A signal feature compression method based on codebook discrete quantization and multi-task learning

Abstract

The application discloses a signal feature compression method based on codebook discrete quantization and multi-task learning, and belongs to the technical field of cross between signal processing and artificial intelligence. In view of the problem that the prior art is difficult to consider compression efficiency, semantic reservation and computational complexity in high-dimensional IQ signal compression, the application extracts high-dimensional continuous features of the original IQ signal through an encoder; a learnable codebook is used for vector quantization to generate discrete indexes and calculate quantization loss; the discrete features are input into a decoder branch to reconstruct the signal and calculate reconstruction loss, and are simultaneously input into a classifier branch to predict categories and calculate classification loss; a multi-layer perceptron is used to process traditional features and coding features, a cosine similarity calculation is used to calculate comparison loss; the model is trained by jointly optimizing quantization loss, reconstruction loss, classification loss and comparison loss; and finally, the discrete indexes are output as compressed features. The method realizes efficient compression and classification semantic reservation, and is suitable for wireless communication, Internet of Things and other scenes.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of signal processing and artificial intelligence, and particularly relates to a signal feature compression method based on codebook discrete quantization and multi-task learning. BACKGROUND

[0002] In the field of wireless communication technology, with the large-scale deployment of 5G networks and the explosive growth of Internet of Things devices, massive high-dimensional IQ signal data puts a huge pressure on storage and transmission bandwidth. Traditional signal feature compression methods have significant defects: vector quantization technology realizes feature compression through codebook mapping, but its computational complexity is extremely high, resulting in real-time signal processing delay; the large quantization error causes loss of signal details, which is difficult to meet the needs in scenes such as radar signal analysis that require strict accuracy; more importantly, this method cannot effectively preserve the classification discriminative information in the signal, directly affecting the accuracy of subsequent identification tasks.

[0003] Although the product quantization technology reduces the computational burden through feature space decomposition, its semantic preservation ability is insufficient in complex wireless environments (such as multipath interference scenarios), and the compressed features are difficult to support high-precision modulation recognition and other applications. Deep learning methods based on autoencoders focus on signal reconstruction, but the single reconstruction objective ignores the classification semantic demand, resulting in a decrease in the effectiveness of compressed features in device identity authentication and other tasks.

[0004] Existing improvement schemes attempt to combine multi-task learning to jointly optimize reconstruction and classification loss, but they still do not solve the fusion problem of traditional artificial features and deep learning features, and the comprehensive nature of feature representation is insufficient, limiting the generalization ability in scenes such as electromagnetic spectrum monitoring that require comprehensive multi-dimensional features. Contrastive learning improves feature discriminativeness in the image field, but its combination with discrete quantization technology still faces challenges: how to balance the relationship between contrastive loss and quantization loss while ensuring quantization efficiency, and avoid semantic information loss, which becomes a bottleneck for actual deployment.

[0005] Therefore, the current signal feature compression technology faces three contradictions: 1) the balance between computational efficiency and quantization accuracy in the process of discrete quantization of high-dimensional continuous features; 2) the conflict between compression operation and classification discriminative information preservation; 3) the lack of discriminativeness caused by the split of traditional features and deep learning feature representation. These defects are particularly prominent in low-bandwidth Internet of Things edge nodes, real-time military communication, and other scenarios, and urgent breakthrough solutions are needed. SUMMARY

[0006] To solve the above technical problems, the present application proposes a signal feature compression method based on codebook discrete quantization and multi-task learning to solve the problems existing in the prior art.

