Space optical communication decoding method based on binary search pruning
By using a binary search pruning method in the spatial optical communication system, the weight matrix of the bidirectional recurrent neural network is pruned, which solves the problems of low system performance, high computational complexity and redundancy, and achieves more efficient decoding and lower computational complexity.
Patent Information
- Application Number
- CN202510133007.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-06
- Publication Date
- 2025-05-30
AI Technical Summary
The existing spatial optical communication methods have problems such as low system performance, high computational complexity, low accuracy and system redundancy.
The spatial optical communication decoding method based on binary search and pruning is adopted. By extracting the weight matrix of the bidirectional recurrent neural network as the initial weight matrix, the mask matrix and pruning weight matrix are constructed, and the sparse binary search algorithm is used for iterative training until the end of the loop is to replace the weight matrix of the bidirectional recurrent neural network in the decoder and reduce the calculation complexity of the decoder.
On the premise of ensuring system accuracy, the calculation complexity of the bidirectional recurrent neural network decoder is reduced, the system redundancy is reduced, and the performance of the spatial optical communication system is improved.
Smart Images
Figure CN120074667A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of free-space optical communication, and particularly to a free-space optical communication decoding method based on binary search pruning. Background Art
[0002] Free-space optical communication, as an important part of the space-air-ground integrated strategy, has received extensive attention. Free-space optical communication is a wireless access technology using laser as the carrier, loading digital signals onto laser beams, transmitting signals through the atmospheric channel, and receiving optical signals using photodetectors at the receiving end. Compared with traditional radio frequency communication technologies, it has the advantages of large communication capacity, no need for spectrum authorization, high transmission rate, strong anti-interference ability, high confidentiality, etc.
[0003] At present, neural networks based on deep learning, relying on their feature extraction capabilities, have gradually been adopted in the field of free-space optical communication, especially in long-distance and heavy turbulence scenarios to suppress the degradation of system performance caused by atmospheric turbulence effects. However, since deep learning requires the ability to learn complex non-linear models when implementing prediction or classification problems, as the dimension of the neural network model increases, the training cost gradually rises, and the computational complexity increases significantly. For free-space optical communication systems, the deployment of neural networks requires hardware support, and high complexity will affect the operation speed, storage, and energy consumption of the receiver. Summary of the Invention
[0004] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is the problems of low system performance, large computational complexity, low accuracy, and system redundancy in the existing free-space optical communication methods. The present invention provides a free-space optical communication decoding method based on binary search pruning, which reduces the computational complexity of the bidirectional recurrent neural network decoder and reduces the system redundancy on the premise of ensuring system accuracy.
[0005] To achieve the above object, the present invention provides a free-space optical communication decoding method based on binary search pruning, including the following steps:
[0006] Obtain the bidirectional recurrent neural network of the decoder using a receiver, and extract the weight matrices of each layer of the bidirectional recurrent neural network as the initial weight matrices from the bidirectional recurrent neural network of the decoder;
[0007] Construct a mask matrix and a pruned weight matrix using the initial weight matrices;
[0008] Construct a weight permutation matrix using the initial weight matrices, and set an upper bound and a lower bound for the sparsity;
[0009] Use the sparsity binary search algorithm to set the pruning sparsity to perform iterative training, and perform binary search pruning until the loop ends;
[0010] After the loop ends, the pruned weight matrix is extracted to replace the weight matrix in the bidirectional recurrent neural network in the decoder, obtaining a pruned free-space optical communication decoder to reduce the computational complexity of the decoder.
[0011] Further, the receiver includes an optical filter, a photodiode, an analog-to-digital converter, and a receiver decoder. After the optical signal received by the receiver passes through the optical filter and the photodiode, it is subjected to analog-to-digital conversion by the analog-to-digital converter, and the signal calculates the bit error rate after passing through the decoder bidirectional recurrent neural network.
[0012] Further, the decoder bidirectional recurrent neural network is obtained by using the receiver, and the weight matrices of each layer of the bidirectional recurrent neural network are extracted as the initial weight matrix, specifically including the following steps:
[0013] Step 1: Build a free-space optical communication system based on a decoder bidirectional recurrent neural network. Among them, the receiver decoder uses a bidirectional recurrent neural network. The input dimension n of the bidirectional recurrent neural network is 12, the output dimension m is 16, the dimension of the hidden layer bidirectional recurrent neural network is 16, and the number of layers is 1. Among them, for the hidden layer bidirectional recurrent neural network, the states at the forward time t1, the backward output time t2, and the overall output time t are respectively expressed as:
[0014]
[0015] where, h t is the state at the current time, and h t-1 is the state at the previous time. α t is the activation function, W t and W h are the weight matrices at different times and different dimensions respectively, and b t is the bias vector.
