the number of hidden layer nodes of the first restricted Boltzmann machine is determined at the same time; the number of visible layer nodes of subsequent restricted Boltzmann machines of the deep belief network is successively equal to the number of hidden layer nodes of the previous restricted Boltzmann machine, and the number of hidden layer nodes decreases; Randomly initialize the connection weights w in all restricted Boltzmann machines ij and the visible layer bias a i , hidden layer bias b j , connection weight w ij The initialization range is set between [-0.1, 0.1], and the visible layer bias a i and the hidden layer bias b j Initialized to a value of [-0.05, 0.05]; Training of deep belief networks, including: For the first restricted Boltzmann machine: Input data preparation: The zero intermediate frequency signal y of each subchannel k [n] (k = 1, 2, ..., M) is organized into training data batches according to time series, each batch contains B time sample points, and the input matrix dimension is M × B; Forward propagation calculates the hidden layer state probability: using the formula Calculate the activation probability of each node in the hidden layer, where is a dynamic adjustment factor related to the signal frequency. f k,j It represents the frequency component of the k-th subchannel associated with the j-th hidden layer node, reflecting the frequency correlation between a specific subchannel and the hidden layer node; represents the sum of the frequency components associated with all subchannels and all hidden layer nodes, which is used to calculate a single f k，j Normalize it so that The value of can be within a reasonable relative range. By combining the signal frequency factor, the role of the connection weight in calculating the activation probability of the hidden layer node is dynamically changed, allowing the deep belief network to learn and capture features based on the different frequency characteristics of the signal; ij is the connection weight between the visible layer and the hidden layer, v i Enter a value for the visible layer node, b j is the hidden layer node bias; this dynamic adjustment factor is used to dynamically adjust the effect of the connection weight according to the signal frequency characteristics, so that the network has different sensitivities to signals with different frequency components to mine signal features; where P(h j =1|v) represents the probability that the state of the jth node in the hidden layer takes the value of 1 under the condition of a given visible layer state vector v. It is an indicator to measure the possibility of the node being activated, and its value range is between 0 and 1. j represents the state variable of the jth node in the hidden layer, and its final value can only be 0 or 1; v is the visible layer state vector, and its dimension is the same as the number of visible layer nodes. The element v in the vector i Corresponding to the input value of the i-th node in the visible layer, this input value comes from the specific sample data of the zero intermediate frequency signal of each sub-channel, and is the basic data passed to the hidden layer for calculating the activation probability; Hidden layer state sampling: According to the calculated probability P(h j =1|v) samples the hidden layer state h to generate a hidden layer state value of 0 or 1; Back propagation reconstructs the visible layer: According to the hidden layer state h, through the formula Calculate the probability of reconstructing the visible layer state, and then sample to obtain the reconstructed visible layer state v′, where a i is the visible layer node bias, This factor comprehensively considers the influence of hidden layer node states and signal frequency on reconstruction so that the reconstruction process reflects signal characteristics; Calculate the reconstruction error: Use the formula to calculate the reconstruction error E: , where v i，b and v′ i，b are the original and reconstructed values ​​of the visible layer node i at the bth sample point, h j，b is the actual state value of the jth node in the hidden layer at the bth sample point, h′ j，b is the state value of the hidden layer at the bth sample point calculated again according to the reconstructed visible layer state, and λ is the regularization parameter; the reconstruction error is weighted and adjusted from different angles through the exponential term and the cosine term. The exponential term dynamically scales the visible layer reconstruction error according to the hidden layer state difference, and the cosine term adjusts the regularization of the