[0007] The first aspect, to achieve the above object, the present application provides a kind of signal feature compression method based on codebook discrete quantization and multi-task learning, comprising the following steps:

[0008] Input original IQ signal, extract high-dimensional continuous features by encoder;

[0009] Vector quantization is carried out on high-dimensional continuous features using codebook, and the nearest neighbor search is mapped to discrete index, and the corresponding codebook vector is output and the quantization error loss is calculated;

[0010] Discrete feature representation is input into decoder and classifier respectively, decoder reconstructs signal and calculates reconstruction loss, and classifier predicts signal class and calculates classification loss;

[0011] Multi-layer perceptron is used to process traditional features of original signal and high-dimensional continuous features extracted by encoder respectively, and comparison loss is calculated;

[0012] Quantization loss, reconstruction loss, classification loss and comparison loss are summed to obtain total loss function, and model is trained based on total loss function;

[0013] After training, discrete index is output as compressed feature for storage, transmission or classification task.

[0014] Optionally, the process of extracting high-dimensional continuous features by the encoder comprises:

[0015] IQ signal input in complex form is received;

[0016] The signal is processed by three-layer convolution structure, wherein the first two layers use down-sampling convolution, and the third layer keeps the size unchanged;

[0017] ReLU activation function is used and residual module is stacked to enhance feature representation, and high-dimensional continuous feature vector of fixed dimension is output.

[0018] Optionally, the process of vector quantization comprises:

[0019] A learnable codebook is maintained, and the Euclidean distance between high-dimensional continuous features and all codebook vectors is calculated;

[0020] The nearest neighbor codebook vector is selected to replace the original feature, and discrete index representation is generated;

[0021] The mean square error between the original feature and the selected codebook vector is calculated as quantization loss.

[0022] Optionally, the process of reconstructing signal by the decoder comprises:

[0023] Quantized discrete features are processed by convolution layer;

[0024] Signal representation is reconstructed through multi-layer residual module.

[0025] The original IQ signal is reconstructed by transposed convolution upsampling, and a reconstruction mean square error is calculated.

[0026] Optionally, the process of calculating the contrast loss comprises:

[0027] Traditional features of the original signal are extracted, including statistical features, instantaneous features and spectral features;

[0028] The traditional features and the encoder features are respectively mapped to a common space by a multi-layer perception;

[0029] Cosine similarity of the two is calculated, and a InfoNCE contrast loss function is used to optimize feature alignment.

[0030] Optionally, the process of jointly training the model comprises:

[0031] The quantization loss, the reconstruction loss, the classification loss and the contrast loss are aggregated to form a total optimization target;

[0032] A back propagation algorithm is used to update parameters of the encoder, the codebook, the decoder, the classifier and the multi-layer perception;

[0033] A cosine annealing learning rate scheduling strategy is used to adjust the training process.

[0034] In a second aspect, the present application further provides a signal feature compression system based on codebook discrete quantization and multi-task learning, which is used to implement a signal feature compression method based on codebook discrete quantization and multi-task learning, and the system comprises:

[0035] A feature extraction module is configured to receive an original IQ signal and extract high-dimensional continuous features by an encoder;

[0036] A vector quantization module is configured to perform vector quantization on the high-dimensional continuous features by using a codebook, map the high-dimensional continuous features to discrete indexes by nearest neighbor search, output corresponding codebook vectors and calculate a quantization error loss;

[0037] A multi-task processing module comprises a decoding unit and a classification unit, the decoding unit is configured to reconstruct the signal and calculate a reconstruction loss, and the classification unit is configured to predict a signal category and calculate a classification loss;

[0038] A feature alignment module is configured to process traditional features of the original signal and the high-dimensional continuous features extracted by the encoder by a multi-layer perception respectively, and calculate a contrast loss;

[0039] A joint optimization module is configured to sum the quantization loss, the reconstruction loss, the classification loss and the contrast loss to obtain a total loss function, and jointly train model parameters based on the total loss function;

[0040] The compression output module is configured to output the discrete index as a compressed feature after the training is completed.

[0041] In a third aspect, the present application further provides a computer terminal device, comprising:

[0042] one or more processors;

[0043] a memory coupled to the processors, for storing one or more programs;

[0044] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the signal feature compression method based on codebook discrete quantization and multi-task learning in the first aspect.

