Wireless bridge fixed link channel feature decoupling and abnormal state detection method and system based on dual-domain contrast learning

Abstract

The application discloses a wireless bridge fixed link channel feature decoupling and abnormal state detection method and system based on double-domain contrast learning, which constructs double-feature pairs of environment domain and channel domain, uses a cross-modal contrast learning mechanism to force double-domain features of the same samples to converge closely in a unified space and significantly separate different samples, can improve the inter-class distinguishability and intra-class cohesion of the feature space, and solves the problem that traditional single-modal features are sensitive to environmental interference. Based on the optimized double-domain feature pairs, the application uses a nonlinear decoupling network to separate environment-related components and environment-independent components, strips the interference of the environment on the channel features, highlights the inherent characteristics of the channel, provides purer feature representation for anomaly detection, improves the accuracy and interpretability of anomaly state detection, realizes the detection of the abnormal state of the wireless channel, and reduces the dependence on labeled data.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of wireless communication, in particular to a wireless bridge fixed link channel feature decoupling and abnormal state detection method and system based on dual-domain contrast learning. BACKGROUND

[0002] In a wireless communication scene in a signal-free area, the stability of a wireless bridge fixed link as key communication infrastructure directly affects the communication quality, and wireless channel feature decoupling and abnormal state detection in a complex environment face significant technical bottlenecks.

[0003] On the one hand, environmental factors such as terrain and weather are strongly coupled with wireless channel features, and traditional methods cannot effectively separate the mixed information of environmental domain and channel domain features, resulting in that similar samples in the feature space are dispersedly distributed due to environmental differences, and dissimilar samples are misassociated due to local feature similarity; on the other hand, device failures (such as hardware aging) and environmental mutations (such as newly added obstacles) are easily confused in the feature space, in addition, field deployment scenes lack labeled data, and traditional supervised learning is difficult to apply. These problems make it difficult for existing methods to accurately detect the abnormal state of the wireless channel, and an efficient decoupling and detection scheme is urgently needed.

[0004] Therefore, a wireless bridge fixed link channel feature decoupling and abnormal state detection method and system based on dual-domain contrast learning are needed. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a wireless bridge fixed link channel feature decoupling and abnormal state detection method and system based on dual-domain contrast learning. The method extracts wireless channel dual-domain feature pairs under an unsupervised learning framework, maps them to a unified space through cross-modal contrast learning, separates the environment-related and irrelevant components using a nonlinear decoupling, trains an anomaly detection model based on an autoencoder, extracts decoupled features from real-time data, and determines the abnormal type through reconstruction error.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] The wireless bridge fixed link channel feature decoupling and abnormal state detection method based on dual-domain contrast learning provided by the present application comprises the following steps:

[0008] S1, using a feature extractor to extract environment domain features and channel domain features of a wireless bridge fixed link channel to construct dual-domain feature pairs;

[0009] S2, generating positive and negative sample pairs, optimizing the feature extractor through cross-modal contrast learning, and mapping the environment domain features and the channel domain features to a unified semantic space;

[0010] S3, extracting channel decoupling features from the optimized dual-domain feature pairs by using a nonlinear decoupling network to separate the environment-dependent components and the environment-independent components;

[0011] S4, training an auto-encoder network based on the channel decoupling features of normal samples to construct an anomaly detection model;

[0012] S5, extracting channel decoupling features from real-time data, calculating anomaly scores through reconstruction error, and determining abnormal types by layering;

[0013] Further, in the step S1, the environment domain features of the wireless bridge fixed link channel are extracted by the following processing steps:

[0014] Input environment domain data , wherein represents an environment map, represents a metadata parameter tensor composed of metadata parameters of all wireless communication links under the environment map , the metadata parameters including weather type, Euclidean distance, line-of-sight state, path gain, two-dimensional base station coordinates and user coordinates, and L represents the total number of wireless communication links under the environment map;

[0015] The invalid areas of the environment map are removed by using an effective area mask:

[0016]

[0017]

[0018] , wherein represents the effective area, represents the mask value at the pixel in the effective area, represents the gray value of the environment map at the pixel , and represents pixel-by-pixel multiplication, represents the environment map after removing the invalid areas;

[0019] A multi-scale feature extractor based on the U-net architecture is constructed to extract a spatial feature vector:

[0020] Input tensor , wherein represents a one-hot encoding function;

[0021] The encoder layer adopts a 4-layer convolutional neural network to extract semantic features by stepwise downsampling:

[0022]

[0023] , wherein represents the The feature map of the layer output, Indicates the number of layers, Indicates the The number of layer channels, , express Convolution kernel, stride 2, output The convolution operation of the channel, ReLu represents the activation function, and BatchNorm represents the normalization operation;

[0024] The decoder uses a 3-layer transposed convolutional neural network to upsample step by step and concatenate it with the feature maps of the corresponding layers of the encoder to restore spatial details:

[0025]

[0026]

[0027]

[0028] in Represents the encoder layer 4 output, Indicates the number of layers, Indicates the The number of layer channels, , express Transposed convolution kernel, stride 2, output Channel transposed convolution operation, Represents the channel dimension splicing operation, Represents the output feature map after upsampling of the lth layer, BatchNorm represents, represents the feature map output by the lth layer, Represents the feature map of the final output of the decoder, express Convolution kernel, stride 1, output Channel convolution operation;