connection weight according to the average state of the hidden layer, so that the deep belief network can balance the fitting ability and generalization ability; Connection weight and bias update: Use the gradient descent algorithm to update the connection weight w according to the following formula ij , visible layer bias a i and the hidden layer bias b j ; , Δa i =η( <v i >- <v′ i >), Δb j =η( <h j >- <h′ j 〉), where Δw ij Represents the connection weight w ij The update amount, that is, each time the training is iterated, the connection weight needs to be adjusted to change the value, and the current connection weight is updated by this update amount; <v i h j > represents the expectation of the product of the input value of the ith node in the visible layer and the state value of the ith node in the hidden layer over the entire training batch data. It is obtained by performing corresponding calculations and averaging the data of all sample points, reflecting the average performance of the association between the two nodes on the training data under the current network parameters; <v′ i h′ j > is the expectation of the product of the i-th node value of the reconstructed visible layer and the j-th node state value of the hidden layer calculated again based on the reconstructed visible layer, which is also calculated based on the average of all sample point data and used for <v i h j >Compare and reflect the differences before and after reconstruction to determine the direction and magnitude of the connection weight adjustment; It is an exponential adjustment term based on the difference between the hidden layer state before and after reconstruction, which dynamically adjusts the connection weight update amount; It is the cosine function term calculated by combining the original state and the reconstructed state of the hidden layer, and works together with the regularization parameter and the connection weight to further adjust the update of the connection weight according to the state of the hidden layer; η is the learning rate; Δa i Indicates the update amount of the bias of the i-th node in the visible layer, which determines the value of the bias that needs to be adjusted in each iteration and is used to gradually optimize the bias parameters; <v i > is the expectation of the input value of the i-th node in the visible layer over the entire training batch data, reflecting the average input data of the node; <v′ i >To reconstruct the expectation of the value of the i-th node in the visible layer over the entire training batch data, <v i >Comparison, reflecting the average difference between the visible layer nodes before and after reconstruction, in order to determine the direction and magnitude of the bias adjustment; Δb j Represents the update amount of the bias of the jth node in the hidden layer. Its function is to determine the specific value of the bias adjustment of the node in the hidden layer at each iteration; <h j > is the expectation of the state value of the jth node in the hidden layer on the entire training batch data, reflecting the average state of the node in the hidden layer on the training data; <h′ j > represents the expectation of the j-th node state value of the hidden layer calculated again based on the reconstructed visible layer over the entire training batch data, and <h j >Comparison reflects the average difference in the state of the node in the hidden layer before and after reconstruction, based on which the direction and magnitude of the adjustment of the hidden layer bias are determined, so that the network can continuously optimize the activation behavior of the hidden layer nodes during the training process to learn and mine signal features; Subsequent restricted Boltzmann machine training: The hidden layer output obtained from the previous restricted Boltzmann machine training is used as the visible layer input of the next restricted Boltzmann machine, and the above training process is repeated until all restricted Boltzmann machines are trained. During the training process, the multipath effect characteristics are reflected through the frequency-related responses of the hidden layer nodes to different sub-channel signals at different time delays, and the interference signal characteristics are mined from the hidden layer's response to abnormal frequency components and special signal patterns. The signal features of each sub-channel of the zero intermediate frequency signal are deeply extracted by using the trained deep belief network formed by stacking a plurality of restricted Boltzmann machines to obtain signal features.

