Large neural network decomposition optimization method based on multiband decomposition and waveform shaping
By decomposing large neural networks into multiple small networks and decomposing and processing signals in the frequency domain, the problem of high computing complexity and resource requirements in the communication system is solved, and the effect of reducing training complexity and energy consumption is achieved.
Patent Information
- Application Number
- CN202510043777.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-10
- Publication Date
- 2025-05-23
AI Technical Summary
Large neural networks have high computational complexity and resource requirements in communication systems, resulting in long training time and high energy consumption.
The large neural network decomposition optimization method based on multi-band decomposition and waveform shaping is adopted to decompose the large neural network into multiple small networks, and the signals are decomposed in the frequency domain through filtered block neural networks, and each sub-band signal is independently processed and trained.
It significantly reduces the training complexity of neural networks in large-scale multi-carrier systems, reduces resource energy consumption requirements, and improves the processing efficiency of neural networks.
Smart Images
Figure CN120031094A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of neural network decomposition optimization, and in particular to a large neural network decomposition optimization method based on multi-band decomposition and waveform shaping. Background Art
[0002] The application of large neural networks (LNN) in communication systems has become a research hotspot in recent years. They use the powerful nonlinear modeling capabilities of deep learning to optimize multiple modules in communication systems and improve system performance and efficiency. The application of large neural networks in communication systems has demonstrated its great potential in complex channels, dynamic environments, and efficient optimization. However, in current wireless communications, the number of carriers carrying data in the wireless signals of actual systems is huge. Such a large-scale system will bring significant training complexity and require a lot of time and hardware resources to complete neural network training.
[0003] To this end, we designed a large-scale neural network decomposition optimization method based on multi-band decomposition and waveform shaping to solve the above problems. Summary of the invention
[0004] The purpose of the present invention is to solve the shortcomings of large neural networks in the prior art in the communication system, such as high computational complexity and high resource requirements. A large neural network decomposition and optimization method based on multi-band decomposition and waveform shaping is proposed, and a filtering block neural network is designed. The large neural network can be decomposed into multiple small networks, which effectively reduces the computational complexity and resource energy consumption requirements in the communication process, and ensures the performance of the neural network in the communication system.
[0005] In order to achieve the above object, the present invention adopts the following technical solutions:
[0006] A large neural network decomposition optimization method based on multi-band decomposition and waveform shaping, the method comprising:
[0007] Waveform shaping: use filters to smooth the edge of each sub-band signal, adjust the waveform characteristics of the sub-band signal, cut off the out-of-band radiation part of the sub-band signal, and reduce the out-of-band power leakage in the sub-band signal;
[0008] Decompose the signal in the frequency domain, divide the entire signal into multiple sub-bands, convert the time domain signal into the frequency domain using fast Fourier transform, and separate and process each sub-band signal;
[0009] Decomposition of the neural network: decompose the neural network into multiple sub-neural networks, each sub-band signal corresponds to an independent sub-neural network, and connect multiple sub-neural networks in parallel in a modular manner to form a block neural network for training and learning.
[0010] As a further preferred embodiment of the present invention, the following method is used to smooth the edge of each sub-band signal by using a filter:
[0011] Determine the multi-band signal to be waveform shaped, and use a raised cosine filter to adjust the waveform characteristics of the sub-band signal. The mathematical expression of the time domain signal of the raised cosine filter is:
[0012]
[0013] In the formula, T represents the symbol period, t is the time variable, β is the roll-off coefficient, and sinc() is the function form; in order to add the raised cosine filter to the communication system, filter the sub-band, and cut off the out-of-band radiation part, the convolution formula s(t) is used to perform convolution operation on each symbol:
[0014]
[0015] In the formula, a k represents the amplitude of the kth symbol, k is the symbol index, and h(t-kT) represents the impulse response of the raised cosine filter;
[0016] Taking into account the characteristics of the spectrum, the mathematical expression of the raised cosine filter in the frequency domain is:
[0017]
[0018] Where f is the frequency variable; the roll-off coefficient β determines the out-of-band roll-off speed and bandwidth of the raised cosine filter. When β = 0, the raised cosine filter becomes an ideal Nyquist filter with a frequency response of The inner part is a rectangle with no transition band. When β=1, the bandwidth is expanded to 1 / T, and there will be very wide out-of-band interference. Therefore, the total bandwidth B of the raised cosine filter is:
[0019]
[0020] The out-of-band bandwidth is corrected by adjusting the roll-off factor β to obtain different sub-band shapes.
