Wireless communication link fault identification and positioning method based on multi-domain joint representation and mixed attention

Abstract

The invention discloses a wireless communication link fault identification and positioning method based on multi-domain joint representation and mixed attention, and relates to the field of wireless communication link fault identification and positioning, and the method comprises the steps: determining a wireless communication network topology structure as each link number; collecting various quality evaluation index time sequence data of the wireless communication network links under the conditions of link faults and normal work; extracting time domain, spatial domain and frequency domain features of the whole network based on a mixed attention mechanism to construct feature vectors of multi-domain joint representation; extracting feature vectors of multi-domain joint representation under various evaluation indexes, and obtaining weighted feature vectors by using a global attention mechanism; and outputting the predicted fault link category by the weighted feature vector through a full connection layer so as to realize positioning. According to the method, a mixed attention mechanism is introduced to extract multi-domain features to construct a joint representation vector, multiple quality evaluation index time series data are collected, weighting is performed by using a global attention mechanism, the classification precision is improved, and a fault link is positioned more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication link fault identification and positioning, and in particular to a wireless communication link fault identification and positioning method based on multi-domain joint representation and hybrid attention. Background Art

[0002] With the rapid development of wireless communication technology, the scale and complexity of communication networks are increasing. Timely identification and location of link faults are crucial to ensuring network stability and service quality. Traditional methods often rely on a single evaluation indicator, which has limitations and is difficult to fully and accurately reflect the link status. At the same time, there is a lack of comprehensive consideration of time, space, and frequency domains, insufficient feature extraction, and difficulty in adaptively adjusting feature weights, resulting in low accuracy in fault identification and location.

[0003] Therefore, a method is needed to accurately identify and locate faulty links. Summary of the invention

[0004] In view of this, the object of the present invention is to provide a wireless communication link fault identification and location method based on multi-domain joint representation and hybrid attention, which can locate the faulty link more accurately.

[0005] In order to achieve the above object, the present invention provides the following technical solutions: The wireless communication link fault identification and location method based on multi-domain joint representation and hybrid attention provided by the present invention comprises the following steps: S1, determining the wireless communication network topology structure as each link number; S2, collecting time series data of various quality evaluation indicators of wireless communication network links under each link failure and normal operation; S3, based on the hybrid attention mechanism, extracts the time domain, space domain and frequency domain features of the overall network to construct a feature vector of multi-domain joint representation; S4, extract the feature vectors of multi-domain joint representation under multiple evaluation indicators respectively and use the global attention mechanism to obtain the weighted feature vector; S5, the weighted feature vector is output through the fully connected layer to predict the fault link category for location.

[0006] Further, in step S1, numbering each link includes numbering each edge of the wireless communication link network, and the network topology structure of the wireless communication link includes N edges, then a total of N+1 wireless communication link failure categories are included, and an additional one is a normal working category.

[0007] Further, in step S2, the wireless communication link multiple evaluation index time series data includes M evaluation indexes, and the multiple evaluation index time series data are constructed: Q={} ; in, It is the time series data matrix of the overall wireless communication network link under one evaluation index: = ; Wherein, N is the link dimension of the network topology structure of the wireless communication link, T is the time dimension of the time series data, and any one for Evaluation index: Time series data of link a at time point b.

[0008] Further, in step S3, the time domain, space domain and frequency domain features of the overall network are extracted based on the hybrid attention mechanism to construct a feature vector of multi-domain joint representation, which specifically includes the following steps: The Each row is input into the Transformer encoder to extract the time characteristics of each link timing data under an evaluation index to obtain the time feature vector ; Extraction using FCN fully convolutional network and hybrid attention mechanism The spatial eigenvector of ; Extraction using STFT The frequency domain eigenvector of ; The time feature vector , spatial eigenvector and frequency domain eigenvector Concatenate the feature vectors that form a joint representation of multiple domains , where the splicing process is .

[0009] Furthermore, the FCN fully convolutional network and hybrid attention mechanism apply the CBAM hybrid attention mechanism after the output of the first two layers of one-dimensional temporal convolution blocks of the FCN fully convolutional network to enhance the relational features between different links and the spatial features of key features at different time steps.