[0016] Step 2: Perform deep learning training on the free-space optical communication system based on the decoder bidirectional recurrent neural network. Based on the ADAM gradient descent optimizer, after 5000 iterations, ensure the transmission reliability performance of the decoder, and extract the weight matrices of different dimensions in the decoder bidirectional recurrent neural network as the initial weight matrix for pruning.
[0017] Further, the photoelectric conversion sensitivity of the photodiode is 1 A / W.
[0018] Further, a mask matrix and a pruned weight matrix are constructed using the initial weight matrix. Specifically, a mask matrix and a pruned weight matrix are constructed using the initial weight matrix for binary search pruning to determine its dimension, initial value, and corresponding relationship. Among them, the dimensions of the mask matrix and the pruned weight matrix are the same as those of the initial weight matrix.
[0019] Further, determine the initial values of the mask matrix and the pruning weight matrix. Specifically, the initial value of each element of the mask matrix is 1, and the value of each element of the pruning weight matrix is the same as the value of the corresponding element of the initial weight matrix. The initial value of the corresponding element of the initial weight matrix is extracted from the decoder bidirectional recurrent neural network after deep learning training.
[0020] Further, determine the correspondence between the elements of the mask matrix and the pruning weight matrix. Specifically, each element of the mask matrix is 0 or 1. If the corresponding element of the weight matrix is 0, then the element of the mask matrix is 0; if the corresponding element of the weight matrix is 1, then the element of the mask matrix is 1. The construction method of the pruning weight matrix is
[0021] Further, construct a weight permutation matrix using the initial weight matrix, and set the upper bound and lower bound of the sparsity. Specifically, use the initial weight matrix w to construct a weight permutation matrix v by arranging the second-order norm values of the elements of w from small to large; then determine the correspondence between each row of each column of the weight permutation matrix; finally, to ensure that the best performance of the bidirectional recurrent neural network is within the range of the upper and lower bounds of the sparsity, and to speed up the search speed of the binary search pruning algorithm, determine the initial lower bound of the sparsity to be 0 and the initial upper bound to be 0.64.
[0022] Further, the dimension size of the weight permutation matrix has 3 rows and the number of columns is the number of elements of the weight matrix.
[0023] Technical effects
[0024] A space optical communication decoding method based on binary search pruning provided by the present invention uses a binary search pruning scheme to compress the model of the decoder based on the recurrent neural network, and reduces the computational complexity of the decoder on the premise of ensuring that the sensitivity performance does not decrease.
[0025] The following will further illustrate the concept, specific structure and technical effects generated by the present invention with reference to the accompanying drawings, so as to fully understand the purpose, features and effects of the present invention. Brief description of the drawings
[0026] Figure 1 is a flowchart of a space optical communication decoding method based on binary search pruning according to a preferred embodiment of the present invention;
[0027] Figure 2 is a flowchart of a progressive pruning method based on sparsity binary search of a space optical communication decoding method according to a preferred embodiment of the present invention;
[0028] Figure 3Bit error rate, sparsity, and progressivity analysis diagrams for each pruning round of the binary search pruning method of a space optical communication decoding method based on binary search pruning in a preferred embodiment of the present invention;
[0029] Figure 4 Computational complexity analysis diagram for each pruning round of the binary search pruning method of a space optical communication decoding method based on binary search pruning in a preferred embodiment of the present invention. Detailed implementation manners
[0030] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.
[0031] In the following description, specific details such as specific internal programs and technologies are proposed for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present invention. However, those skilled in the art should clearly understand that the present invention can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits and methods are omitted to avoid unnecessary details from interfering with the description of the present invention.
[0032] In the model compression based on the neural network method, parameter pruning is mainly divided into two parts: unstructured pruning and structured pruning. Among them, unstructured pruning is a connection-level fine-grained pruning, which requires specific algorithms or hardware platforms to support due to its high precision, including weight pruning and neuron pruning. Structured pruning has a coarser granularity and lower precision, and can be directly run on a deep learning framework. Block-level pruning can reconstruct the model, has a regular structure and does not require hardware support such as a sparse convolution library. It mainly includes vector-level pruning, kernel-level pruning, group-level pruning and filter pruning. Since in this embodiment, the bidirectional recurrent neural network has a fixed number of network layers and hidden layer dimensions, coarser-grained pruning will cause the system to be underfitted and affect the sensitivity performance of the decoder. Therefore, the pruning scheme in this embodiment is based on weight pruning in fine-grained pruning.