[0045] In a fourth aspect, the present application further provides a computer readable storage medium, having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the signal feature compression method based on codebook discrete quantization and multi-task learning in the first aspect.

[0046] In a fifth aspect, the present application further provides a computer program product, comprising a computer program, wherein the computer program is executed by a processor to implement the steps of the signal feature compression method based on codebook discrete quantization and multi-task learning in the first aspect.

[0047] Compared with the prior art, the present application has the following advantages and technical effects:

[0048] The signal feature compression method based on codebook discrete quantization and multi-task learning provided by the present application realizes efficient compression of high-dimensional IQ signals through codebook discrete quantization, generates low-bit discrete indexes to significantly reduce storage and transmission overhead, combines a multi-task learning mechanism to simultaneously optimize reconstruction loss and classification loss during the compression process, ensures that the compressed features retain the classification discriminative semantics of the original signals, aligns traditional features and deep learning features through contrastive learning to enhance the discriminativeness and robustness of feature representation, and finally guarantees high compression rate while meeting the stringent requirements of low-bandwidth communication and edge computing scenarios for signal classification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0049] The accompanying drawings, which form a part of the present application, are intended to provide further understanding of the present application, and the illustrative embodiments thereof and their descriptions are used to explain the present application without imposing undue limitation on the same. In the drawings:

[0050] Figure 1 is a method flowchart of the embodiment of the present application;

[0051] Figure 2 is a model training framework diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0052] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0053] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0054] To overcome the core contradiction in the current signal processing field that high-dimensional feature compression and classification information reservation are difficult to balance, that is, traditional compression methods (such as PCA and conventional autoencoder) are prone to cause loss of classification semantics, and continuous feature representation is faced with high storage cost and weak anti-interference ability, etc. bottleneck, the present application proposes a signal feature compression method based on vector quantization codebook and multi-task collaborative training. The method realizes the extreme compression of high-dimensional signal features by constructing a learnable discrete codebook space; at the same time, it innovatively integrates three optimization objectives of reconstruction learning, classification supervision and feature comparison alignment, and accurately reserves the signal category information in the compression process. Through the quantization architecture of encoder-codebook-decoder, combined with the cross-domain knowledge transfer of traditional features and deep learning features, the method reserves the semantic features of the signal itself for classification under the premise of ensuring high compression rate, and provides an efficient and scalable signal processing solution for complex scenarios such as low-bandwidth communication and edge computing.

[0055] Embodiment one

[0056] As shown in Figure 1 , the present embodiment provides a signal feature compression method based on codebook discrete quantization and multi-task learning, comprising:

[0057] Input the original IQ signal, and extract high-dimensional continuous features through the encoder;

[0058] Vector quantize the high-dimensional continuous features using the codebook, map them to discrete indexes through nearest neighbor search, output the corresponding codebook vectors and calculate the quantization error loss;

[0059] Input the discrete feature representation into the decoder and the classifier respectively, the decoder reconstructs the signal and calculates the reconstruction loss, and the classifier predicts the signal category and calculates the classification loss;

[0060] Use a multi-layer perceptron to process the traditional features of the original signal and the high-dimensional continuous features extracted by the encoder respectively, and calculate the comparison loss;

[0061] Sum the quantization loss, reconstruction loss, classification loss and comparison loss to obtain a total loss function, and train the model based on the total loss function;

[0062] After training, the discrete index is output as a compressed feature for storage, transmission or classification tasks.

[0063] S1: Input the original IQ signal and extract it through an encoder.

[0064] S2: Vector quantize the high-dimensional continuous feature using a codebook, map it to a discrete index through nearest neighbor search, and output the corresponding codebook vector, and calculate the quantization error loss.

[0065] S3: Input the discrete feature representation into the decoder and classifier branches respectively, the decoder branch is used to reconstruct the signal and calculate the reconstruction loss, and the classifier branch is used for signal class prediction and calculates the classification loss.