[0010] Further, the STFT extraction The frequency domain characteristics of the , specifically including the following steps: Will Slice the time dimension into K subsequences to get the slice matrix: ; Among them, any subsequence ,L is the length of each subsequence; For each sub-matrix in the slice matrix Calculate the STFT matrix: ; Among them, any one , the formula for calculating STFT is ; in, subsequence The sequence obtained after STFT calculation is Indicates frequency, is an imaginary unit, is a window function with a length of L; The complex form of the STFT matrix Convert into amplitude matrix and phase matrix respectively; Use the Transformer encoder to extract the amplitude feature vectors of the amplitude matrix respectively and the phase eigenvector of the phase matrix ; The amplitude eigenvector and the phase eigenvector are concatenated to form the frequency domain eigenvector. The concatenation process is: .

[0011] Further, in S4, obtaining a weighted feature vector using a global attention mechanism includes the following steps: The feature vectors jointly representing multiple domains under the various evaluation indicators are: As input, calculate the global attention weight W = ; Among them, each Use the global attention weight to perform weighted summation on the feature vector to obtain the weighted feature vector .

[0012] Furthermore, in step S5, the weighted feature vector is output through the fully connected layer to predict the fault link category to achieve positioning, which is specifically performed in the following manner: The fully connected layer is ,in, For the predicted results, is the weight of the fully connected layer, is the bias of the fully connected layer; The model training uses a cross entropy loss function to measure the difference between the predicted result and the true category label. The cross entropy loss function is: ; in, represents the feature vector after weighted summation of feature vectors using global attention weights; N represents the number of samples; K represents the number of categories; Indicates the probability of the model predicting the i-th sample; Represents the one-hot encoding of the true label of the i-th sample.

[0013] The beneficial effects of the present invention are: The present invention provides a method for identifying and locating wireless communication link faults based on multi-domain joint representation and hybrid attention. The method relates to the field of wireless communication link fault identification and positioning, including determining the wireless communication network topology structure as each link number; collecting time series data of various quality evaluation indicators of wireless communication network links under each link failure and normal operation; extracting the time domain, space domain and frequency domain features of the overall network based on the hybrid attention mechanism to construct a feature vector of multi-domain joint representation; extracting the feature vectors of multi-domain joint representation under various evaluation indicators respectively and obtaining weighted feature vectors using the global attention mechanism; and locating the weighted feature vector by outputting the predicted fault link category through the fully connected layer. The present invention introduces a hybrid attention mechanism to extract multi-domain features to construct a joint representation vector, collects time series data of various quality evaluation indicators, and uses the global attention mechanism for weighting to improve the classification accuracy and realize more accurate positioning of the faulty link.

[0014] This method extracts the time domain features of the link timing data of the wireless communication network, the space domain features of the timing data between different links, and the frequency domain features of the non-stationary timing data converted to the frequency domain, and constructs a multi-domain joint representation feature vector that can comprehensively and comprehensively reflect the status of the wireless communication network link. Compared with the traditional single-domain analysis method, multi-domain can more comprehensively capture the operating status of the wireless communication network link and improve the accuracy of link fault identification.

[0015] The present invention introduces a hybrid attention mechanism when extracting spatial domain features. The model can better capture the relationship features between non-adjacent links and the spatial features of key features at different time steps, further improving the accuracy of fault classification.

[0016] The present invention collects time series data of various quality evaluation indicators under each link failure and normal operation, providing a rich source of information for fault location. Different evaluation indicators reflect the status of the link from different angles, and the comprehensive use of these indicators can more accurately determine the location of the faulty link.

[0017] The present invention uses a global attention mechanism to weight the multi-domain joint characterization feature vectors under multiple evaluation indicators, and can adaptively adjust the weights of different indicators and features, highlighting the evaluation indicators that are more valuable for positioning according to actual conditions. This makes fault location more accurate, helps to quickly locate and repair faulty links, and reduces the impact of faults on wireless communication networks.

[0018] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration: Figure 1 This is a simplified schematic diagram of the process.

[0020] Figure 2 This is the model framework diagram of this method.