[0021] As a further preferred embodiment of the present invention, when decomposing the neural network into a plurality of sub-neural networks, the sizes of the sub-neural networks are decomposed according to the sizes of the sub-band signals and the limitations of existing resources.
[0022] As a further preferred embodiment of the present invention, the input neurons and output neurons of the block neural network are independent of each other and are not connected.
[0023] As a further preferred embodiment of the present invention, each input neuron and output neuron of the block neural network is only connected to neurons on the left and right sides.
[0024] As a further preferred embodiment of the present invention, the block neural network adopts any one of a convolutional neural network, a deep neural network or a long short-term memory artificial neural network (LSTM network) for training and learning.
[0025] As a further preferred embodiment of the present invention, the sub-neural network comprises an input layer, a hidden layer and an output layer;
[0026] Input layer: receives sub-band signals and selects partial connection input or full connection input;
[0027] Hidden layer: Select different network architectures according to signal characteristics to process sub-band signals;
[0028] Output layer: outputs the processed sub-band signal.
[0029] As a further preferred embodiment of the present invention, the training and learning of the block neural network includes:
[0030] Normalize and extract features of multi-band signals to ensure the consistency of input signals in frequency and time domains, and construct annotated data sets of sub-band signals, including ideal waveforms and target spectra;
[0031] Each sub-neural network is trained one by one to learn and process the characteristics of the sub-band signals, and the block neural networks are jointly trained using the back-propagation algorithm.
[0032] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention can significantly reduce the complexity of neural network training in large-scale multi-carrier systems through independent training and parallel processing of sub-bands, while supporting the flexibility of hardware implementation. By dividing the signal into multiple independent sub-bands through filtering and shaping, a dedicated sub-neural network (subnet) can be designed for each sub-band. The independence between the sub-bands allows the training of each sub-neural network to be carried out separately, reducing the complexity of overall network training. The architectural design of the sub-neural network naturally supports parallel processing, can utilize hardware resources to achieve efficient neural network operation, reduce energy consumption, and significantly improve the processing efficiency of the neural network. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 A schematic diagram of an application of an artificial intelligence neural network in a communication system according to an embodiment of the present invention;
[0034] Figure 2 This is a schematic diagram of a large neural network decomposition optimization method based on multi-band decomposition and waveform shaping proposed by the present invention;
[0035] Figure 3 A schematic diagram of spectrum resource distribution corresponding to different signal designs in an embodiment of the present invention;
[0036] Figure 4 This is a neural network connection architecture diagram for the large neural network decomposition optimization based on multi-band decomposition and waveform shaping proposed in the present invention.
[0037] In the figure: 1. Bit data; 2. QAM modulator at the transmitting end; 3. Multi-carrier OFDM modulator; 4. RF unit at the transmitting end; 5. Transmitting antenna; 6. Receiver antenna; 7. RF unit at the receiving end; 8. OFDM demodulator at the receiving end; 9. QAM demodulator at the receiving end; 10. Bit data stream. DETAILED DESCRIPTION
[0038] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.
[0039] Figure 1 It is the application of artificial intelligence neural network in the communication system, which includes a transmitter and a receiver, as well as a specific neural network application scenario. Bit data 1 is the bit data stream of the transmitter, which is the lowest-level original data that needs to be transmitted. The QAM modulator 2 at the transmitter is responsible for converting the bit data stream into a symbol data stream. The multi-carrier OFDM modulator 3 is responsible for modulating the symbol data of a single carrier onto different subcarriers. The RF unit 4 at the transmitter is responsible for converting the digital signal into an analog signal that can be transmitted through the transmitting antenna 5. The RF unit 7 at the receiving end is responsible for converting the analog signal received by the receiver antenna 6 into a digital signal. The OFDM demodulator 8 at the receiving end is responsible for demodulating the data modulated on each subcarrier into the symbol data of a single carrier. The QAM demodulator 9 at the receiving end is responsible for decomposing the symbol data into a bit data stream 10.