[0066] S4: Use a multi-layer perceptron (MLP) to process the traditional features of the original signal and the high-dimensional continuous feature representation extracted by the encoder respectively, extract their contrast features, and then calculate the contrast loss.

[0067] S5: Sum the codebook loss, reconstruction loss, classification loss and contrast loss to obtain the total loss function, based on the total loss function, use optimization algorithms such as back propagation to jointly train the model and optimize the parameters of each module.

[0068] S6: After training, the discrete index is output as a compressed feature, which can be efficiently stored / transmitted while retaining the semantic information required for classification.

[0069] As an embodiment in this embodiment, the process of the encoder extracting high-dimensional continuous features includes:

[0070] Receive IQ signal input in complex form;

[0071] Process the signal through a three-layer convolution structure, of which the first two layers use down-sampling convolution and the third layer keeps the size unchanged;

[0072] Use the ReLU activation function and stack the residual module to enhance the feature representation, and output a high-dimensional continuous feature vector with a fixed dimension.

[0073] Specifically, the S1 includes the following contents:

[0074] S1.1: Encode the original signal to obtain a high-dimensional continuous feature.

[0075] Input the original IQ signal (N is the number of sampling points), extract the high-dimensional continuous feature through the encoder

[0076] (1)​

[0077] where D is the feature dimension, and is the high-dimensional continuous feature extracted by the encoder.

[0078] The more specific operation process is as follows: referring to Figure 1 , Figure 2 , receiving a complex number form IQ signal input containing in-phase components and quadrature components, and performing signal feature extraction through a convolutional neural network encoder module. The encoder module includes three layers of convolution processing structure: the first layer of convolution kernel size is 4 and the down-sampling factor is 2; the second layer of convolution kernel size is 4 and the down-sampling factor is 2; the third layer of convolution kernel size is 3 and the size is unchanged. Under the configuration of compression factors of 8 and 16, 1-2 additional down-sampling convolution layers are added. ReLU activation function is used after all convolution layers, and finally the residual module is stacked to enhance the feature representation ability, and a high-dimensional continuous feature vector of fixed dimension is output.

[0079] As an embodiment in this embodiment, the process of vector quantization includes:

[0080] Maintain a learnable codebook, and calculate the Euclidean distance between the high-dimensional continuous feature and all codebook vectors.

[0081] Select the nearest neighbor codebook vector to replace the original feature to generate a discrete index representation.

[0082] Calculate the mean square error between the original feature and the selected codebook vector as the quantization loss.

[0083] Specifically, the S2 includes the following contents:

[0084] S2.1: Vector quantization is performed on the high-dimensional continuous feature to output a discrete feature.

[0085] Discrete quantization of the high-dimensional continuous feature ( using a codebook of size , calculating the Euclidean distance between and all codebook vectors , and selecting the nearest neighbor index :

[0086] (2)

[0087] Output the discrete feature , where is the codebook vector indexed by q.

[0088] S2.2 Calculate the codebook quantization loss:

[0089] Use the square of the Euclidean distance before and after quantization as the loss function to calculate the codebook quantization loss :

[0090] (3)

[0091] where denotes the stop-gradient operation, is the weight coefficient.

[0092] The more specific operation process is as follows: referring to the accompanying drawings Figure 1 , Figure 2 , the high-dimensional feature output by the encoder is input into a vector quantization module. The module maintains a learnable embedding space as a codebook, which contains a preset number of discrete embedding vectors. By calculating the Euclidean distance between the input feature and all codebook vectors, the nearest neighbor codebook vector is selected to replace the original feature. This process generates a discrete index representation, and the mean square error between the original feature and the selected codebook vector is calculated as the quantization loss.

[0093] As an embodiment in this embodiment, the process of the decoder reconstructing the signal includes:

[0094] The quantized discrete feature is processed by a convolutional layer;

[0095] The signal representation is reconstructed by a multi-layer residual module;

[0096] The original IQ signal is reconstructed by upsampling using transpose convolution, and the reconstruction mean square error is calculated.