[0021] Figure 3 Comparison chart of recognition accuracy curve during training Figure 4 It is the confusion matrix under time domain features.

[0022] Figure 5 Confusion matrix for joint representation of multiple domains.

[0023] Figure 6 Confusion matrix for multi-domain joint representation and mixed attention. DETAILED DESCRIPTION

[0024] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0025] refer to Figure 1 and Figure 2 , Figure 1 It is a simplified schematic diagram of the process of the present invention, Figure 2 The model framework diagram of the present invention is as follows. The embodiment of the present invention discloses a method for identifying and locating wireless communication link faults based on multi-domain joint representation and hybrid attention, comprising the following steps: S1, determining the wireless communication network topology structure as each link number; It should be noted that the link numbering in S1 includes the numbering of each edge of the wireless communication link network, and the network topology structure of the wireless communication link includes N edges, which includes a total of N+1 wireless communication link failure categories, including N failure categories and one normal working category; In this embodiment, only the fault type of link disconnection is considered, and the category output by the model is the number of the disconnected link for positioning. If other fault types are considered, such as link reconnection faults, it is necessary to collect data of the corresponding fault type and increase the classification category.

[0026] S2, collecting time series data of various quality evaluation indicators of wireless communication network links under each link failure and normal operation; It should be noted that the wireless communication link multiple evaluation index time series data in S2 includes M evaluation indexes, and multiple evaluation index time series data are constructed: Q={ } ; in, It is the time series data matrix of the overall wireless communication network link under one evaluation index: = ; Wherein, N is the link dimension of the network topology structure of the wireless communication link, T is the time dimension of the time series data, and any one for The evaluation index is the time series data of link a at time point b; Q represents the timing data of the overall wireless communication network link under various evaluation indicators; represents the time series data of the overall wireless communication network link under the Mth evaluation index; Indicates the type of evaluation indicator; represents a real vector space; represents the time series data of the overall wireless communication network link under the i-th evaluation index; Expressed as The time series data of link N at time point 1 under the evaluation index; It should be noted that in this embodiment, network traffic data is selected as the basis for determining wireless network communication link failures, and the three evaluation indicators of network traffic transmission rate, delay, and bit error rate are mainly adopted, that is, M=3. 7890 samples of a network of 25 nodes with 29 edges are collected under normal working conditions and different edge link disconnection scenarios, that is, N=29, and the classification categories are N+1=30 categories, and the training set and test set are divided into 8:2. Each sample contains time series data points under the three evaluation indicators of transmission rate, delay, and bit error rate of the captured network traffic, and each indicator contains 28,000 data time series points, that is, a total of about 2 million data points are used in the training process, and 500,000 data points are used in the test.

[0027] S3, based on the hybrid attention mechanism, extracts the time domain, space domain and frequency domain features of the overall network to construct a feature vector of multi-domain joint representation; It should be noted that in S3, the time domain, space domain and frequency domain features of the overall network are extracted based on the hybrid attention mechanism to construct a feature vector of multi-domain joint representation, including the following steps: S301, the Each row is input into the Transformer encoder to extract the time characteristics of each link timing data under an evaluation index to obtain the time feature vector ; S302, extracting using FCN fully convolutional network and hybrid attention mechanism The spatial eigenvector of ; S303, extracted using STFT The frequency domain eigenvector of ; S304: Time feature vector , spatial eigenvector and frequency domain eigenvector Concatenate the feature vectors that form a joint representation of multiple domains , where the splicing process is .

[0028] in, represents the time series data of the overall wireless communication network link under the i-th evaluation index; express The time feature vector of express The spatial eigenvector of express The frequency domain eigenvector of express The feature vector of the multi-domain joint representation; It should be noted that the FCN fully convolutional network and hybrid attention mechanism are obtained by applying the CBAM hybrid attention mechanism after the output of the first two layers of one-dimensional temporal convolution blocks of the FCN fully convolutional network to enhance the relational characteristics of time series data between different links under the same evaluation index and the spatial characteristics of key features at different time steps.