[0040] The application of neural networks at the transmitter end can learn new modulation modes and new error correction code structures to replace the existing QAM modulation mode and error correction mechanism. It can also learn new multi-carrier modulation to replace the existing OFDM multi-carrier mode. It can also learn new digital precoding mechanisms to replace the existing minimum mean square error (MMSE) or zero forcing (ZF) digital precoding. The application of neural networks in detecting classified channels can learn signal classification identifiers to distinguish different received signals, monitor signals, and classify them. The application of neural networks at the receiver end can learn better signal detection and channel estimation methods, and can learn better modulation signal classification and decision methods.
[0041] In wireless communications, channels become highly dynamic due to multipath effects and mobility, and traditional channel estimation and equalization methods (such as minimum mean square error) may be inefficient in complex scenarios. Large neural networks are able to learn the statistical characteristics of the channel from the received signal and automatically adapt to nonlinear, non-Gaussian channel conditions. Under extreme conditions (such as fast channel fading or strong interference), large neural networks LNN can provide more accurate estimates, such as Figure 1 In the application of neural networks in signal detection and channel estimation, classic application cases include convolutional neural networks (CNN) for channel feature extraction and recurrent neural networks (RNN) for dynamic channel prediction. In massive MIMO and millimeter wave communications, beamforming and resource allocation are crucial to improving signal coverage and system throughput. LNN can learn the complex relationship between channel state information (CSI) and beamforming coefficients, and generate optimization solutions in real time, such as Figure 1 It learns new digital precoding mechanisms and supports dynamic resource allocation, taking into account changes in user locations, channel conditions, and service requirements, such as Figure 1 In the application of detection and classification channels, classic application cases include the use of deep reinforcement learning (DRL) combined with LNN for dynamic beam optimization and user scheduling. In efficient error correction codes (such as LDPC and Turbo codes), traditional algorithms have high computational complexity and are difficult to cope with dynamic scenarios. LNN can implement "neural learning error correction codes" (such as Neural Decoding) through neural network learning, and can achieve near-optimal performance under low signal-to-noise ratio and complex channel conditions. And it can be further optimized on the basis of traditional coding structures to improve coding gain, such as Figure 1 In the process of learning new modulation modes and new error correction codes, as well as classifying QAM signals and signal judgments, a typical case is a sequence decoder implemented with LSTM or Transformer network, which can quickly decode long codewords. In order to adapt to the randomly changing communication environment, LNN can design an end-to-end communication model through a large autoencoder to directly optimize the performance of the overall system (such as bit error rate, bandwidth utilization). It realizes multi-module joint optimization from the physical layer to the application layer (such as the entire process of signal sending, transmission, and reception). Figure 1It can learn new modulation modes, new error correction codes and new multi-carrier modes, effectively solving the suboptimal problem caused by the independent design of different modules. In cellular networks, satellite communications and the Internet of Things, resource management and network optimization are core challenges. LNN can predict changes in network traffic, optimize switching strategies, reduce energy consumption, improve network reliability, and support distributed network optimization to adapt to dynamic scenarios in heterogeneous networks. For example, generative adversarial networks (GANs) are used to simulate network traffic and deep learning is combined to achieve intelligent network management. In densely deployed networks (such as 5G small base stations or satellite cluster communications), interference management is the key to improving system performance. LNN can learn interference patterns and design collaborative communication strategies to maximize system capacity. In collaborative communications, large neural networks can efficiently coordinate information sharing and joint processing between nodes. Use graph neural networks (GNNs) to handle interference relationships between nodes.
[0042] Large neural networks such as convolutional neural networks (CNNs) or large prediction models (GPTs) usually contain millions or even billions of parameters. A large number of matrix operations are required during training and reasoning, which is extremely computationally intensive. The training of large neural networks requires high-performance hardware devices (such as GPUs and TPUs), which places high demands on cost and energy efficiency. When deployed in actual communication equipment (such as base stations or terminals), they may not be able to support complex large neural networks due to limited hardware resources (such as storage, processing power, and energy consumption). The training and reasoning of large neural networks often require a lot of energy consumption, which is particularly prominent in communication systems with limited energy (such as satellite communications and IoT devices). High energy consumption is contrary to the goals of energy conservation, emission reduction, and green communications pursued in today's communications field.