[0029] Generate spatial feature vectors through global average pooling and fully connected layers: ,

[0030] in represents global average pooling, FC represents fully connected layer;

[0031] Extract metadata parameter features:

[0032] in is the metadata parameter tensor, is the weight matrix, flatten is the flattening function;

[0033] The spatial features And metadata parameter features Concatenate the output final environment domain features:

[0034] ;

[0035] Further, the step S1 of extracting the channel domain features of the wireless channel comprises the following processing steps:

[0036] Input the complex channel matrix sequence set of all L links in the target wireless communication environment within the time window T , wherein is the number of receive antennas, is the number of transmit antennas, is the complex channel matrix of the ith link, C represents the complex domain, and T represents the number of time sampling points, represents any one time point within the time T;

[0037] Split the original complex channel matrix set sequence into a real part matrix set sequence and an imaginary part matrix set sequence , wherein and respectively represent the real part and the imaginary part of the channel matrix of the ith link at the tth time point;

[0038] Use a shared complex convolution kernel to perform a complex convolution operation on the real part matrix and the imaginary part matrix of each time slice of each link:

[0039]

[0040] , wherein respectively represent the real part and the imaginary part weight of the convolution kernel, k represents the size of the convolution kernel, represents two-dimensional convolution, represents the real part and the imaginary part feature map of the ith link after complex convolution at the tth time point;

[0041] Perform average pooling on the real part and the imaginary part feature map of the ith link at the tth time point:

[0042]

[0043]

[0044] Construct the real part time sequence feature and the imaginary part time sequence feature of the ith link:

[0045]

[0046]

[0047] Real part timing feature of full link and imaginary part timing feature :

[0048]

[0049]

[0050] Real part time domain feature matrix of full link is constructed by using BiLSTM network to capture time domain feature of each link and imaginary part feature matrix of full link :

[0051]

[0052]

[0053] Real part time domain feature matrix Maximum pooling is performed along the link dimension: ;

[0054] Imaginary part time domain feature matrix Average pooling is performed along the link dimension: ;

[0055] Real part feature vector and imaginary part feature vector are respectively subjected to nonlinear transformation through full connection layer:

[0056]

[0057]

[0058] Wherein L represents the total number of links, represents the maximum pooling operation, represents the average pooling operation, respectively represent the full link real part feature vector and imaginary part feature vector after nonlinear transformation, respectively represent the full connection layer weight of real part and imaginary part, respectively represent the bias vector of real part and imaginary part.

[0059] and are spliced to obtain 256-dimensional full link channel feature ;

[0060] Further, in the step S2, generating positive and negative sample pairs includes the following processing steps:

[0061] Selecting original sample , keeping the environment map​ and metadata parameter tensor invariant, extract adjacent set of complex channel matrix sequence sets of time intervals , get and original sample constitute a set of positive sample pairs , wherein P is the total number of sample pairs;

[0062] select samples environmental domain data and sample set of complex channel matrix sequence sets mismatched combination, constitute a negative sample set: , wherein environmental domain data of sample p, channel domain data of sample q;

[0063] Further, in the step S2, the feature extractor is optimized by cross-modal contrast learning to map the environmental domain features and the channel domain features to a unified semantic space includes the following processing steps:

[0064] Obtain positive and negative sample pairs, and extract environmental domain features and channel domain features of the sample pairs using the feature extractor;

[0065] Obtain multimodal interaction features through a multimodal attention mechanism, and obtain environmental domain features and channel domain features after the attention enhanced features are obtained by fusing the original features through a residual gating mechanism, respectively;

[0066] The fused environmental domain features and channel domain features are mapped to a unified feature space through 2-layer MLP:

[0067]

[0068]

[0069] wherein respectively represent the cross-domain semantic alignment feature vectors of the environmental domain and the channel domain, is a weight matrix, is a bias;

[0070] Calculate the cross-modal contrast loss: =

[0071] wherein N represents the number of batch samples, represents the cross-domain semantic alignment feature vector from the positive sample pair, represents the cross-domain semantic alignment feature vector from the negative sample pair, For the function of calculating the cosine similarity score, a and b represent two inputs, represents a temperature parameter, represents a cross-domain semantic alignment feature vector of the environment domain of the i-th sample, respectively represent cross-domain semantic alignment feature vectors of the channel domain of the i-th and k-th samples;

[0072] Calculate the feature space alignment loss:

[0073] wherein is a learnable alignment matrix, initialized as an identity matrix, is the environment domain feature of sample i;

[0074] Calculate the joint loss function: wherein represents a weight coefficient;

[0075] Iteratively perform forward propagation and back propagation to optimize the parameters of the feature extractor and the projection network until convergence;

[0076] Further, the multi-modal interaction features are obtained through the multi-modal attention mechanism and are fused with the original features respectively, including the following steps:

[0077] The environment domain feature and the channel domain feature are constructed into a joint feature matrix: ;

[0078] The joint feature matrix is mapped into a query matrix , a key matrix , and a value matrix , wherein represents a learnable projection matrix;

