A method, system and device for generating a large-capacity mutually orthogonal complementary sequence set

Abstract

The present invention discloses a method, system, and device for generating a large-capacity mutually orthogonal complementary sequence set, belonging to the field of data processing. The method comprises: using a long short-term memory (LSTM) network for training to generate an initial sequence set; reshaping the initial sequence into a symmetric weight matrix through a Hopfield neural network and performing asynchronous state updates; grouping and cutting the updated sequence set to generate a pending sequence set; and evaluating the pending sequence set based on the MC indicator to screen and obtain a final mutually orthogonal complementary sequence set. By leveraging the LSTM's ability to model long-term sequence dependencies and combining iterative optimization of orthogonal complementary constraints with the Hopfield neural network, the present invention breaks through the number theory limitations of traditional algebraic constructions and can rapidly generate large-scale sequence sets that meet zero cross-correlation and zero autocorrelation sidelobe conditions in a data-driven manner.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method, system and device for generating a large-capacity mutually orthogonal complementary sequence set. Background Art

[0002] Mutually Orthogonal Complementary Sequence Sets (MOCSS) are a family of sequences with strict complementarity. They have important applications in fields such as communications, radar, sonar, and bioinformatics. Their core value lies in their ability to simultaneously satisfy both complementarity and orthogonality. In Code Division Multiple Access (CDMA), Non-Orthogonal Multiple Access (NOMA), and 6G communications, large numbers of users share the same frequency band. Therefore, mutually orthogonal or low-correlation spreading codes must be used to distinguish users and avoid interference.

[0003] In cutting-edge explorations of modern communications and radar technologies, mutually orthogonal complementary sequence sets, with their unique zero cross-correlation and zero autocorrelation sidelobe characteristics, have become a core element supporting breakthrough system performance. In 6G communications, the development of high-frequency spectrum resources, the widespread adoption of massive Multiple-Input Multiple-Output (MIMO) technology, and the demand for high-density user access in non-orthogonal multiple access (NOMA) systems all place stringent demands on the capacity and generation efficiency of mutually orthogonal complementary sequence sets. Only orthogonal complementary sequence sets of sufficient size can achieve ideal isolation of thousands or even tens of thousands of user signals within a limited spectrum, avoiding signal misjudgment and system capacity bottlenecks caused by Multiple Access Interference (MAI).

[0004] Take active phased array radars, for example. They must simultaneously track hundreds of dynamic targets and mitigate enemy interference in complex electromagnetic environments. This relies on massive orthogonal coding sequences to achieve beamforming and target resolution. High-resolution imaging in synthetic aperture radar (SAR) requires long sequences to maintain strict orthogonality in the face of multipath fading to avoid misidentification of false targets. Therefore, generating large-capacity MOCSS sequences is crucial. With the continuous advancement of communications, radar, and sonar, the demand for large-capacity MOCSS is increasing. In communications, whether CDMA, NOMA, or 6G, large-capacity MOCSS is required to support more users and improve system performance. In radar and sonar systems, large-capacity MOCSS can better handle complex target detection scenarios. Only by achieving large-capacity MOCSS generation can its advantages in various fields be fully realized, meeting growing technological demands, and promoting further development in related fields.

[0005] This contradiction is particularly evident in the implementation of this technology: 6G test network measurements show that MOCSS generated by traditional methods can only support 256 simultaneous users, far below the theoretically designed capacity of 10,000. In a strong clutter environment, the multi-target recognition accuracy of a certain type of shipborne radar plummeted from 95% to 68% due to insufficient orthogonal sequence size. Clearly, overcoming the constraints of algebraic construction and achieving efficient generation of large-scale, arbitrary-length MOCSS has become a key obstacle to the next generation of communications and radar technologies.

[0006] Existing MOCSS construction methods primarily include Hadamard matrix construction and HP-GAN-based generation schemes. Hadamard matrix construction methods, including direct construction, Kronecker product-based construction, and recursive construction, rely heavily on algebraic structures such as finite fields and group theory. Sequence lengths must meet specific requirements (such as powers of 2), and computational complexity increases exponentially with scale. When practical requirements exceed the limitations of traditional algebraic structures (for example, requiring sequences with a length of 1023 or non-powers of 2), existing methods often fail to achieve construction failure, limiting user access to the communication system and reducing radar target detection accuracy.