[0033] This embodiment provides a space optical communication decoding method based on binary search pruning, in which the decoder uses a bidirectional recurrent neural network, and the binary search pruning method is implemented on the weight matrix of the bidirectional recurrent neural network by using the sparsity binary search algorithm, so as to reduce the computational complexity of the decoder without losing the sensitivity performance of the decoder.
[0034] As Figure 1 shown, the embodiment of the present invention provides a space optical communication decoding method based on binary search pruning, including the following steps:
[0035] Step 1: Use the receiver to obtain the decoder bidirectional recurrent neural network, and extract the weight matrices of each layer of the bidirectional recurrent neural network from the decoder bidirectional recurrent neural network as the initial weight matrices; specifically including:
[0036] (1). Build a free-space optical communication system based on the decoder of the bidirectional recurrent neural network. Among them, the receiver decoder uses a bidirectional recurrent neural network. The input dimension n of the bidirectional recurrent neural network is 12, the output dimension m is 16, the dimension of the hidden layer bidirectional recurrent neural network is 16, and the number of layers is 1. Among them, for the hidden layer bidirectional recurrent neural network, the states at the forward time t1, the backward output time t2, and the overall output time t are respectively expressed as:
[0037]
[0038] Among them, h t is the state at the current time, and h t-1 is the state at the previous time. α t is the activation function, W t and W h are the weight matrices at different times and different dimensions respectively, and b t is the bias vector.
[0039] (2). For the free-space optical communication system based on the decoder of the bidirectional recurrent neural network, implement deep learning training. Based on the ADAM gradient descent optimizer, after 5000 iterations, ensure the transmission reliability performance of the decoder, and extract the weight matrices of different dimensions in the decoder bidirectional recurrent neural network as the initial weight matrices for pruning.
[0040] In this embodiment, the receiver includes an optical filter, a photodiode, an analog-to-digital converter, and a receiver decoder. After the optical signal received by the receiver passes through the optical filter and the photodiode, it is subjected to analog-to-digital conversion by the analog-to-digital converter, and the signal calculates the bit error rate after passing through the decoder bidirectional recurrent neural network. Among them, the photoelectric conversion sensitivity of the photodiode is 1 A / W (ampere / watt). The decoder uses a bidirectional recurrent neural network structure, the neural network input dimension is 12, the output dimension is 16, the number of hidden layers is 2 layers, and the hidden layer dimension is 16.
[0041] Step 2: Use the initial weight matrix to construct a mask matrix and a pruned weight matrix;
[0042] In this embodiment, use the initial weight matrix to construct a mask matrix and a pruned weight matrix for binary search pruning, and determine its dimension, initial value, and corresponding relationship.
[0043] Determine the dimensions of the mask matrix u and the pruned weight matrix , specifically, the dimensions of the mask matrix and the pruned weight matrix are the same as those of the initial weight matrix w.
[0044] Determine the initial values of the mask matrix and the pruning weight matrix. Among them, the initial value of each element of the mask matrix is 1, and the value of each element of the pruning weight matrix is the same as the value of the corresponding element of the initial weight matrix.
[0045] Determine the corresponding relationship between the elements of the mask matrix and the pruning weight matrix. Specifically, each element of the mask matrix is 0 or 1. If the corresponding element of the weight matrix is 0, then the element of the mask matrix is 0; if the corresponding element of the weight matrix is 1, then the element of the mask matrix is 1. The construction method of the pruning weight matrix is
[0046] Step 3: Use the initial weight matrix to construct a weight permutation matrix, and set the upper bound and lower bound of the sparsity. Specifically, use the initial weight matrix w to construct a weight permutation matrix v according to the ascending order of the second-order norm values of the elements of w.
[0047] Among them, the dimension size of the weight permutation matrix has 3 rows and the number of columns is the number of elements of the weight matrix.
[0048] Determine the corresponding relationship between each row of each column of the weight permutation matrix. Among them, the first row is the second-order norm of the element, the second row is the number of the weight matrix where the element is located, and the third row is the position of the element in the matrix. The second row and the third row together constitute the index of the element.