[0043] like Figure 2 In this embodiment, a large neural network decomposition optimization method based on multi-band decomposition and waveform shaping is proposed. The method mainly includes waveform shaping and neural network decomposition.
[0044] Waveform shaping uses filters to smooth the edge of each sub-band signal, adjust the waveform characteristics of the sub-band signal, remove the out-of-band radiation part of the sub-band signal, optimize the spectrum shape, and thus reduce interference with adjacent bands. This lays the foundation for independent processing of each signal band, making it more efficient to use neural networks to analyze and optimize signals; reduce out-of-band power leakage in sub-band signals to reduce mutual interference between sub-bands. The purpose is to reduce Figure 3 After pulse shaping, the interference between adjacent sub-bands is significantly weakened, making the training effect of the decomposed sub-neural network better. Figure 3In c, the mutual interference between different frequency bands is reduced, and the filtered signal is more "clean". Since the correlation between sub-bands is greatly reduced, the training complexity of the neural network is significantly reduced, which makes it easier for the neural network to train and analyze in each sub-band, thereby achieving Figure 4 The block neural network-A structure of II and the block neural network-B structure of III.
[0045] The filter can be of any structure. The most widely used raised cosine filter function is used in the present invention. Filter bandwidth: determined by system requirements, such as the need to meet the Nyquist criterion or communication standards. Raised cosine roll-off coefficient: by adjusting the size of the coefficient between 0 and 1, the filter shape is controlled, thereby controlling the out-of-band attenuation characteristics of the signal and reducing out-of-band power leakage.
[0046] Sub-band decomposition: Use a filter (such as the raised cosine filter used in the present invention) to perform signal edge smoothing on the input signal sub-band, so that the signal can be decomposed in the frequency domain and the entire signal can be divided into multiple sub-bands. Frequency domain transformation: Use FFT (Fast Fourier Transform) to convert the time domain signal into the frequency domain, and separate the signal of each frequency band. Separation accuracy control: The separation of sub-bands must ensure the orthogonality between adjacent frequency bands, which is achieved by selecting a reasonable raised cosine transition bandwidth and raised cosine roll-off coefficient.
[0047] Decomposition of the neural network: Decompose the size of the sub-band signal according to the size of the sub-neural network and the limitations of existing resources, and decompose the neural network into multiple sub-neural networks. Each sub-band signal corresponds to an independent sub-neural network. Each sub-neural network contains an input layer, a hidden layer, and an output layer. Input layer: Receives specific sub-band signals, and can choose partial connection input (extract local features) or full connection input (extract global features). Hidden layer: Select different network architectures according to signal characteristics: Convolutional layer (CNN) is suitable for frequency domain feature extraction. Recurrent layer (RNN / LSTM) is suitable for signals with strong temporal correlation. Output layer: Outputs processed sub-band signals to support tasks such as fitting and classification. Independent optimization of sub-networks: Each sub-neural network is trained independently to optimize its ability to process specific frequency band signals. Each sub-neural network only processes sub-bands, and multiple sub-neural networks are connected in parallel in a modular way to form a block neural network, and then trained and learned. All sub-neural networks process input signals independently, and finally integrate the results through the aggregation layer. Shared parameters: Share certain parameters (such as convolution kernels or biases) between adjacent sub-networks to reduce model complexity. The input neurons and output neurons of the block neural network are independent and unconnected. Each input neuron and output neuron of the block neural network is only connected to the neurons on the left and right sides. The block neural network uses any of the convolutional neural networks, deep neural networks or long short-term memory artificial neural networks (LSTM networks) for training and learning.
[0048] Training of blocky neural networks:
[0049] (1) Data preprocessing:
[0050] Normalize and extract features of multi-band signals to ensure that the input signals are consistent in frequency and time domains. Build an annotated dataset of sub-band signals, including ideal waveforms and target spectra.
[0051] (2) Training process:
[0052] Single block training: Train each sub-neural network one by one to enable it to effectively learn and process the characteristics of the sub-band signal. Global optimization: Use the back propagation algorithm to jointly train the entire block neural network. Generally, the initial learning rate is 0.01 and gradually decays. The reverse optimization uses the Adam optimizer or the SGD (stochastic gradient descent) algorithm. Verification indicators: Evaluate out-of-band radiation suppression, signal reconstruction accuracy and operation efficiency.