[0007] The Hp-GAN generation scheme combines a generative adversarial network with a Hopfield neural network to approximate the generation of MOCSS. It uses a Hopfield neural network to screen sequences, iteratively updates the sequence state, and evaluates the sequence set using the metric function MC. Sequence sets with an MC value of 0 are selected as MOCSS. Hp-GAN generates MOCSS by alternating between training the GAN and the Hopfield neural network for iterative optimization. However, it requires many rounds of training before it can even achieve the first set of sequences with an MC value of 0. This means that multiple rounds of training are required before generating MOCSS. The training process also suffers from poor convergence characteristics, and its feature extraction and expression capabilities are insufficient. This is particularly true when dealing with long sequences and sequences with complex structures, resulting in low generated sequence quality and search efficiency.

[0008] Therefore, there is an urgent need for a MOCSS generation technology with faster and more stable training, stronger feature learning adaptability, and higher sequence quality and search efficiency. Summary of the Invention

[0009] The purpose of the present invention is to overcome the technical problems existing in the prior art and provide a method, system and device for generating a large-capacity mutually orthogonal and complementary sequence set. By leveraging the LSTM modeling capability of the long-term dependencies of sequences and combining iterative optimization of the orthogonal and complementary constraints with the Hopfield neural network, the present invention breaks through the number theory limitations of traditional algebraic constructions and can quickly generate a large-scale sequence set that meets the conditions of zero cross-correlation and zero autocorrelation sidelobes in a data-driven manner.

[0010] The object of the present invention is achieved through the following technical solutions:

[0011] In a first aspect, a method for generating a large-capacity mutually orthogonal complementary sequence set is provided, comprising:

[0012] Use long short-term memory network LSTM training to generate the initial sequence set;

[0013] Reshaping the initial sequence set into a symmetric weight matrix through a Hopfield neural network and performing asynchronous state updates;

[0014] Grouping and cutting the updated sequence set to generate a pending sequence set;

[0015] The candidate sequence set is evaluated based on the MC index, and the final mutually orthogonal and complementary sequence set is screened.

[0016] In some embodiments, the long short-term memory network LSTM is defined as:

[0017]

[0018] Where y represents the output and X represents the input tensor. For LSTM layer operation, is the weight matrix of the LSTM layer, is the weight matrix of the fully connected layer, is the bias vector, represents the hyperbolic tangent activation function.

[0019] In some embodiments, the generating an initial sequence set using a long short-term memory network (LSTM) training comprises:

[0020] Use the Adam optimizer to minimize the mean squared error loss and use the cosine annealing learning rate scheduler to dynamically adjust the learning rate to generate a continuous value sequence;

[0021] The continuous value sequence is binarized by applying a sign function.

[0022] In some embodiments, the symmetric weight matrix is:

[0023]

[0024] in, Reshape the matrix. Represents a vector of length 1024, i represents a continuous sequence The i-th index in .

[0025] In some embodiments, the asynchronous status update includes:

[0026] Define the Hopfield neural network update function and randomly select neurons for update; the update function is:

[0027]

[0028] in, represents the state of node i at the next time t+1, Represents a sign function, which outputs +1 when the input is positive, -1 when it is negative, and usually 0 or +1 when it is 0. represents the connection weight from node j to i, represents the state of node j at time t, represents the state of node j at time t+1, j represents the node sequence number, and I represents the index set. In some embodiments, the MC index is calculated as:

[0029]

[0030] in, represents the autocorrelation index, represents the cross-correlation index, Represents a rollover indicator.

[0031] In some embodiments, the screening to obtain a final set of mutually orthogonal and complementary sequences comprises:

[0032] like If the sequence does not exist in the historical optimal sequence set S, it will be included in the historical optimal sequence set S.