9. The signal processing system of the phased array S-band narrowband digital processor as claimed in claim 6, characterized in that: The optimization module is specifically used for: The particle swarm optimization algorithm is pre-trained in the following way: Particle initialization process: First, determine the particle encoding method, that is, make it clear that each particle represents a set of beamforming weights, and the number of weights corresponds to the number of phased array antenna elements; then, within a given reasonable value range, randomly initialize the position and speed of the particle, set the initial value of the individual optimal position of each particle to its current position, and randomly select the position of a particle as the initial value of the global optimal position, in preparation for the subsequent iterative optimization based on the fitness function, specifically including: Determine particle encoding: Each particle in the particle swarm optimization algorithm represents a set of beamforming weights w p =[w p1 , w p2 ,…,w pN ] T , where N is the number of elements of the phased array antenna; Randomly initialize particle positions and velocities: used as beamforming weights w p The particle position x p In [-w min , w max ] Random value within the range; speed v p In [-v min , v max ] is randomly set within the range; at the same time, the individual optimal position p of each particle is initialized p =x p , and randomly select the position of a particle as the initial value of the global optimal position g; Fitness function design and calculation process: Construct a fitness function that integrates multiple indicators. This function is a weighted sum of related indicators such as beam pointing accuracy, sidelobe level suppression, signal-to-interference ratio, beam shape consistency, and beam energy concentration. By analyzing the beam signal synthesized based on the current beamforming weight, the corresponding data of beam pointing angle, main lobe power, sidelobe power, signal power, and interference signal power are obtained, and the specific values ​​are substituted into the calculation formula of each indicator to obtain the specific values, which are then multiplied by the corresponding weight coefficients and summed to obtain the fitness function value of each particle. This value is compared with the individual optimal fitness value of the particle and the global optimal fitness value. If it is better than the former, the corresponding optimal position is updated, thereby guiding the particle to search for a better solution in subsequent iterations, including: Construct fitness function: F(w p )=αF1+βF2+γF3+δF4+∈F5, where α, β, γ, δ, ∈ are weight coefficients; Calculate the beam pointing accuracy F1: Get the actual beam pointing angle θ through the spatial spectrum estimation algorithm a , and the desired beam pointing angle θ d In contrast, the formula Calculate the beam pointing accuracy, where K is the preset number of measurements and κ is the preset focusing factor; θ a，k is the actual beam pointing angle obtained by the kth measurement, which is calculated by analyzing the beam signal synthesized based on the current beamforming weights through the spatial spectrum estimation algorithm; θ d It is the preset expected beam pointing angle, which is the angle value of the ideal beam pointing direction pre-set according to system design or application requirements. It is used as a benchmark for comparison with the actual beam pointing angle to determine the error size of the beam pointing and guide the particles to adjust the beam forming weight during the iteration process. Calculate the sidelobe level suppression F2: Based on the current beamforming weight w p The synthesized beam signal is subjected to power spectrum analysis to obtain the main lobe power P main and the sidelobe average power P sidlobe ,calculate ∈ is a preset minimum value to prevent the denominator from being 0, μ is the weight factor, and f sidelobe represents the sidelobe frequency range; P(f) is the power spectrum function of the beam signal, which represents the power value at frequency f; Calculate the signal-to-interference ratio F3: Use wavelet transform combined with energy detection, an interference detection and signal power estimation method based on time-frequency analysis, to obtain the signal power P s and interference signal power P i ,calculate ν is the enhancement coefficient, which is a preset minimum value to prevent the denominator from being 0; Calculate beam shape consistency F4: Define the desired beam shape function G(f) by calculating the beam shape consistency based on the current beamforming weights w p The synthesized beam signal is subjected to power spectrum analysis to obtain the actual beam shape function P(f), and then calculated f totdl represents the total frequency range, ω is the shape adjustment frequency; G(f) is the power distribution function that the ideal beam should have at different frequency points, which is pre-set according to system design or application requirements and serves as a standard template for comparison with the actual beam shape; Calculate beam energy concentration F5: Calculate total beam energy Main lobe energy f main represents the main lobe frequency range, then ρ is the preset enhancement factor; Calculate the fitness function value F(w p ), and compare it with the individual optimal fitness value F(p p ). If F(w p ) < F(p p ), then update the individual optimal position p p = x p ; At the same time, compare F(w p ) with the global optimal fitness value F(g). If F(w p ) < F(g), then update the global optimal position g = x p ; Particle update process: According to the particle speed update formula, the new speed of the particle is calculated by combining the inertia weight, learning factor, random number and dynamic