[0029] The time domain features capture the dynamic changes in link performance, the spatial domain features reveal the relationship between links and the impact of network topology on fault propagation, and the frequency domain features identify potential periodic fault modes by analyzing the frequency components of time series data. This cross-domain fusion not only enriches feature information, but also improves the comprehensiveness and robustness of the model, thereby improving the accuracy of fault identification and location.

[0030] It should be noted that the CBAM hybrid attention mechanism processing process is to first use the channel attention to obtain the output result of the one-dimensional time convolution block and then obtain the output through spatial attention.

[0031] S4, extract the feature vectors of multi-domain joint representation under multiple evaluation indicators respectively and use the global attention mechanism to obtain the weighted feature vector; It should be noted that in S4, STFT is used to extract The frequency domain features include the following steps: S401, Slice the time dimension into K subsequences to get the slice matrix: ; Among them, any subsequence ,L is the length of each subsequence; S402: for each sub-matrix in the slice matrix Calculate the STFT matrix: ; Among them, any one , the formula for calculating STFT is ; S403, where subsequence The sequence obtained after STFT calculation is Indicates frequency, is an imaginary unit, is a window function with a length of L; S404, the complex form of the STFT matrix Convert into amplitude matrix and phase matrix respectively; S405, using the Transformer encoder to extract the amplitude feature vectors of the amplitude matrix and the phase eigenvector of the phase matrix ; S406, concatenate the amplitude feature vector and the phase feature vector to form a frequency domain feature vector. The concatenation process is: .

[0032] in, express The Kth subsequence obtained by slicing the Nth link timing data under the evaluation index; express The bth subsequence obtained by slicing the time series data of the ath link under the evaluation index; express The STFT matrix obtained by STFT calculation after slicing the time series data under the evaluation index; Represents a subsequence The sequence obtained after STFT calculation; express The second data point in Represents a subsequence The sequence obtained after STFT calculation; Represents a subsequence Perform STFT calculations; Indicates the length of each subsequence obtained by slicing; It should be noted that obtaining the weighted feature vector using the global attention mechanism in S4 includes the following steps: S411, the feature vectors of the multi-domain joint representation under the multiple evaluation indicators are: As input, calculate the global attention weight W = ; Among them, each ; S412 uses the global attention weight to perform weighted summation on the feature vector to obtain the weighted feature vector: .

[0033] S5, the weighted feature vector is output through the fully connected layer to predict the fault link category for location.

[0034] It should be noted that the process of passing through the fully connected layer in S5 is ,in, is the predicted probability distribution, is the weight of the fully connected layer, is the bias of the fully connected layer.

[0035] It should be noted that the model training in this embodiment uses the cross entropy loss function to measure the difference between the prediction result and the true category label, where the cross entropy loss function is in, represents the feature vector after weighted summation of feature vectors using global attention weights; The feature vector representing the joint representation of multiple domains under the Mth evaluation index; The feature vector representing the joint representation of multiple domains under various evaluation indicators; express The weight coefficient of Indicates the number of types of evaluation indicators; The feature vector representing the joint representation of multiple domains under the kth evaluation index; N represents the number of samples; K represents the number of categories; Indicates the probability of the model predicting the i-th sample; Represents the one-hot encoding of the true label of the i-th sample; refer to Figure 3 , Figure 4 , Figure 5 and Figure 6 , Figure 3This is a comparison of the recognition accuracy curves during the training process. It can be seen from the figure that the recognition accuracy of using multi-domain feature joint representation and mixed attention is better than that of multi-domain joint representation using only time domain features and without mixed attention. Figure 4 , Figure 5 and Figure 6 They are the confusion matrix under time domain features, the confusion matrix of multi-domain joint representation, and the confusion matrix of multi-domain joint representation and mixed attention. It can be seen from the figure that multi-domain joint representation and mixed attention have better classification effects for link faults.

[0036] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.

Claims

1. A wireless communication link fault identification and location method based on multi-domain joint representation and hybrid attention, characterized by: The following steps are involved: S1, determining the wireless communication network topology structure as each link number; S2, collecting time series data of various quality evaluation indicators of wireless communication network links under each link failure and normal operation; S3, based on the hybrid attention mechanism, extracts the time domain, space domain and frequency domain features of the overall network to construct a feature vector of multi-domain joint representation; S4, extract the feature vectors of multi-domain joint representation under multiple evaluation indicators respectively and use the global attention mechanism to obtain the weighted feature vector; S5, the weighted feature vector is output through the fully connected layer to predict the fault link category for location.