[0033] In a second aspect, a system for generating a large-capacity mutually orthogonal complementary sequence set is provided, comprising:

[0034] LSTM sequence generation module, used to generate the initial sequence set using long short-term memory network LSTM training;

[0035] A Hopfield neural network optimization module, configured to reshape the initial sequence into a symmetric weight matrix through a Hopfield neural network and perform asynchronous state updates;

[0036] The sequence grouping and cutting module is used to group and cut the updated sequence set to generate a pending sequence set;

[0037] The index evaluation module is used to evaluate the pending sequence set based on the MC index, and screen out the final mutually orthogonal and complementary sequence set.

[0038] In a third aspect, an electronic device is provided, comprising a memory and a processor, wherein the memory stores computer instructions that can be executed on the processor, and when the processor executes the computer instructions, the method for generating a large-capacity mutually orthogonal complementary sequence set described in the first aspect is executed.

[0039] It should be further explained that the technical features corresponding to the above embodiments can be combined or replaced with each other to form a new technical solution if there is no conflict.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] The present invention uses a long short-term memory network (LSTM) to train and generate an initial sequence set, and uses a Hopfield neural network for optimization. Specifically, the initial sequence set is reshaped into a symmetric weight matrix and asynchronously updated. Finally, the pending sequence set is evaluated based on the MC indicator to screen out the final mutually orthogonal and complementary sequence set. With the LSTM's ability to model long-term sequence dependencies, combined with the Hopfield network's iterative optimization of orthogonal and complementary constraints, it breaks through the number theory limitations of traditional algebraic constructions and can quickly generate large-scale sequence sets that meet the conditions of zero mutual correlation and zero autocorrelation sidelobes in a data-driven manner. At the same time, compared with the Hp-GAN network, its generation quantity has increased by about 21%, the convergence speed has increased by about 96%, and the first round of MOCSS generation has been advanced from 2,300 rounds to the first round. The training stability has been significantly improved, the feature learning adaptability has been stronger, and the sequence quality and search efficiency have been higher. This innovation provides a key technical path for the spectrum efficiency revolution of 6G communications and the upgrade of the anti-interference capability of radar systems, and provides technical support for large-scale MOCSS application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of a flow chart of a method for generating a large-capacity mutually orthogonal complementary sequence set according to the present invention;

[0043] Figure 2 Schematic diagram of the HP-LSTM network structure of the present invention;

[0044] Figure 3 Schematic diagram of the training process of Hp-LSTM of the present invention;

[0045] Figure 4 Schematic diagram comparing Hp-LSTM and Hp-GAN of the present invention;

[0046] Figure 5is the generator loss graph of Hp-GAN;

[0047] Figure 6 is the discriminator loss graph of Hp-GAN;

[0048] Figure 7 This is the loss graph of Hp-LSTM. DETAILED DESCRIPTION

[0049] The technical solutions of the present invention are described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings herein can be arranged and designed in various different configurations. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0050] It should be noted that the defects existing in the solutions in the above-mentioned prior art are the results obtained by the inventor after practice and careful research. Therefore, the discovery process of the above-mentioned problems and the solutions proposed in the embodiments of this application below for the above-mentioned problems should be the contributions made by the inventor to this application in the process of invention and creation, and should not be understood as technical contents known to technical personnel in this field.

[0051] In response to the technical problems pointed out in the background technology, the embodiments provided by the present invention are as follows:

[0052] Reference Figure 1 In an exemplary embodiment, a method for generating a large-capacity mutually orthogonal complementary sequence set is provided, comprising:

[0053] Use long short-term memory network LSTM training to generate the initial sequence set;

[0054] Reshaping the initial sequence set into a symmetric weight matrix through a Hopfield neural network and performing asynchronous state updates;

[0055] Grouping and cutting the updated sequence set to generate a pending sequence set;

[0056] The candidate sequence set is evaluated based on the MC index, and the final mutually orthogonal and complementary sequence set is screened.