[0049] In addition, considering the sensitivity performance of the bidirectional recurrent neural network and the binary search pruning method comprehensively, the initial lower bound S l of the sparsity in this application is set to 0, and the initial upper bound S r is set to 0.64.
[0050] Step 4: Use the sparsity binary search algorithm to set the pruning sparsity to perform iterative training, and perform binary search pruning until the loop ends. After the loop ends, extract the pruning weight matrix to replace the weight matrix in the bidirectional recurrent neural network in the decoder, and obtain the pruned space optical communication decoder to reduce the computational complexity of the decoder.
[0051] Next, the progressive pruning method based on sparsity binary search in step 4 of the embodiment of this application will be introduced, as Figure 2 shown.
[0052] In the embodiment of this application, S i is used to represent the sparsity, S i retains two decimal places, and the minimum progressive degree is set to 0.01. Set the initial lower bound S l to 0, and the initial upper bound S ris 0.64. In this application, the sparsity is set as the pruning measurement criterion for the bidirectional recurrent neural network. The pruning sparsity is the ratio of the number of elements with a second-order norm of 0 in the first row of v to the total number of elements. Set the intermediate pruning sparsity for binary search as S m , and let S m =(S l +S r ) / 2.
[0053] Step 4.1: Update the mask matrix using the weight matrix, and prune the values of the elements in the weight matrix and the weight permutation matrix.
[0054] Among them, the calculation relationship between the weight matrix and the pruned weight matrix is If the element in the weight matrix is not 0, the value of the corresponding element in the mask matrix is 1, and the value of the corresponding element in the pruned weight matrix is equal to the value of the corresponding element in the weight matrix; if the value of the element in the weight matrix is 0, the value of the corresponding element in the mask matrix is 0, and the value of the corresponding element in the pruned weight matrix is equal to 0. In addition, if the corresponding element in the weight matrix is 0, set the second-order norm value of the corresponding element in the weight permutation matrix to 0 according to the index, and rearrange the weight permutation matrix from largest to smallest according to the second-order norm.
[0055] Step 4.2: According to the pruning sparsity S i for each round, set the elements at the corresponding positions in the mask matrix of the first S i % of the elements with the smallest second-order norm in the weight permutation matrix from 1 to 0. Among them, set the intermediate pruning sparsity for binary search as S m , and let S m =(S l +S r ) / 2.
[0056] Step 4.3: For the weight permutation matrix under a certain set sparsity S i , calculate the corresponding pruned weight matrix. The calculation formula is Construct a pruned neural network decoder by replacing the weight matrix with the pruned weight matrix, and perform 500 iterations of deep learning training to fine-tune the active weight elements that have not been pruned.
[0057] Step 4.4: Extract the pruned neural network after iterative training, and compare the bit error rates of the output signals before and after pruning. Use the 7% HD-FEC decision threshold (3.8×10 -3 ) as the standard basis for system reliability. By comparing the bit error rates before and after pruning, the difference in sensitivity of the decoder to reach the HD-FEC decision threshold before and after pruning can be calculated. If the sensitivity loss after pruning is less than 1 dB, then execute S l =S m+0.01, and replace the weight matrix with the current pruned weight; if the sensitivity loss after pruning exceeds 1 dB, execute S r = S m -0.01, and the weight matrix is not updated at this time.
[0058] The embodiment of the present application cycles through steps one to four based on the progressive pruning scheme of sparsity binary search, respectively implementing the four steps of matrix generation, weight pruning, iterative training, and reliability judgment until S l > S r , and the pruning based on the binary search method ends.
[0059] Next, the feasibility of a space optical communication decoding method based on binary search pruning provided by the present application is analyzed through specific implementation cases.
[0060] Extract the decoder bidirectional recurrent neural network from the receiver using a deep learning-based space optical communication system, and extract the weight matrices of each layer of the bidirectional recurrent neural network as the initial weight matrix to build a space optical communication system based on the bidirectional recurrent neural network decoder. Among them, the receiver decoder uses a bidirectional recurrent neural network. The input dimension n of the bidirectional recurrent neural network is 12, the output dimension m is 16, the dimension of the hidden layer bidirectional recurrent neural network is 16, and the number of layers is 1. For the space optical communication system based on the bidirectional recurrent neural network decoder, deep learning training is implemented. Based on the ADAM gradient descent optimizer, after 5000 iterations, the transmission reliability performance of the decoder is ensured, and the weight matrices of different dimensions in the decoder bidirectional recurrent neural network are extracted as the initial weight matrix for pruning.