[0053] Specifically, the filter is used to smooth the edge of each sub-band signal in the following manner:
[0054] Determine the multi-band signal to be waveform shaped, and use a raised cosine filter to adjust the waveform characteristics of the sub-band signal. The mathematical expression of the time domain signal of the raised cosine filter is:
[0055]
[0056] Where T represents the symbol period, t is the time variable, β is the roll-off coefficient, which is between 0 and 1, and sinc() is a function expressed as sin(x) / x. In order to add the raised cosine filter to the communication system, filter the sub-band and cut off the out-of-band radiation, the convolution formula s(t) is used to perform convolution operation on each symbol:
[0057]
[0058] In the formula, a k represents the magnitude of the kth symbol, k is the symbol index, and h(t-kT) represents the impulse response of the raised cosine filter;
[0059] Taking into account the characteristics of the spectrum, the mathematical expression of the raised cosine filter in the frequency domain is:
[0060]
[0061] Where f is the frequency variable; the roll-off coefficient β determines the out-of-band roll-off speed and bandwidth of the raised cosine filter. When β = 0, the raised cosine filter becomes an ideal Nyquist filter with a frequency response of The inner part is rectangular, without transition band, which is a perfect rectangular filter. Figure 3 For the rectangular sub-band c, when β = 1, the bandwidth is expanded to 1 / T, and there will be very wide out-of-band interference. Therefore, the total bandwidth B of the raised cosine filter is:
[0062]
[0063] The out-of-band bandwidth is corrected by adjusting the roll-off factor β to obtain different sub-band shapes.
[0064] The bandwidth of the frequency band signal in the frequency domain is closely related to β. The smaller the β value, the better the frequency domain characteristics. The larger the β value, the more the frequency band will be expanded, but the time domain will be smoother. Therefore, the bandwidth utilization and anti-interference performance can be compromised by adjusting the roll-off factor β. In actual systems, the raised cosine filter is easier to implement in communication systems, but for the sake of algorithm and hardware simplification. In actual systems, there will still be interference between each sub-band, but it will be greatly reduced. Figure 3 The interference between sub-bands only comes from the left and right sub-bands, so a new type of block neural network-C, such as Figure 4The block neural network-C structure in IV can more effectively reduce the training complexity of large neural networks.
[0065] In practical applications, Figure 3 Design corresponding spectrum resource distribution for different signals. a is the existing single-band OFDM signal spectrum. This signal occupies all spectrum resources and requires a large neural network to perform corresponding analysis and processing. Therefore, in order to effectively train data, the neural network needs a large number of neurons to assist, and each neuron must have a link to communicate with each other to build a fully connected neural network. Figure 4 Middle I is the currently commonly used fully connected neural network, in which each neuron is connected to each other. Due to the large number of neurons and connections between neurons, this neural network architecture has a long training time and high energy consumption.
[0066] Figure 3 In b, it is the spectrum of multi-band OFDM signal. The signal is divided into different small signals. Correspondingly, a small neural network can be used to perform corresponding analysis and processing on each sub-band. In theory, it can reduce the complexity of the neural network and speed up the training. However, there will be mutual interference between each frequency band, which will seriously affect the performance of neural network training and learning. c is the filtering block multi-band OFDM signal structure proposed by the invention. It can be seen that each sub-band is independent of each other and there is no overlapping and interfering area. In this way, a completely independent small neural network can be used to perform learning and training on each sub-band, such as Figure 3 In the figure d, the filtering block multi-band OFDM signal structure is proposed in the context that out-of-band interference cannot be completely eliminated in the actual system. It can be seen that there is still some overlapping interference between each sub-band, but it is not very serious. In this way, the sub-neural network can still be used to learn and train each sub-band, but some fine-tuning of the neural network is required, such as Figure 4 The blocky neural network-C structure of Figure Ⅳ.