[0057] Long Short-Term Memory (LSTM) is a special recurrent neural network (RNN) architecture that aims to solve the gradient vanishing and gradient exploding problems faced by traditional RNNs when processing long sequence data, thereby effectively capturing long-term dependencies in the sequence. The core structure of LSTM consists of a memory cell, an input gate, a forget gate, and an output gate. The gating mechanism of LSTM can adaptively and selectively retain or forget historical information, effectively alleviating the defects of traditional RNNs in long sequence modeling. At the same time, compared with the Nash equilibrium that the GAN network needs to achieve, the training difficulty and time cost are greatly reduced. Based on this, the present invention uses LSTM's ability to model long-term sequence dependencies, combined with the iterative optimization of the orthogonal complementary constraints of the Hopfield neural network. A better generation algorithm for mutually orthogonal complementary sequence sets (MOCSS) is implemented, referred to as Hp-LSTM, with a structure as follows: Figure 2 shown.

[0058] Specifically, the LSTM input is a three-dimensional tensor ,in, is the number of samples, the time step is set to 1, and the feature dimension is 1024. The long short-term memory network LSTM is defined as:

[0059]

[0060] Where y represents the output and X represents the input tensor. For LSTM layer operation, is the weight matrix of the LSTM layer, is the weight matrix of the fully connected layer, is the bias vector, represents the hyperbolic tangent activation function.

[0061] Furthermore, the model training process adopts the mean square error (MSE) loss function:

[0062]

[0063] Where L is the loss, represents the batch size, is the true value, is the predicted value, is the square of the norm.

[0064] Use Adam optimizer with learning rate To minimize the loss, due to the limited characteristics of the binary MOCSS sequence, the present invention uses a cosine annealing learning rate scheduler to dynamically adjust the learning rate to avoid falling into the local optimum as much as possible, and finally generates a continuous value sequence, and binarizes the continuous value sequence by applying a sign function.

[0065] Furthermore, in the sequence screening stage, after a certain number of training rounds (e.g., 500 rounds, as HP-LSTM converges quickly, the number of evaluation rounds can be reduced), the trained LSTM is used to predict the input data (e.g., using LSTM to predict 200 sequences of length 1024). The shape is then reshaped to [200, 32, 32] as the symmetric weight matrix of the Hopfield neural network. The symmetric weight matrix is:

[0066]

[0067] in, Reshape the matrix. Represents a vector of length 1024, i represents a continuous sequence The i-th index / element in . Further, define the Hopfield neural network update function:

[0068]

[0069] in, represents the state of node i at the next time t+1, Represents a sign function, which outputs +1 when the input is positive, -1 when it is negative, and usually 0 or +1 when it is 0. represents the connection weight from node j to i, represents the state of node j at time t, Represents the state of node j at time t+1, j represents the node number, and I represents the index set. By randomly initializing a binary sequence of length 32 , and perform a certain number of iterative updates (such as 200 rounds). At the same time, the evaluation function MC is used to evaluate the sequence, and the MC index is calculated as:

[0070]

[0071] in, represents the autocorrelation index, represents the cross-correlation index, Represents a rollover indicator.

[0072] Ruoruo If the sequence does not exist in the historical optimal sequence set S, it will be included in the historical optimal sequence set S. Finally, the set of historical optimal sequence sets S with the longest length is saved to generate a MOCSS with a larger capacity.

[0073] In another exemplary embodiment, a specific example of the above method is given, which includes the following steps:

[0074] Step 1: Generate the initial sequence set of Hp-LSTM

[0075] 1) Given training data (500 samples, time step 1, feature 1024), every 100 rounds of training, through the HP-LSTM model Generate a sequence of 100 consecutive values:

[0076] ,in, are the parameters of the fully connected layer.

[0077] 2) Apply a sign function to the output:

[0078]

[0079] Step 2: Construct a symmetric weight matrix

[0080] Reshape the 1024-dimensional sequence into a symmetric weight matrix.

[0081] Step 3: Asynchronous state updates

[0082] Randomly select 8 neurons to update, each time selecting the index set . Update using the above Hopfield neural network update function.