[0079] The modal interaction weight matrix is calculated using the scaled dot-product attention:

[0080]

[0081] wherein represents the self-attention strength in the environment domain, represents the attention strength of the environment domain query to the channel domain key, represents the self-attention strength in the channel domain, represents the attention strength of the channel domain query to the environment domain;

[0082] The outputs of each head attention are weighted and summed, spliced, and projected to the original dimension to obtain the multi-modal interaction features:

[0083]

[0084] wherein is an output projection matrix, represents a channel concatenation operation;

[0085] extracting environment domain attention enhanced features and channel domain attention enhanced features from the multi-modal interaction features

[0086]

[0087]

[0088] implementing conflict-aware fusion of the original features and the environment domain attention enhanced features and the channel domain attention enhanced features:

[0089]

[0090]

[0091]

[0092]

[0093] wherein represents an activation function, represents a shared gating weight matrix, represents an element-wise product, and respectively represent the environment domain features and the channel domain features after fusion of the attention enhanced features;

[0094] Further, in the step S3, extracting channel decoupled features from the optimized dual-domain feature pair using the nonlinear decoupling network comprises the following processing steps:

[0095] fixing the network parameters of the optimized feature extractor and the cross-modal mechanism, and extracting optimized features with cross-modal semantic alignment capability:

[0096]

[0097]

[0098] wherein and respectively represent the extracted optimized environment domain features and the channel domain features, represents an optimized environment domain feature extractor, represents an optimized channel domain feature extractor, represents input environment data, including an environment map and a metadata parameter tensor, ​represents an input complex channel matrix sequence set;

[0099] The environment domain feature after the fusion attention enhanced feature is obtained through the cross-modal attention mechanism And the channel domain feature ;

[0100] The environment-related component is extracted by using a nonlinear decoupling network composed of two layers of MLP: , wherein represents an environment-related component, respectively represent weight matrices, , respectively represent bias vectors;

[0101] The environment-independent component is calculated: ;

[0102] The decoupling loss function is calculated:

[0103]

[0104] , wherein represents the calculation of the Euclidean distance, N represents the batch size, represents the weight coefficient;

[0105] The optimization feature extractor is frozen, and the optimization nonlinear decoupling network parameters are updated by using the decoupling loss function;

[0106] Further, in the step S4, training the double-branch autoencoder network based on the channel decoupling features of normal samples includes the following steps:

[0107] Input the channel decoupling features of normal samples: , wherein and respectively represent the environment-related component and the environment-independent component;

[0108] A double-branch independent encoder is constructed, including an environment branch encoder And a device branch encoder ;

[0109] , wherein , respectively represent the weight matrices of the environment branch and the device branch, and respectively represent the bias vectors of the environment branch and the device branch;

[0110] Feature fusion is performed in the hidden layer:

[0111] A joint decoder is constructed, including reconstructing the environment-related component and the reconstruction environment independent component

[0112] wherein , and and respectively represent the weight matrix of the reconstruction environment dependent component and independent component,

[0113] The dual-channel reconstruction loss function is calculated as:

[0114]

[0115] wherein N represents the batch size;

[0116] The dual-branch independent encoder and joint decoder parameters are optimized by a back propagation algorithm;

[0117] Further, the step S5 of extracting channel decoupling features from real-time data and generating abnormal score by reconstructing channel decoupling features to determine the abnormal type includes the following steps:

[0118] Input real-time channel decoupling features: , wherein and represent the real-time environment dependent component and the environment independent component;

[0119] Reconstruct the real-time channel decoupling features to obtain the reconstructed real-time environment dependent component and the environment independent component

[0120] Calculate the dual-channel abnormal score:

[0121]

[0122]

[0123] wherein and respectively represent the abnormal score of the environment dependent component and the environment independent component;

[0124] Set the threshold value based on the reconstruction error distribution of the historical normal sample:

[0125]

[0126]

[0127] wherein and respectively represent the mean and standard deviation of and in the normal sample;

[0128] Determine the type of exception by layered logic rules:

[0129] If and , determine normal;

[0130] If and , determine mixed exception;

[0131] If only , determine environmental mutation;

[0132] If only , determine device failure.

[0133] The application also provides a wireless bridge fixed link channel feature decoupling and abnormal state detection system based on dual-domain contrast learning, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to realize the above method.

[0134] The beneficial effects of the application are:

[0135] The wireless bridge fixed link channel feature decoupling and abnormal state detection method and system based on dual-domain contrast learning provided by the application, by constructing a dual-feature pair of environment domain and channel domain, using a cross-modal contrast learning mechanism to force the dual-domain features of the same samples to converge closely in a unified space and significantly separate the different samples, the inter-class discrimination and intra-class cohesion of the feature space can be improved, and the problem of sensitivity to environmental interference of traditional single-modal features is solved.

[0136] Based on the optimized dual-domain feature pair, the application uses a nonlinear decoupling network to separate the environment-related components and environment-independent components, strips the interference of the environment on the channel features, highlights the inherent characteristics of the channel, and provides a more pure feature representation for abnormal detection.