[0097] Specifically, the S3 includes the following contents:

[0098] S3.1: Decoder branch.

[0099] The quantized discrete feature is input into the decoder to obtain a reconstructed signal with the same shape as the original IQ input , , where the decoder calculation process is denoted, and the reconstruction loss is calculated using the original signal and the reconstructed signal:

[0100] (4)

[0101] S3.2: Classifier branch.

[0102] The quantized discrete feature is input into the classifier to output a class prediction result and calculate a classification loss:

[0103] (5)

[0104] where, is the true class label.

[0105] More specifically, the operation process is as follows: refer to the accompanying drawings Figure 1 、 Figure 2 The quantized feature vectors are input into two branches in parallel: a decoder branch and a classifier branch. The decoder first processes the features through a convolutional layer, then reconstructs the signal representation through multiple residual modules, and finally reconstructs the original IQ signal through multiple transposed convolutions to achieve upsampling, and calculates the reconstruction mean square error. The classifier branch contains two layers of convolutional processing structure, which extracts features through convolutional layers, compresses the space-time dimension through global average pooling, and finally outputs the classification probability distribution through a fully connected layer and calculates the cross-entropy loss.

[0106] As an embodiment in this embodiment, the process of calculating the contrast loss includes:

[0107] Extracting traditional features of the original signal, including statistical features, instantaneous features and spectral features;

[0108] Mapping the traditional features and the encoder features to a common space through multiple layer perceptrons, respectively;

[0109] Calculating the cosine similarity of the two, and using the InfoNCE contrast loss function to optimize the feature alignment.

[0110] Specifically, the S4 includes the following contents:

[0111] S4.1: Extracting traditional features of the original signal (including statistical features, instantaneous features and spectral features), and M is the dimension of the traditional feature vector. The traditional features and the high-dimensional continuous features output by the encoder are processed through two MLP projection heads, respectively :

[0112] (6)

[0113] wherein, is the vector representation of the projected traditional features, is the vector representation of the projected high-dimensional continuous features.

[0114] S4.2: Calculating the contrast loss using the projected traditional features and the encoded features :

[0115] (7)

[0116] wherein, is the cosine similarity, is the temperature coefficient, is the batch size.

[0117] More specifically, the operation process is as follows: refer to the accompanying drawings Figure 1 、Figure 2 The encoding feature vector and the traditional feature vector are processed by a projection network composed of two fully connected layers to map them into a 228-dimensional feature representation. The cosine similarity of the encoding features and the traditional features is calculated, and the InfoNCE contrastive loss function is used in the feature space to make the features of the same class sample close to each other and the features of different class samples far away from each other.

[0118] As an embodiment in the present embodiment, the process of jointly training the model includes:

[0119] The aggregation of the quantization loss, the reconstruction loss, the classification loss, and the contrastive loss forms a total optimization objective;

[0120] The parameters of the encoder, the codebook, the decoder, the classifier, and the multi-layer perceptron are updated using a backpropagation algorithm;

[0121] A cosine annealing learning rate scheduling strategy is used to adjust the training process.

[0122] Specifically, the S5 includes the following contents:

[0123] S5.1: The quantization loss, the reconstruction loss, the classification loss, and the contrastive loss calculated in the above steps are weighted and summed to obtain a total loss :

[0124] (8)

[0125] wherein, is the weight coefficient of each loss.

[0126] S5.2: Based on the total loss function training optimization, the parameters of the full model are updated by gradient backpropagation, including the encoder, the codebook, the decoder, the classifier, and the MLP of the contrastive learning module. The cosine annealing learning rate scheduling strategy is used to accelerate convergence and avoid local optimum.

[0127] The more specific operation process is as follows: refer to the attached Figure 1 、 Figure 2 The aggregation of the quantization loss, the reconstruction loss, the classification loss, and the contrastive loss forms a comprehensive optimization objective. The backpropagation algorithm with straight-through gradient estimation is used to update the parameters of all modules, including the convolutional layer weights, the residual module parameters, the codebook vectors, and the projection network parameters. The cosine annealing learning rate scheduling strategy is used, with an initial learning rate of 0.001, and the learning rate is periodically adjusted during the training process to accelerate convergence and avoid local optimum.