[0083] Step 4: Sequence grouping and cutting

[0084] Split the 1024-dimensional sequence into , forming a MOCSS pending sequence set with L = 8, M = 2, and K = 2. L is the sequence length, M is the number of tuples, which determines the number of users, and K is the number of group sequences, which determines the number of subcarriers.

[0085] Step 5: MC indicator calculation

[0086] 1) Calculate autocorrelation, autocorrelation function (delay ) is expressed as:

[0087]

[0088] in, Representation sequence The autocorrelation value of Indicates delay, Represents the index of the sequence element; the autocorrelation values ​​of the remaining sequences are calculated in the same way.

[0089] From this we can get the autocorrelation index:

[0090]

[0091] 2) Calculate cross-correlation, cross-correlation function (delay ) is expressed as:

[0092] ,in, Representation sequence and The cross-correlation value of Represents the index of the sequence element; and so on to calculate the mutual correlation value of any two sequences;

[0093] From this, the mutual correlation index can be obtained as:

[0094]

[0095] 3) Calculate flip correlation, flip dot product (delay ) is expressed as:

[0096] ,in, Representation sequence The flip correlation value of Represents the index of the sequence element, and so on to calculate the flip correlation value of the remaining sequences;

[0097] From this we can get the reversal index:

[0098]

[0099] The total MC index is calculated as:

[0100]

[0101] If the sequence set evaluation computes its , if this sequence set has never appeared (does not exist in the historical sequence cluster), it will be added to the candidate sequence cluster.

[0102] Step 6: When the maximum number of training rounds (e.g., 10,000) is reached, the training is terminated and the candidate sequence clusters are saved. Finally, 95 new MOCSSs (not present in the historical sequence set) are obtained.

[0103] The effect of the HP-LSTM of this application is evaluated below.

[0104] Figure 3The main training process of HP-LSTM is shown. The horizontal axis represents the training round, and the vertical axis represents the MC index. After each round of training, 100 sets of pending sequences are generated, and the average value of MC in this round is calculated ( ) and the minimum value ( ).

[0105] From the figure we can see that Hp-LSTM in the 8th round, The first time it reaches 0, it means that the model has found The sequence set is effectively learned and has been found since then. The ability of the HP-LSTM to extract sequence features is demonstrated in Figure 2. Meanwhile, its MC average value gradually decreases, reaching a plateau at 40 rounds, and then fluctuates slightly between 10 and 20 rounds. This shows that HP-LSTM can quickly extract sequence features in a very short training round, and its MC average value converges to a low value very quickly, indicating that HP-LSTM is suitable for sequence generation tasks.

[0106] like Figure 5-Figure 6 As shown, the training process of Hp-LSTM is compared with that of Hp-GAN. Figure 5 and Figure 6 These are the generator loss graph and discriminator loss graph of Hp-GAN, respectively. The horizontal axis represents the training round, and the vertical axis represents the binary cross entropy loss. Figure 7 This is the mean squared error loss plot for the HP-LSTM. The horizontal axis represents training rounds, and the vertical axis represents mean squared error loss. A comparison shows that the HP-LSTM converges very quickly, completing convergence in around 100 rounds, while the HP-GAN barely converges after 3000 rounds, while the generator loss continues to fluctuate and rise. This demonstrates that the HP-LSTM is significantly superior to the HP-GAN in sequence generation and feature extraction. In terms of convergence speed, it is far more stable than the HP-GAN network.

[0107] like Figure 4 As shown, by comparing the MC curve, it can be found that Hp-GAN found the first For a set of sequences, HP-LSTMM was able to generate MOCSS from the very first round, demonstrating that LSTM has advantages over GAN in matrix feature learning, especially when generating longer and more complex sequences. Furthermore, the average MC value of HP-LSTM reached the level of HP-LSTM trained for 6000 rounds at the beginning of training, and its average MC value remained lower than that of HP-GAN throughout the entire training process, further confirming the advantages of LSTM.

[0108] Finally, using the same training set, compared to Hp-GAN which found 119 new MOCSS sequence sets in 10,000 rounds of training, our Hp-LSTM method found 144 MOCSS sequence sets in 4 training rounds.