[0137] The application trains an autoencoder based on the channel decoupled features of normal samples, constructs an abnormal score mechanism through reconstruction error, captures the normal state feature distribution under the condition of no label using an unsupervised learning framework, realizes the detection of the abnormal state of the wireless channel, and reduces the dependence on labeled data. At the same time, the environment-related and independent components of the channel decoupled features are extracted in real time, the double-channel abnormal scores are calculated and the statistical threshold is set, the four types of "normal", "mixed exception", "environmental mutation" and "device failure" are distinguished by combining the layered logic rules, and the accuracy and explainability of the abnormal state detection are improved.

[0138] Additional advantages, objects, and features of the application will be apparent to those skilled in the art upon examination of the following specification. It is intended to be covered by the following claims. BRIEF DESCRIPTION OF DRAWINGS

[0139] In order to make the purposes, technical solutions and beneficial effects of the present application clearer, the present application provides the following drawings for illustration.

[0140] Figure 1 Simplified flowchart of the method.

[0141] Figure 2 Model framework diagram of the method.

[0142] Figure 3 Channel domain feature extractor of the method.

[0143] Figure 4 Environment domain feature extractor of the method. DETAILED DESCRIPTION

[0144] The present application will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present application and implement it. However, the embodiments are not intended to limit the present application.

[0145] REFERENCE Figure 1 and Figure 2 , Figure 1 Simplified flowchart of the method, Figure 2 Model framework diagram of the method, the embodiment discloses a wireless bridge fixed link channel feature decoupling and abnormal state detection method based on dual-domain contrast learning, comprising the following steps:

[0146] S1, using a feature extractor to extract environment domain features and channel domain features of the wireless bridge fixed link channel to construct a dual-domain feature pair;

[0147] It should be noted that the feature extractor includes an environment domain feature extractor, a channel domain feature extractor, and a multi-modal attention mechanism;

[0148] It should be noted that in S1, the extraction of the environment domain features of the wireless bridge fixed link channel comprises the following processing steps:

[0149] S100, input environment domain data , wherein represents an environment map, represents a metadata parameter tensor composed of metadata parameters of all wireless communication links under the environment map The metadata parameters include weather type, Euclidean distance, line-of-sight state, path gain, two-dimensional base station coordinates and user coordinates, L represents the total number of wireless communication links under the environment map;

[0150] S101, using the valid area mask to remove the invalid area of the environment map:

[0151]

[0152]

[0153] wherein, represents the valid area, represents the mask value at the pixel of the valid area, represents the gray value of the environment map at the pixel , represents pixel-by-pixel multiplication, represents the environment map after removing the invalid area;

[0154] S102, constructing a multiscale feature extractor based on the U-net architecture to extract a spatial feature vector:

[0155] The input tensor , wherein represents a one-hot encoding function;

[0156] The encoder layer uses a 4-layer convolutional neural network to extract semantic features by progressively downsampling:

[0157]

[0158] wherein, represents the feature map output by the i-th layer, represents the number of layers, represents the number of channels of the i-th layer, , , , represents convolution operation with a convolution kernel, a step of 2, and output channels, ReLu represents an activation function, and BatchNorm represents a normalization operation; The decoder uses a 3-layer transposed convolutional neural network to progressively upsample and splice the feature maps corresponding to the encoder layers to restore spatial details:

[0159]

[0160]

[0161]

[0162] ​

[0163] wherein, represents the 4th layer output of the encoder, represents the number of layers, represents the number of channels of the 4th layer, , represents transpose convolution kernel, step 2, output channel transpose convolution operation, represents the channel dimension splicing operation, represents the output feature map after upsampling of the lth layer, BatchNorm represents, represents the feature map output by the lth layer, represents the feature map output by the lth layer, represents convolution kernel, step 1, output channel convolution operation;

[0164] S103, generate a spatial feature vector through global average pooling and a fully connected layer: ,

[0165] wherein represents global average pooling, and FC represents a fully connected layer;

[0166] S104, extract metadata parameter features:

[0167] wherein is a metadata parameter tensor, is a weight matrix, and flatten is a flattening function;

[0168] S105, concatenate the spatial feature and the metadata parameter feature to output the final environmental domain feature:

[0169]

[0170] It should be noted that in S1, the channel domain feature of the wireless channel includes the following processing steps:

[0171] S106, input a complex channel matrix sequence set of all L links in a target wireless communication environment within a time window T wherein is the number of receive antennas, is the number of transmit antennas, is the complex channel matrix of the ith link, C represents the complex domain, and T represents the number of time sampling points, represents any one time point within the time T.​

[0172] S107, split the original complex channel matrix set sequence into real part matrix sequence set and imaginary part matrix sequence set , where and respectively represent the real part and imaginary part of the channel matrix of the ith link at the tth time point;

[0173] S108, perform complex convolution operation on the real part matrix and imaginary part matrix of each time slice of each link using the shared complex convolution kernel :

[0174]

[0175] where, respectively represent the real part and imaginary part weights of the convolution kernel, k represents the size of the convolution kernel, represents two-dimensional convolution, represents the real part and imaginary part feature maps of the ith link after complex convolution at time point t;

[0176] S109, perform average pooling on the real part and imaginary part feature maps of the ith link at time point t:

[0177]

[0178]

[0179] S110, construct the real part time sequence feature and imaginary part time sequence feature of the ith link:

[0180]

[0181]

[0182] S111, construct the real part time sequence feature and imaginary part time sequence feature of all links:

[0183]

[0184]

[0185] S112, construct the all-link real part time domain feature matrix and all-link imaginary part feature matrix using the BiLSTM network to capture the time domain features of each link:

[0186]

[0187]

[0188] S113, the real time domain characteristic matrix Perform max pooling along the link dimension: ;

[0189] S114, for the imaginary time domain characteristic matrix Perform average pooling along the link dimension: ;

[0190] S115, the real eigenvector and the imaginary eigenvector Nonlinear transformation is performed through the fully connected layer respectively:

[0191]

[0192]

[0193] Where L represents the total number of links, represents the maximum pooling operation, represents the average pooling operation, represent the real eigenvector and imaginary eigenvector of the full link after nonlinear transformation, denote the fully connected layer weights of the real and imaginary parts respectively, Bias vectors representing the real and imaginary parts respectively;

[0194] S116, will and Splicing to obtain 256-dimensional full-link channel features ;

[0195] S2, generates positive and negative sample pairs, optimizes the feature extractor through cross-modal contrastive learning, and maps the environment domain features and channel domain features into a unified semantic space; Figure 3 and Figure 4 As shown, Figure 3 is the channel domain feature extractor of this method, Figure 4 It is the environmental domain feature extractor of this method.

[0196] It should be noted that in S2, generating positive and negative sample pairs includes the following processing steps:

[0197] S200, select the original sample , maintain the environment map and metadata parameter tensors Unchanged, extract adjacent The set of complex channel matrix sequences of time intervals , obtained With the original sample constitute a positive sample pair set , wherein P is the total number of sample pairs;

[0198] S201, select samples environmental domain data and sample complex channel matrix sequence set mismatch combination, constitute a negative sample set: , wherein represents the environmental domain data of sample p, represents the channel domain data of sample q;

[0199] It should be noted that in S2, the environmental domain features and the channel domain features are mapped to a unified semantic space by optimizing the feature extractor through cross-modal contrastive learning, including the following processing steps:

[0200] S202, obtain positive and negative sample pairs, and extract environmental domain features and channel domain features of the sample pairs using the feature extractor;

[0201] S203, obtain multi-modal interaction features through a multi-modal attention mechanism, and respectively fuse the original features through a residual gating mechanism to obtain attention-enhanced environmental domain features and channel domain features;

[0202] S204, map the fused environmental domain features and channel domain features to a unified feature space through 2-layer MLP:

[0203]

[0204]

[0205] wherein respectively represent the cross-domain semantic alignment feature vectors of the environmental domain and the channel domain, is a weight matrix, is a bias;

[0206] S205, calculate the cross-modal contrastive loss:

[0207] =

[0208] wherein N represents the number of batch samples, represents the cross-domain semantic alignment feature vector from the positive sample pair, represents the cross-domain semantic alignment feature vector from the negative sample pair, is a function for calculating the cosine similarity score, a and b represent two inputs, represents a temperature parameter, a cross-domain semantic aligned feature vector representing the environment domain of the i-th sample, a cross-domain semantic aligned feature vector representing the channel domain of the i-th, k-th sample, respectively;

[0209] S206, calculating a feature space alignment loss:

[0210] wherein is a learnable alignment matrix, initialized as an identity matrix, an environment domain feature of sample i;

[0211] S207, calculating a joint loss function: wherein represents a weight coefficient;

[0212] S208, iteratively performing forward propagation and back propagation to optimize the parameters of the feature extractor and the projection network until convergence;

[0213] It should be noted that in S203, obtaining the multi-modal interaction feature through the multi-modal attention mechanism and fusing it with the original feature respectively includes the following steps:

[0214] S209, constructing the environment domain feature and the channel domain feature into a joint feature matrix: ;

[0215] S210, mapping the joint feature matrix into a query matrix , a key matrix , and a value matrix , wherein represents a learnable projection matrix;

[0216] S211, calculating a modal interaction weight matrix using scaled dot-product attention:

[0217]

[0218] wherein represents an intra-environment domain attention strength, represents an environment domain query attention strength on a channel domain key, represents an intra-channel domain attention strength, represents a channel domain query attention strength on an environment domain;

[0219] S212, concatenating and projecting the weighted sum of the outputs of each head attention to the original dimension to obtain a multi-modal interaction feature:

[0220]

[0221] wherein is an output projection matrix, represents a channel concatenation operation;

[0222] S213, extracting environment domain attention enhanced features and channel domain attention enhanced features from the multi-modal interaction features respectively:

[0223]

[0224]

[0225] S214, realizing conflict perception fusion of the original features and the environment domain attention enhanced features and the channel domain attention enhanced features:

[0226]

[0227]

[0228]

[0229]

[0230] wherein represents an activation function, represents a shared gating weight matrix, represents an element-level product, and respectively represent the environment domain features and the channel domain features after fusion of the attention enhanced features;

[0231] S3, extracting channel decoupling features from the optimized dual-domain feature pair by using a nonlinear decoupling network to separate environment-related components and environment-independent components;