[0128] As an additional embodiment in the present embodiment, the step S6 includes the following contents:

[0129] S6.1: After the training is completed, the discrete index is output as a compressed feature. only need bit storage, meet compression efficiency; contains the classification discriminant information of the original signal, which can be directly used for downstream tasks, and the category semantic reservation is achieved.

[0130] The more specific operation process is as follows: refer to the accompanying Figure 1 、 Figure 2 After the model training is completed, the decoder and the contrast learning module are discarded in the deployment stage. The original signal is processed by the encoder to generate high-dimensional features, which are converted into discrete index sequences through vector quantization. In actual application, only the discrete index sequence needs to be stored or transmitted, and the receiver directly inputs the classifier after obtaining the codebook vector through table lookup to complete signal recognition.

[0131] Based on this, the signal feature compression method based on codebook discrete quantization and multi-task learning provided by the embodiment of the application realizes efficient compression of high-dimensional IQ signals through codebook discrete quantization, generates low-bit discrete indexes to significantly reduce storage and transmission overhead; combined with the multi-task learning mechanism, the reconstruction loss and the classification loss are optimized simultaneously in the compression process, ensuring that the compressed features retain the classification discriminant semantics of the original signal; through contrast learning, the traditional features and the deep learning features are aligned to enhance the discriminability and robustness of the feature representation; finally, while ensuring high compression rate, the stringent requirements of low-bandwidth communication and edge computing scenarios for signal classification accuracy are met.

[0132] As described above, the application is an application example of the application in actual situation for realizing feature compression and category semantic reservation of IQ signals. The application generates continuous feature vectors by inputting the original signal into the encoder, and obtains low-dimensional and compact discrete feature representation through codebook discrete quantization mapping; the decoder and the classifier respectively perform signal reconstruction and category prediction on the discrete features, and jointly optimize the reconstruction loss and the classification loss to balance information compression and semantic reservation; at the same time, a cross-modal contrast learning mechanism is designed, the deep features output by the encoder and the traditional features artificially extracted are respectively mapped to a common space through MLP, the semantic consistency of the two is constrained through the contrast loss, and the feature discriminability is enhanced; finally, the total loss is obtained by weighted summation of the codebook quantization loss, the reconstruction loss, the classification loss and the contrast loss, and multi-objective joint optimization is performed based on the total loss, so that the maximum reservation of the category key information of the signal in the efficient compression process is realized, and the discrete feature representation with high compression rate and strong separability is generated. The application is merely illustrative, not restrictive. Those skilled in the art understand that many changes, modifications, or even equivalents can be made within the spirit and scope defined by the claims of the application, but all will fall within the protection scope of the application.

[0133] Embodiment two

[0134] In this embodiment, a computer terminal device is provided, comprising:

[0135] one or more processors;

[0136] a memory coupled to the processors, storing one or more programs;

[0137] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-mentioned signal feature compression method based on codebook discrete quantization and multi-task learning.

[0138] In the embodiment, a computer readable storage medium is also provided, which stores a computer program. When the computer program is executed by a processor, the steps of the above-mentioned signal feature compression method based on codebook discrete quantization and multi-task learning are implemented.

[0139] In the embodiment, an electronic device is also provided, which includes a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to execute the steps of the above-mentioned signal feature compression method based on codebook discrete quantization and multi-task learning.

[0140] In the embodiment, a computer program product is also provided, which includes a computer program. When the computer program is executed by a processor, the steps of the above-mentioned signal feature compression method based on codebook discrete quantization and multi-task learning are implemented.