[0109] In summary, the Hp-GAN algorithm demonstrates significant performance advantages over the Hp-GAN algorithm. Judging from the convergence characteristics of the training process, its stability and speed of training are significantly superior to those of the Hp-GAN. Comparing the MC (sequence metric) curves of the two shows that the Hp-GAN only achieved its first set of sequences with an MC value of 0 after approximately 2500 rounds of training, indicating that it began to generate sequences that met the strict requirements of MOCSS. In contrast, the Hp-LSTM demonstrated the ability to generate MOCSS from the very first round of training, demonstrating the LSTM network's greater adaptability in matrix feature learning. In particular, its feature extraction and representation capabilities significantly outperform the GAN architecture when processing long sequences and sequences with complex structures.

[0110] HP-LSTM also demonstrates significant advantages in terms of sequence quality and search efficiency. The average MC value of the HP-LSTM during initial training reaches the level of HP-GAN after 6,000 rounds of training, fully demonstrating its superior sequence optimization efficiency. Under the same training set conditions, the original HP-GAN paper only discovered 119 MOCSS sequence sets after 10,000 rounds of training, while the HP-LSTM successfully obtained 144 MOCSS sequence sets after four independent training rounds, significantly reducing the number of training rounds and increasing the number of sequence discoveries by approximately 21%. This result demonstrates that HP-LSTM, through its deep fusion of LSTM and discrete Hopfield neural networks, has achieved a dual breakthrough in sequence search efficiency and quality, making it more adaptable to large-scale communication scenarios in the future. It also provides a more efficient and stable solution for CSS sequence generation and related applications.

[0111] In another exemplary embodiment, based on the same inventive concept as the method embodiment, a system for generating a large-capacity mutually orthogonal complementary sequence set is provided, comprising:

[0112] LSTM sequence generation module, used to generate the initial sequence set using long short-term memory network LSTM training;

[0113] A Hopfield neural network optimization module is used to reshape the initial sequence set into a symmetric weight matrix through a Hopfield neural network and perform asynchronous state update;

[0114] The sequence grouping and cutting module is used to group and cut the updated sequence set to generate a pending sequence set;

[0115] The index evaluation module is used to evaluate the pending sequence set based on the MC index, and screen out the final mutually orthogonal and complementary sequence set.

[0116] In another exemplary embodiment, based on the same inventive concept as the method embodiment, an electronic device is provided, including a memory and a processor, wherein the memory stores computer instructions that can be executed on the processor, and when the processor executes the computer instructions, it executes a method for generating a large-capacity mutually orthogonal complementary sequence set provided in an embodiment of the present invention.

[0117] The processor may be a single-core or multi-core central processing unit or a specific integrated circuit, or one or more integrated circuits configured to implement the present invention.

[0118] Embodiments of the subject matter and functional operations described in this specification may be implemented in: tangibly embodied computer software or firmware, computer hardware including the structures disclosed in this specification and their structural equivalents, or a combination of one or more thereof. Embodiments of the subject matter described in this specification may be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory program carrier for execution by a data processing apparatus or to control the operation of the data processing apparatus. Alternatively or in addition, the program instructions may be encoded on an artificially generated propagated signal, such as a machine-generated electrical, optical, or electromagnetic signal, that is generated to encode and transmit information to a suitable receiver apparatus for execution by the data processing apparatus.

[0119] The processes and logic flows described in this specification can be performed by one or more programmable computers executing one or more computer programs to perform the corresponding functions by operating on input data and generating output. The processes and logic flows can also be performed by, and apparatus can be implemented as, special purpose logic circuitry, such as an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit).

[0120] Processors suitable for executing computer programs include, for example, general-purpose and / or special-purpose microprocessors, or any other type of central processing unit. Typically, a central processing unit will receive instructions and data from a read-only memory and / or random access memory. The basic components of a computer include a central processing unit for implementing or executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include one or more mass storage devices for storing data, such as magnetic disks, magneto-optical disks, or optical disks, or the computer will be operably coupled to such a mass storage device to receive data from it or to transmit data to it, or both. However, a computer does not necessarily have such a device. In addition, a computer can be embedded in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device such as a universal serial bus (USB) flash drive, to name a few.