[0232] It should be noted that in S3, extracting channel decoupling features from the optimized dual-domain feature pair by using a nonlinear decoupling network includes the following processing steps:

[0233] S300, fixing the network parameters of the optimized feature extractor and the cross-modal mechanism, and extracting optimized features with cross-modal semantic alignment capability:

[0234]

[0235]

[0236] wherein and respectively represent the extracted optimized environment domain features and channel domain features, ​denotes the optimized environment domain feature extractor, denotes the optimized channel domain feature extractor, denotes the input environment data, including environment map and metadata parameter tensor, denotes the input complex channel matrix sequence set;

[0237] S301, obtaining the environment domain feature after fusion attention enhanced feature through the cross-modal attention mechanism and channel domain feature ;

[0238] S302, extracting environment related components by using a nonlinear decoupling network composed of two layers of MLP: wherein denotes the environment related component, denotes the weight matrix, , denotes the bias vector, respectively;

[0239] S303, calculating the environment independent component: ;

[0240] S304, calculating the decoupling loss function:

[0241]

[0242] wherein denotes the calculation of Euclidean distance, N denotes the batch size, denotes the weight coefficient;

[0243] S305, freezing the optimized feature extractor, and updating the optimized nonlinear decoupling network parameters by using the decoupling loss function;

[0244] S4, training the auto-encoder network based on the channel decoupling features of normal samples to construct an anomaly detection model;

[0245] It should be noted that in S4, training the double-branch auto-encoder network based on the channel decoupling features of normal samples includes the following steps:

[0246] S401, inputting the channel decoupling features of normal samples: , wherein and denote the environment related component and the environment independent component, respectively;

[0247] S402, constructing a double-branch independent encoder, including an environment branch encoder and a device branch encoder ;

[0248] wherein, Wenv and Wdev represent the weight matrix of the environment branch and the device branch, respectively, benv and bdev represent the bias vector of the environment branch and the device branch, respectively;

[0249] S403, feature fusion at the hidden layer:

[0250] S404, constructing a joint decoder, including reconstructing environment-related components and environment-independent components

[0251] wherein, Wenv and Wdev represent the weight matrix of the environment branch and the device branch, respectively, benv and bdev represent the bias vector of the environment branch and the device branch, respectively;

[0252] S405, calculating a double-channel reconstruction loss function:

[0253]

[0254] wherein, N represents the batch size;

[0255] S406, optimizing the parameters of the double-branch independent encoder and the joint decoder through a back propagation algorithm;

[0256] S5, extracting channel decoupling features from real-time data, calculating an abnormal score through reconstruction error, and determining the abnormal type by layering;

[0257] It should be noted that in S5, the channel decoupling features are extracted from real-time data, and the abnormal type is determined by layering the abnormal score generated by reconstructing the channel decoupling features, including the following steps:

[0258] S500, inputting real-time channel decoupling features: wherein, represent real-time environment-related components and environment-independent components;

[0259] S501, reconstructing real-time channel decoupling features to obtain reconstructed real-time environment-related components and environment-independent components

[0260] S502, calculating a double-channel abnormal score:

[0261]

[0262] ​​​​​

[0263] wherein, and respectively represent the anomaly scores of the environment-related component and the environment-independent component;

[0264] S503, setting the threshold value based on the reconstruction error distribution of the historical normal samples:

[0265]

[0266]

[0267] wherein, and respectively represent the mean and standard deviation of and in the normal samples;

[0268] S504, determining the anomaly type by the hierarchical logic rule:

[0269] if and , it is determined as normal;

[0270] if and , it is determined as mixed anomaly;

[0271] if only , it is determined as environment mutation;

[0272] if only , it is determined as device failure.

[0273] The above-described embodiments are only the preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. The equivalent substitutions or transformations made by the person skilled in the art on the basis of the present application are within the protection scope of the present application. The protection scope of the present application is subject to the claims.

Claims

1. A wireless bridge fixed link channel feature decoupling and abnormal state detection method based on dual-domain comparative learning, characterized by: The following steps are involved: S1, using the feature extractor to extract the environment domain features and channel domain features of the wireless bridge fixed link channel respectively to construct a dual-domain feature pair; S2, generates positive and negative sample pairs, optimizes the feature extractor through cross-modal contrastive learning, and maps the environment domain features and channel domain features into a unified semantic space; S3, using a nonlinear decoupling network to extract channel decoupling features from the optimized dual-domain feature pairs to separate the environment-related components from the environment-irrelevant components; S4, trains the autoencoder network based on the channel decoupling features of normal samples to build an anomaly detection model; S5, extract channel decoupling features from real-time data, calculate anomaly scores based on reconstruction errors, and determine anomaly types in layers; In step S2, the feature extractor is optimized through cross-modal contrastive learning to map the environment domain features and the channel domain features into a unified semantic space, which includes the following processing steps: Obtain positive and negative sample pairs, and use a feature extractor to extract environmental domain features and channel domain features of the sample pairs; The multimodal interaction features are obtained through the multimodal attention mechanism and fused with the original features through the residual gating mechanism to obtain the environmental domain features and channel domain features after the attention enhancement features; The fused environmental domain features and channel domain The features are mapped to a unified feature space through a 2-layer MLP: ， ， in The feature vectors representing the cross-domain semantic alignment of the environment domain and the channel domain, is the weight matrix, is bias; Calculate cross-modal contrast loss: = Where N represents the number of batch samples, represents the feature vector of cross-domain semantic alignment from positive sample pairs, Represents the feature vector of cross-domain semantic alignment from negative sample pairs, To calculate the cosine similarity score function, a, b represent two inputs, represents the temperature parameter, The feature vector representing the cross-domain semantic alignment of the environment domain of the i-th sample, Represent the feature vectors of cross-domain semantic alignment of the channel domain of the i-th and k-th samples respectively; Calculate the feature space alignment loss: in is a learnable alignment matrix, initialized to the identity matrix, Environmental domain characteristics of sample i; Calculate the joint loss function: ,in represents the weight coefficient; Forward and backpropagation are iteratively performed to optimize the parameters of the feature extractor and projection network until convergence.