adjustment matrix related to the particle's historical position and speed; the inertia weight decreases with the number of iterations, so that the particle has a larger exploration range in the early stage and tends to search locally in the later stage; the learning factor controls the degree to which the particle approaches the individual optimal and global optimal positions; the dynamic adjustment matrix makes fine adjustments to the particle update direction according to the historical changes in the particle position and speed; then, according to the particle position update formula, the new speed is added to the current position to obtain the new position of the particle; the fitness function calculation and evaluation process and the particle update process are repeated after multiple iterations, which specifically include: Particle velocity update: According to the formula Update the particle velocity, where ω is the inertia weight, initially set to 0.95, and Regularly decreasing, t is the current iteration number, T PSO is the total number of iterations; c1 and c2 are learning factors; r1 and r2 are random numbers between [0, 1]; ⊙: represents element-by-element multiplication, T1 and T2 are two matrices related to the particle position and velocity history, whose elements and Represent the jth element of the velocity and position of particle p at the tth iteration respectively; represents the velocity vector of particle p at the t+1th iteration, each dimension of which corresponds to the velocity of the particle in the direction of that dimension when the beamforming weights of the phased array antenna element are updated; Particle position update: According to the formula Update particle positions; Iterative optimization: Repeat the fitness function calculation and evaluation, particle update steps, and after a preset T PSO Iterations; The signal features are input into a pre-trained particle swarm optimization algorithm. Through the particle position update and velocity update operations, as the iteration proceeds, the particles continuously search in the feature space and gradually move to the area with the optimal fitness function value. Finally, the optimal beamforming weight is searched in the feature space, so that the beam can achieve better performance in related performance indicators; the performance indicators include: pointing accuracy, sidelobe level suppression, anti-interference ability, beam shape and energy concentration.

10. The signal processing system of the phased array S-band narrowband digital processor as claimed in claim 9, characterized in that: The beamforming module is specifically used for:

1. Beamforming Beam signal calculation: Let the optimized beamforming weight be w opt =[w opt，1 , w opt，2 ,…,w opt，M ] T , the zero intermediate frequency signal of each subchannel is y k [n](k=1, 2, ..., M); synthesized beam signal Where L is the number of signal delay samples considered, f c is the center frequency, T s is the sampling period; Distributed computing architecture application: Using distributed computing architecture, the sub-channels are divided into P groups, each group contains sub-channels, Indicates rounding up; suppose the pth processing unit is responsible for processing the weighted summation task of subchannel (p-1)Q+1 to min(pQ, M), p = 1, 2, ..., P; The intermediate results calculated by processing unit p Finally, the results of all processing units are added to obtain the final beam signal (II) Dynamic beam adjustment: Signal strength monitoring: calculate the average power of the signal Where N s To calculate the number of samples for the average power, we compare P at different times. avg To monitor signal strength changes; Interference signal spectrum analysis: Perform fast Fourier transform on the beam signal B[n] to obtain the spectrum Assume that the spectrum range of the normal signal is [f min , f max ], by detecting the obvious peaks outside this range in the spectrum, the existence of new interference sources and their frequency characteristics can be determined; if there is a frequency f i Satisfy S(f i )>T h and T h is the spectrum peak detection threshold, it is determined that a new interference source appears; Beam optimization model updates, including: Trigger condition judgment: When |P avg -P avg，pre |>3dB or when a new interference source is detected, restart the beam optimization model; P avg,prev is the average power at the previous moment; Deep belief network feature extraction: preprocess and organize the current signal features as input data for the deep belief network; among them, the signal strength and spectrum energy distribution information in different frequency bands are combined into a feature vector in is the energy of frequency band k, obtained by integrating the spectrum in the corresponding frequency band; then the deep belief network is trained according to the training method, and the connection weights and biases are updated using the contrast divergence algorithm to extract the deep features of the signal; the signal features include: signal strength and interference spectrum; Particle swarm optimization algorithm optimizes beamforming weights: Particle encoding is constructed in the training mode of particle swarm optimization algorithm, and each particle represents a new set of beamforming weights; when calculating the fitness function value, the current signal feature data and the following beamforming formula are used; where w p,k is the beamforming weight corresponding to particle p; after multiple iterations, the new optimal beamforming weight w is obtained new，opt ; Beamforming update: According to the new optimal beamforming weight W new，opt Perform beam synthesis, that is, according to the beam synthesis formula Calculate new beam signals to achieve dynamic tracking and adaptive adjustment of the beam.

Images