[0141] The above-mentioned program can be run in a processor, or can also be stored in a memory (or called a computer readable medium). The computer readable medium includes permanent and non-permanent, removable and non-removable media, and can be implemented by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0142] These computer programs can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer implemented process, so that the instructions executed on the computer or other programmable device provide a process for implementing the flow Figure 1 one flow or multiple flows and / or blocks Figure 1The steps of the functions specified in one block or multiple blocks can correspond to different steps, which can be implemented by different modules.

[0143] Such a device or system is provided in this embodiment. The system is called a signal feature compression system based on codebook discrete quantization and multi-task learning, which includes:

[0144] The feature extraction module is used to receive the original IQ signal and extract high-dimensional continuous features through an encoder.

[0145] The vector quantization module is used to vector quantize the high-dimensional continuous features using a codebook, map them to discrete indexes through nearest neighbor search, output corresponding codebook vectors, and calculate quantization error loss.

[0146] The multi-task processing module includes a decoding unit and a classification unit. The decoding unit reconstructs the signal and calculates the reconstruction loss, and the classification unit predicts the signal category and calculates the classification loss.

[0147] The feature alignment module is used to process the traditional features of the original signal and the high-dimensional continuous features extracted by the encoder using a multi-layer perception respectively, and calculate the contrast loss.

[0148] The joint optimization module is used to sum the quantization loss, reconstruction loss, classification loss, and contrast loss to obtain a total loss function, and train the model parameters based on the total loss function.

[0149] The compression output module is used to output the discrete index as the compressed feature after training is completed.

[0150] As an embodiment in this embodiment, the feature extraction module includes:

[0151] The signal receiving unit is used to receive the IQ signal in complex form.

[0152] The convolution processing unit includes a three-layer convolution structure. The first two layers perform down-sampling operations, and the third layer keeps the feature size unchanged.

[0153] The feature enhancement unit enhances the feature representation through ReLU activation function and residual module stacking, and outputs a high-dimensional continuous feature vector with fixed dimensions.

[0154] As an embodiment in this embodiment, the vector quantization module includes:

[0155] The codebook storage unit is used to maintain a set of learnable codebook vectors.

[0156] The distance calculation unit is used to calculate the Euclidean distance between the high-dimensional continuous features and all codebook vectors.

[0157] The index generation unit is configured to select a nearest neighbor codebook vector to generate a discrete index and to calculate a mean square error between the original feature and the selected vector as a quantization loss.

[0158] As an embodiment in the present embodiment, the decoding unit of the multi-task processing module comprises:

[0159] The feature reconstruction unit is configured to process the quantized discrete feature through a convolution layer.

[0160] The signal restoration unit is configured to reconstruct a signal representation through a multi-layer residual module.

[0161] The up-sampling output unit is configured to reconstruct the original IQ signal through a transpose convolution operation and to calculate a reconstruction mean square error.

[0162] As an embodiment in the present embodiment, the decoding unit of the multi-task processing module comprises:

[0163] The feature reconstruction unit is configured to process the quantized discrete feature through a convolution layer.

[0164] The signal restoration unit is configured to reconstruct a signal representation through a multi-layer residual module.

[0165] The up-sampling output unit is configured to reconstruct the original IQ signal through a transpose convolution operation and to calculate a reconstruction mean square error.

[0166] As an embodiment in the present embodiment, the joint optimization module comprises:

[0167] The loss aggregation unit is configured to integrate the quantization loss, the reconstruction loss, the classification loss and the contrast loss.

[0168] The parameter updating unit is configured to optimize the parameters of the encoder, the codebook, the decoder, the classifier and the multi-layer perceptron through a back propagation algorithm.

[0169] The learning rate scheduling unit is configured to dynamically adjust a training learning rate based on a cosine annealing strategy.

[0170] The system or device is configured to realize the functions of the method in the above embodiments, each module in the system or device corresponds to each step in the method, and the description has been made in the method and will not be repeated here.

[0171] Through the above embodiments, the problem of signal feature compression based on codebook discrete quantization and multi-task learning in the related art is solved, so as to ensure that the problems in the prior art are solved.