[0121] It should be understood that each box in the flow chart or block diagram can represent a module, program segment or a part of code, and the part of the module, program segment or code comprises one or more executable instructions for realizing the logical function of the provision. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a sequence different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flow chart, and the combination of the boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the function or action of the provision, or can be implemented with a combination of dedicated hardware and computer instructions.

[0122] The above specific implementation methods are detailed descriptions of the present invention. It cannot be considered that the specific implementation methods of the present invention are limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, they can make several simple deductions and substitutions without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A method for generating a large-capacity mutually orthogonal complementary sequence set, characterized in that: include: An initial sequence set is generated using a long short-term memory (LSTM) network for training; the input data of the LSTM network is a three-dimensional tensor of communication and radar field data; The initial sequence set is reshaped into a symmetric weight matrix through the Hopfield neural network and asynchronous state update is performed; the symmetric weight matrix is: ,in, Reshape the matrix. Represents a vector of length 1024, i represents a continuous sequence The i-th index in ; Grouping and cutting the updated sequence set to generate a pending sequence set; The candidate sequence set is evaluated based on the MC index to screen out the final mutually orthogonal and complementary sequence set; the MC index is calculated as: ,in, represents the autocorrelation index, represents the cross-correlation index, Represents a rollover indicator.

2. The method for generating a large-capacity mutually orthogonal complementary sequence set according to claim 1, wherein: The long short-term memory network LSTM is defined as: , where y represents the output and X represents the input tensor. For LSTM layer operation, is the weight matrix of the LSTM layer, is the weight matrix of the fully connected layer, is the bias vector, represents the hyperbolic tangent activation function.

3. The method for generating a large-capacity mutually orthogonal complementary sequence set according to claim 2, characterized in that: The long short-term memory network LSTM is used to train and generate an initial sequence set, including: Use the Adam optimizer to minimize the mean squared error loss and use the cosine annealing learning rate scheduler to dynamically adjust the learning rate to generate a continuous value sequence; The continuous value sequence is binarized by applying a sign function.

4. The method for generating a large-capacity mutually orthogonal complementary sequence set according to claim 1, wherein: The asynchronous status update includes: Define the Hopfield neural network update function and randomly select neurons for update; the update function is: ,in, represents the state of node i at the next time t+1, Represents a sign function, which outputs +1 when the input is positive, -1 when it is negative, and usually 0 or +1 when it is 0. represents the connection weight from node j to i, represents the state of node j at time t, Represents the state of node j at time t+1, j represents the node sequence number, and I represents the index set.

5. The method for generating a large-capacity mutually orthogonal complementary sequence set according to claim 1, wherein: The screening results in a final mutually orthogonal and complementary sequence set, comprising: like If the sequence does not exist in the historical optimal sequence set S, it will be included in the historical optimal sequence set S.

6. A system for generating a large-capacity mutually orthogonal complementary sequence set, characterized in that: include: An LSTM sequence generation module is used to generate an initial sequence set using a long short-term memory network (LSTM) training; the input data of the LSTM is a three-dimensional tensor of communication and radar field data; A Hopfield neural network optimization module is used to reshape the initial sequence set into a symmetric weight matrix through a Hopfield neural network and perform asynchronous state update; the symmetric weight matrix is: ,in, Reshape the matrix. Represents a vector of length 1024, i represents a continuous sequence The i-th index in ; The sequence grouping and cutting module is used to group and cut the updated sequence set to generate a pending sequence set; An index evaluation module is used to evaluate the pending sequence set based on the MC index to screen out the final mutually orthogonal and complementary sequence set; the MC index is calculated as: ,in, represents the autocorrelation index, represents the cross-correlation index, Represents a rollover indicator.

7. An electronic device comprising a memory and a processor, wherein the memory stores computer instructions that can be executed on the processor, characterized in that: When the processor runs the computer instructions, it executes the method for generating a large-capacity mutually orthogonal complementary sequence set as described in any one of claims 1 to 5.

Images