[0067] Figure 4 In the figure, Ⅰ is a fully connected neural network commonly used at present, and each neuron is connected to each other. Ⅱ-Ⅳ are block neural network structures proposed by the present invention. Each input and output neuron of block neural network-A in Ⅱ is independent and not connected to each other, which greatly reduces the complexity of calculation. Block neural network-B in Ⅲ can flexibly modify the scale of input and output neurons. Block neural network-C in Ⅳ is a structure that takes actual signals into consideration. Each input and output neuron is only connected to the neurons on the left and right sides, which reduces the complexity of training while ensuring the performance of training.
[0068] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.
Claims
1. A large neural network decomposition optimization method based on multi-band decomposition and waveform shaping, characterized in that: include: Waveform shaping: use filters to smooth the edge of each sub-band signal, adjust the waveform characteristics of the sub-band signal, cut off the out-of-band radiation part of the sub-band signal, and reduce the out-of-band power leakage in the sub-band signal; Decompose the signal in the frequency domain, divide the entire signal into multiple sub-bands, convert the time domain signal into the frequency domain using fast Fourier transform, and separate and process each sub-band signal; Decomposition of the neural network: decompose the neural network into multiple sub-neural networks, each sub-band signal corresponds to an independent sub-neural network, and connect multiple sub-neural networks in parallel in a modular manner to form a block neural network for training and learning.
2. The large-scale neural network decomposition and optimization method based on multi-band decomposition and waveform shaping according to claim 1 is characterized in that: The filter is used to smooth the edge of each sub-band signal in the following way: Determine the multi-band signal to be waveform shaped, and use a raised cosine filter to adjust the waveform characteristics of the sub-band signal. The mathematical expression of the time domain signal of the raised cosine filter is: In the formula, T represents the symbol period, t is the time variable, β is the roll-off coefficient, and sinc() is the function form; in order to add the raised cosine filter to the communication system, filter the sub-band, and cut off the out-of-band radiation part, the convolution formula s(t) is used to perform convolution operation on each symbol: In the formula, a k represents the amplitude of the kth symbol, k is the symbol index, and h(t-kT) represents the impulse response of the raised cosine filter; Taking into account the characteristics of the spectrum, the mathematical expression of the raised cosine filter in the frequency domain is: Where f is the frequency variable; the roll-off coefficient β determines the out-of-band roll-off speed and bandwidth of the raised cosine filter. When β = 0, the frequency response is The inner part is a rectangle with no transition band. When β=1, the bandwidth is expanded to 1 / T. The total bandwidth B of the raised cosine filter is: The out-of-band bandwidth is corrected by adjusting the roll-off factor β to obtain different sub-band shapes.
3. The large-scale neural network decomposition and optimization method based on multi-band decomposition and waveform shaping according to claim 1 is characterized in that: When a neural network is decomposed into multiple sub-neural networks, the sizes of the sub-neural networks are decomposed according to the size of the sub-band signals and the limitations of existing resources.
4. The large-scale neural network decomposition and optimization method based on multi-band decomposition and waveform shaping according to claim 1 is characterized in that: The input neurons and output neurons of the block neural network are independent and not connected to each other.
5. The large-scale neural network decomposition and optimization method based on multi-band decomposition and waveform shaping according to claim 1 is characterized in that: Each input neuron and output neuron of the block neural network is only connected to neurons on the left and right sides.
6. The large-scale neural network decomposition and optimization method based on multi-band decomposition and waveform shaping according to claim 1 is characterized in that: The block neural network is trained and learned using any one of a convolutional neural network, a deep neural network or a long short-term memory artificial neural network.
7. The large-scale neural network decomposition and optimization method based on multi-band decomposition and waveform shaping according to claim 1 is characterized in that: The sub-neural network contains input layer, hidden layer and output layer; Input layer: receives sub-band signals and selects partial connection input or full connection input; Hidden layer: Select different network architectures according to signal characteristics to process sub-band signals; Output layer: outputs the processed sub-band signal.
8. The large-scale neural network decomposition and optimization method based on multi-band decomposition and waveform shaping according to claim 1 is characterized in that: Training and learning of block neural networks include: Normalize and extract features of multi-band signals to ensure the consistency of input signals in frequency and time domains, and construct annotated data sets of sub-band signals, including ideal waveforms and target spectra; Each sub-neural network is trained one by one to learn and process the characteristics of the sub-band signals, and the block neural networks are jointly trained using the back-propagation algorithm.