[0061] Construct the mask matrix and the elements of the pruned weight matrix. The construction method of the pruned weight matrix is When initializing, all elements of the mask matrix are 1, and the pruned weight matrix is the same as the initial weight matrix.
[0062] Use the initial weight matrix to construct a weight permutation matrix, and set the upper bound and lower bound of sparsity; use the initial weight matrix w to construct a weight permutation matrix v according to the ascending order of the second-order norm values of the w elements.
[0063] In the pruning scheme based on sparsity binary search in the embodiment of the present application, the weight threshold is calculated as the pruning criterion through sparsity, the sparsity of each round of pruning is calculated based on the binary search method, and the weight parameters are gradually pruned and updated during each round of pruning through the bit error rate and sensitivity. Set the initial lower bound S l to 0, and the initial upper bound S r to 0.64. According to the binary search method and the upper and lower bounds of sparsity formulated in the embodiment of the present application, the number of pruning times in the embodiment of the present application is 6 times.
[0064] Since the decoder can reach the 7% hard decision forward error correction (HD-FEC) decision threshold at a signal-to-noise ratio (SNR) of 22 dB before pruning, in order to ensure that the sensitivity performance of the decoder does not degrade before and after pruning, the decoder after pruning needs to meet the HD-FEC threshold at an SNR of 23 dB.
[0065] Figure 3 Shows the bit error rate (BER), sparsity, and progress analysis for each pruning round of the binary search pruning method provided by the embodiments of the present application. In the first and second training rounds, the pruning sparsities are 0.32 and 0.16 respectively, and the BER is greater than 10 at 23 dB -2 , so according to the binary search method, the pruning sparsity needs to be further reduced. The pruning sparsity in the third round is 0.08. Under this condition, reliable signal transmission can be achieved. Therefore, the pruning sparsity in the fourth round can be increased to 0.10. Continuing to implement the pruning method according to this standard, the pruning sparsity in the fifth round is 0.12, and the pruning sparsity in the sixth round is 0.11. After the pruning iterative training is completed, it can be analyzed that the autoencoder meets the HD-FEC threshold at 23 dB in the sixth round of pruning, meeting the sensitivity performance evaluation criteria of the embodiments of the present application, and the final pruning sparsity is 0.11.
[0066] In each pruning round of the embodiments of the present application, based on a specific pruning sparsity, the second-order norm threshold can be calculated through v, and the masks with weights less than this threshold are set to 0. In the first two rounds of pruning, since the system bit error rate exceeds the HD-FEC threshold and is not reliable, the calculated second-order norm threshold is discarded and cleared. The system in the third round of pruning meets the HD-FEC threshold. Therefore, the mask matrix corresponding to the elements in v less than 0.11 is set to 0, and the weight matrix is updated through iterative training. The mask matrix constructed in the fourth round of pruning is not updated because the decoder does not meet the signal transmission reliability. In this round, until the sixth round of pruning, under the condition of a second-order norm threshold of 0.15, the updated pruning weight matrix is used to construct a bidirectional recurrent neural network. From the second-order norm threshold of the weights, it can be seen that the binary search scheme of the embodiments of the present application has progressiveness in weight pruning. In the sixth round of pruning, for the third, fifth, and sixth times, the weights with the smallest second-order norm are gradually pruned until the maximum pruning sparsity is reached.
[0067] Floating Point Operations Per Second (FLOPs) can be used to measure the computational volume and computational time complexity of forward propagation, and serve as a measure of the speed of a neural network model. Floating-point operations are real number operations, and each multiplication and addition in a neural network is counted. To measure the impact of the binary search pruning method on the computational complexity of the decoder, the embodiments of this application provide an analysis diagram of the computational complexity at each pruning round of the binary search pruning method, as shown in Figure 4 shown. In the forward propagation of the weight matrix, without considering the influence of the activation function, a large number of multiplication operations and addition operations will be generated when multiplying the matrices to find the outer product. In the pruning algorithm, some weight elements are not activated, reducing the number of multiplication and addition operations in the matrix transformation. As shown in Figure 4 shown, after 6 rounds of pruning are implemented and the sparsity is 0.11, the FLOPs computational volume is reduced by approximately 11.31% before and after the binary search pruning.