2. The method for decoupling channel characteristics and detecting abnormal states of wireless bridge fixed links based on dual-domain comparative learning according to claim 1 is characterized by: In step S1, extracting the environmental domain features of the fixed link channel of the wireless bridge includes the following processing steps: Entering environmental domain data ,in Indicates the environment map, Metadata parameter tensor representing the metadata parameters of all wireless communication links under the environment map ,The metadata parameters include weather type, Euclidean distance, line-of-sight status, path gain, 2D base station coordinates and user coordinates.,L,represents the total number of wireless communication links under the,environment map; Use the valid area mask to cull invalid areas of the environment map: in, Indicates the effective area, Indicates pixels in the effective area The mask value at Represents the environment map in pixels The gray value at represents pixel-by-pixel multiplication, Represents the environment map after removing invalid areas; Construct a multi-scale feature extractor based on the U-net architecture to extract spatial feature vectors: Input Tensor ,in represents the one-hot encoding function; The encoder layer uses a 4-layer convolutional neural network to extract semantic features by downsampling step by step: in Indicates the The feature map of the layer output, Indicates the number of layers, Indicates the The number of layer channels, , express Convolution kernel, stride 2, output The convolution operation of the channel, ReLu represents the activation function, and BatchNorm represents the normalization operation; The decoder uses a 3-layer transposed convolutional neural network to upsample step by step and concatenate it with the feature maps of the corresponding layers of the encoder to restore spatial details: in, Represents the 4th layer output of the encoder, Indicates the number of layers, Indicates the The number of layer channels, , express Transposed convolution kernel, stride 2, output Channel transposed convolution operation, represents the channel dimension splicing operation, Represents the output feature map after upsampling of the lth layer, BatchNorm represents, represents the feature map output by the lth layer, Represents the feature map of the final output of the decoder, express Convolution kernel, stride 1, output Channel convolution operation; Generate spatial feature vectors through global average pooling and fully connected layers: , in represents global average pooling, FC represents fully connected layer; Extract metadata parameter features: ; The spatial features and metadata parameter characteristics The final environment domain features are stitched and output: ； in, is the metadata parameter tensor, is the weight matrix, and flatten is the flattening function.

3. The method for decoupling channel characteristics and detecting abnormal states of wireless bridge fixed links based on dual-domain comparative learning according to claim 1 is characterized by: In step S1, extracting the channel domain features of the wireless channel includes the following processing steps: Input the complex channel matrix sequence set of all L links in the target wireless communication environment within the time window T ,in is the number of receiving antennas, is the number of transmitting antennas, is the complex channel matrix of the i-th link, C represents the complex domain, T represents the number of time sampling points, Represents any time point within the time range T; Split the original complex channel matrix set sequence into a real part matrix sequence set and the imaginary part matrix sequence set ,in and denote the real and imaginary parts of the channel matrix of the ith link at the tth time point; Using shared complex convolution kernels , perform complex convolution operations on the real and imaginary matrices of each link and each time slice: in Represent the real and imaginary weights of the convolution kernel respectively, k represents the convolution kernel size, represents a two-dimensional convolution, Represents the real and imaginary feature maps of the i-th link after complex convolution at time point t; Perform average pooling on the real and imaginary feature maps of the i-th link at time point t: Construct the real part timing characteristics of the i-th link and imaginary part timing characteristics : Constructing the real part timing characteristics of the entire link and imaginary part timing characteristics : The BiLSTM network is used to capture the time domain features of each link and construct the real time domain feature matrix of the entire link. And the full link imaginary characteristic matrix : The real time domain characteristic matrix Perform max pooling along the link dimension: ; For the imaginary time domain characteristic matrix Perform average pooling along the link dimension: ; The real eigenvector and the imaginary eigenvector Nonlinear transformation is performed through the fully connected layer respectively: Will and Splicing to obtain 256-dimensional full-link channel features ; Where L represents the total number of links, represents the maximum pooling operation, represents the average pooling operation, represent the real eigenvector and imaginary eigenvector of the full link after nonlinear transformation, denote the fully connected layer weights of the real and imaginary parts respectively, Bias vectors representing the real and imaginary parts respectively.