[0172] The above merely describes the preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A signal feature compression method based on codebook discrete quantization and multi-task learning, characterized in that, The method comprises the following steps: inputting an original IQ signal, extracting high-dimensional continuous features through an encoder; vector quantizing the high-dimensional continuous features using a codebook, mapping them into discrete indexes through nearest neighbor search, outputting corresponding codebook vectors and calculating quantization error loss; inputting the discrete feature representation into a decoder and a classifier respectively, reconstructing the signal through the decoder and calculating reconstruction loss, predicting the signal category through the classifier and calculating classification loss; calculating contrast loss by using a multi-layer perceptron to process traditional features of the original signal and high-dimensional continuous features extracted by the encoder respectively; summing the quantization loss, the reconstruction loss, the classification loss and the contrast loss to obtain a total loss function, and jointly training the model based on the total loss function; after the training is completed, outputting the discrete indexes as compressed features for storage, transmission or classification tasks; the process of calculating the contrast loss comprises: extracting traditional features of the original signal, including statistical features, instantaneous features and spectral features; mapping the traditional features and the encoder features into a common space through a multi-layer perceptron respectively; calculating the cosine similarity of the two, and optimizing the feature alignment using an InfoNCE contrast loss function.

2. The method of claim 1, wherein, the process of the encoder extracting high-dimensional continuous features comprises: receiving an IQ signal input in complex form; processing the signal through a three-layer convolution structure, wherein the first two layers use down-sampling convolution, and the third layer keeps the size unchanged; using a ReLU activation function and stacking residual modules to enhance feature representation, and outputting a high-dimensional continuous feature vector with a fixed dimension.

3. The method of claim 1, wherein, the process of vector quantization comprises: maintaining a learnable codebook, and calculating the Euclidean distance between the high-dimensional continuous features and all codebook vectors; selecting the nearest neighbor codebook vector to replace the original features, and generating a discrete index representation; calculating the mean square error between the original features and the selected codebook vector as the quantization loss.

4. The method of claim 1, wherein, the process of the decoder reconstructing the signal comprises: processing the quantized discrete features through a convolution layer; reconstructing the signal representation through multiple residual modules; using transposed convolution upsampling to reconstruct the original IQ signal, and calculating the reconstruction mean square error.

5. The method of claim 1, wherein, the process of jointly training the model comprises: aggregating the quantization loss, the reconstruction loss, the classification loss and the contrast loss to form a total optimization target; updating the parameters of the encoder, the codebook, the decoder, the classifier and the multi-layer perceptron using a backpropagation algorithm; adjusting the training process using a cosine annealing learning rate scheduling strategy.

6. A signal feature compression system based on codebook discrete quantization and multi-task learning, characterized in that, The system for implementing the method of any one of claims 1-5 comprises: a feature extraction module for receiving an original IQ signal and extracting high-dimensional continuous features through an encoder; a vector quantization module for vector quantizing the high-dimensional continuous features using a codebook, mapping them into discrete indexes through nearest neighbor search, outputting corresponding codebook vectors and calculating quantization error loss; a multi-task processing module including a decoding unit and a classification unit, the decoding unit reconstructing the signal and calculating reconstruction loss, and the classification unit predicting the signal category and calculating classification loss; a feature alignment module for calculating contrast loss by using a multi-layer perceptron to process traditional features of the original signal and high-dimensional continuous features extracted by the encoder respectively; The joint optimization module is configured to sum the quantization loss, the reconstruction loss, the classification loss, and the contrast loss to obtain a total loss function, and train the model parameters based on the total loss function; The compression output module is configured to output the discrete index as the compressed feature after the training is completed.

7. A computer terminal device, characterized by The computer program product comprises: one or more processors; a memory coupled to the processors for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method according to any one of claims 1-5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program product comprises:

9. A computer program product comprising a computer program, characterized in that, when the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method according to any one of claims 1-5. The computer program product comprises: when the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method according to any one of claims 1-5.

Images