[0068] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field of this application based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should be within the protection scope determined by the claims.
Claims
1. A spatial optical communication decoding method based on binary search pruning, characterized in that: The following steps are involved: Obtaining a decoder bidirectional recurrent neural network using a receiver, and extracting a weight matrix of each layer of the bidirectional recurrent neural network from the decoder bidirectional recurrent neural network as an initial weight matrix; Use the initial weight matrix to construct the mask matrix and pruning weight matrix; Use the initial weight matrix to construct a weight permutation matrix and set the upper and lower bounds of sparsity; Using the sparsity binary search algorithm, set the pruning sparsity to perform iterative training, and perform binary search pruning until the cycle ends; After the cycle is completed, the pruned weight matrix is extracted to replace the weight matrix in the bidirectional recurrent neural network in the decoder to obtain a pruned spatial optical communication decoder to reduce the computational complexity of the decoder.
2. A spatial optical communication decoding method based on binary search pruning as claimed in claim 1, characterized in that: The receiver includes an optical filter, a photodiode, an analog-to-digital converter and a receiver decoder. The optical signal received by the receiver passes through the optical filter and the photodiode, and then is converted into a digital signal by the analog-to-digital converter. The signal passes through the bidirectional recurrent neural network of the decoder to calculate the bit error rate.
3. A spatial optical communication decoding method based on binary search pruning as claimed in claim 1, characterized in that: The decoder bidirectional recurrent neural network is obtained by using a receiver, and a weight matrix of each layer of the bidirectional recurrent neural network is extracted from the decoder bidirectional recurrent neural network as an initial weight matrix, specifically comprising the following steps: Step 1: Build a spatial optical communication system based on a bidirectional recurrent neural network decoder; Step 2: Implement deep learning training for the spatial optical communication system based on the bidirectional recurrent neural network decoder. Based on the ADAM gradient descent optimizer, after 5000 iterations, ensure the transmission reliability performance of the decoder, and extract the weight matrices of different dimensions in the bidirectional recurrent neural network of the decoder as the initial weight matrix for pruning.
4. A spatial optical communication decoding method based on binary search pruning as claimed in claim 2, characterized in that: The photoelectric conversion sensitivity of the photodiode is 1A / W.
5. The spatial optical communication decoding method based on binary search pruning according to claim 1, characterized in that: The initial weight matrix is used to construct a mask matrix and a pruning weight matrix. Specifically, the initial weight matrix is used to construct a mask matrix and a pruning weight matrix for binary search pruning, and their dimensions, initial values and corresponding relationships are determined, wherein the dimensions of the mask matrix and the pruning weight matrix are the same as those of the initial weight matrix.
6. A spatial optical communication decoding method based on binary search pruning as claimed in claim 5, characterized in that: Determine the initial values of the mask matrix and the pruning weight matrix, specifically, the initial value of the elements of the mask matrix is 1, and the initial value of the elements of the pruning weight matrix is the same as the value of the corresponding element of the initial weight matrix, wherein the initial values of the elements corresponding to the initial weight matrix are extracted from the decoder bidirectional recurrent neural network after deep learning training.
7. A spatial optical communication decoding method based on binary search pruning as claimed in claim 5, characterized in that: Determine the correspondence between the mask matrix and the pruning weight matrix elements. Specifically, the mask matrix elements are 0 or 1. If the corresponding element of the weight matrix is 0, the mask matrix element is 0; if the corresponding element of the weight matrix is 1, the mask matrix element is 1. The construction method of the pruning weight matrix is 8. The spatial optical communication decoding method based on binary search pruning according to claim 1, characterized in that: The initial weight matrix is used to construct a weight permutation matrix, and the upper and lower bounds of sparsity are set. Specifically, the initial weight matrix w is used to construct a weight permutation matrix v according to the second-order norm values of the elements of w. Then, the corresponding relationship between each row of each column of the weight permutation matrix is determined. Finally, in order to ensure that the optimal performance of the bidirectional recurrent neural network is within the range of the upper and lower bounds of the sparsity, and at the same time to speed up the search speed of the binary search pruning algorithm, the initial lower bound of the sparsity is determined to be 0 and the initial upper bound is 0.
64.
9. A spatial optical communication decoding method based on binary search pruning as claimed in claim 8, characterized in that: The dimension size of the weight arrangement matrix is 3 rows and the number of columns is the number of weight matrix elements.