4. The method for decoupling channel characteristics and detecting abnormal states of wireless bridge fixed links based on dual-domain contrastive learning according to claim 1 is characterized by: In step S2, generating positive and negative sample pairs includes the following processing steps: Select the original sample , maintain the environment map and metadata parameter tensors Unchanged, extract adjacent The set of complex channel matrix sequences of time intervals , obtained With the original sample Construct a positive sample pair set ; Select samples Environmental domain data With sample The complex channel matrix sequence set Mismatch combinations constitute the negative sample set: ,in represents the environmental domain data of sample p, Represents the channel domain data of sample q.

5. The method for decoupling channel characteristics and detecting abnormal states of wireless bridge fixed links based on dual-domain contrastive learning according to claim 1 is characterized by: Obtaining multimodal interaction features through the multimodal attention mechanism and fusing them with the original features respectively includes the following steps: Environmental Domain Features and channel domain characteristics Constructed as a joint feature matrix: ; The joint feature matrix Mapping to query matrix , bond matrix , value matrix ,in represents the learnable projection matrix; Calculate the modal interaction weight matrix using scaled dot product attention: in represents the strength of self-attention in the environment domain, represents the attention intensity of the environment domain query on the channel domain key, represents the self-attention strength in the channel domain, Indicates the attention intensity of the channel domain query on the environment domain; The weighted sum of the attention outputs of each head is concatenated and projected into the original dimension to obtain the multimodal interaction features: in is the output projection matrix, Represents a channel splicing operation; Extracting environmental domain attention enhancement features from multimodal interaction features And channel domain attention enhancement features: Realize the conflict-aware fusion of original features, environmental domain attention-enhanced features, and channel domain attention-enhanced features: in represents the activation function, represents the shared gating weight matrix, represents element-wise product, and They represent the environment domain features and channel domain features after fusing the attention enhanced features.

6. The method for decoupling channel characteristics and detecting abnormal states of wireless bridge fixed links based on dual-domain contrastive learning according to claim 1 is characterized by: In step S3, extracting channel decoupling features from the optimized dual-domain feature pairs using a nonlinear decoupling network includes the following processing steps: Fix the optimized feature extractor and network parameters of the cross-modal mechanism to extract optimized features with cross-modal semantic alignment capabilities: in and Respectively represent the extracted optimized environment domain features and channel domain features, represents the optimized environment domain feature extractor, represents the optimized channel domain feature extractor, Represents the input environment data, including the environment map and metadata parameter tensor, represents the set of complex channel matrix sequences of the input; The environmental domain features after integrating attention-enhanced features are obtained through the cross-modal attention mechanism And channel domain characteristics ; A nonlinear decoupling network consisting of two layers of MLP is used to extract environmental related components: ,in represents the environment-related component, Represent the weight matrix, 、 They represent the bias vectors respectively; Calculate the environment-independent component: ; Calculate the decoupling loss function: in Indicates the calculation of Euclidean distance, N indicates the batch size, represents the weight coefficient; The optimized feature extractor is frozen and the decoupling loss function is used to update and optimize the nonlinear decoupling network parameters.

7. The method for wireless bridge fixed link channel feature decoupling and abnormal state detection based on dual-domain contrast learning according to claim 1 is characterized by: In step S4, training a two-branch autoencoder network based on the channel decoupling features of normal samples includes the following steps: Channel decoupling characteristics of input normal samples: ,in and represent the environment-related component and the environment-irrelevant component respectively; Build a dual-branch independent encoder, including an environment branch encoder and device branch encoders ; in , Represent the weight matrices of the environment branch and the device branch respectively, and represent the bias vectors of the environment branch and the device branch respectively; Perform feature fusion in the hidden layer: Constructing a joint decoder, including reconstructing the environment-related components Components that have nothing to do with the reconstruction environment in , Represent the weight matrices of the reconstructed environment-related components and irrelevant components, and Represent the bias vectors of the reconstructed environment-related and irrelevant components respectively; Calculate the dual-channel reconstruction loss function: Where N represents the batch size; The parameters of the dual-branch independent encoder and joint decoder are optimized through the back-propagation algorithm.

8. The method for wireless bridge fixed link channel feature decoupling and abnormal state detection based on dual-domain contrast learning according to claim 1 is characterized by: In step S5, extracting channel decoupling features from real-time data and generating anomaly scores by reconstructing the channel decoupling features to determine the anomaly type in a hierarchical manner include the following steps: Input real-time channel decoupling features: ,in and Represents real-time environment-related components and environment-independent components; Reconstruct the real-time channel decoupling features to obtain the reconstructed real-time environment related components Environmentally independent components Calculate the two-channel anomaly score: in and represent the abnormality scores of the environment-related component and the environment-irrelevant component, respectively; Set the threshold based on the reconstruction error distribution of historical normal samples: in and Represents normal samples and The mean and standard deviation of Determine the exception type through hierarchical logic rules: like and , determined to be normal; like and , determined to be mixed abnormality; If only , determined to be an environmental mutation; If only , it is determined to be a device failure.

9. A wireless bridge fixed link channel feature decoupling and abnormal state detection system based on dual-domain comparative learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

Images