2. According to claim 1, the wireless communication link fault identification and location method based on multi-domain joint representation and hybrid attention is characterized in that: In S1, the numbering of each link includes the numbering of each edge of the wireless communication link network. The network topology structure of the wireless communication link includes N edges, which includes a total of N+1 wireless communication link failure categories, and an additional one is a normal working category.

3. The wireless communication link fault identification and location method based on multi-domain joint representation and hybrid attention according to claim 1 is characterized in that: In S2, the wireless communication link multiple evaluation index time series data includes M evaluation indexes, and the multiple evaluation index time series data are constructed: Q={ } ； in, It is the time series data matrix of the overall wireless communication network link under one evaluation index: = ； Wherein, N is the link dimension of the network topology structure of the wireless communication link, T is the time dimension of the time series data, and any one for The evaluation index is the time series data of link a at time point b.

4. The wireless communication link fault identification and location method based on multi-domain joint representation and hybrid attention according to claim 1 is characterized in that: In S3, the time domain, space domain and frequency domain features of the overall network are extracted based on the hybrid attention mechanism to construct a feature vector of multi-domain joint representation, which specifically includes the following steps: The Each row is input into the Transformer encoder to extract the time characteristics of each link timing data under an evaluation index to obtain the time feature vector ; Extraction using FCN fully convolutional network and hybrid attention mechanism The spatial eigenvector of ; Extraction using STFT The frequency domain eigenvector of ; The time feature vector , spatial eigenvector and frequency domain eigenvector Concatenate the feature vectors that form a joint representation of multiple domains , where the splicing process is .

5. The wireless communication link fault identification and location method based on multi-domain joint representation and hybrid attention according to claim 4 is characterized in that: The FCN fully convolutional network and hybrid attention mechanism apply the CBAM hybrid attention mechanism after the output of the first two layers of one-dimensional temporal convolution blocks of the FCN fully convolutional network to enhance the relational features between different links and the spatial features of key features at different time steps.

6. The wireless communication link fault identification and location method based on multi-domain joint representation and hybrid attention according to claim 4 is characterized in that: The STFT extraction The frequency domain characteristics of the , specifically including the following steps: Will Slice the time dimension into K subsequences to get the slice matrix: ； Among them, any subsequence ,L is the length of each subsequence; For each sub-matrix in the slice matrix Calculate the STFT matrix: ； Among them, any one , the formula for calculating STFT is ; in, subsequence The sequence obtained after STFT calculation is Indicates frequency, is an imaginary unit, is a window function with a length of L; The complex form of the STFT matrix Convert into amplitude matrix and phase matrix respectively; Use the Transformer encoder to extract the amplitude feature vectors of the amplitude matrix respectively and the phase eigenvector of the phase matrix ; The amplitude eigenvector and the phase eigenvector are concatenated to form the frequency domain eigenvector. The concatenation process is: .

7. The wireless communication link fault identification and location method based on multi-domain joint representation and hybrid attention according to claim 1 is characterized in that: In S4, obtaining a weighted feature vector using a global attention mechanism includes the following steps: The feature vectors jointly representing multiple domains under the various evaluation indicators are: As input, calculate the global attention weight W = ; Among them, each Use the global attention weight to perform weighted summation on the feature vector to obtain the weighted feature vector .

8. The wireless communication link fault identification and location method based on multi-domain joint representation and hybrid attention according to claim 1 is characterized in that: In S5, the weighted feature vector is output through the fully connected layer to predict the fault link category to achieve positioning, which is specifically performed in the following manner: The fully connected layer is ,in, For the predicted results, is the weight of the fully connected layer, is the bias of the fully connected layer; The model training uses a cross entropy loss function to measure the difference between the predicted result and the true category label. The cross entropy loss function is: ; in, represents the feature vector after weighted summation of feature vectors using global attention weights; N represents the number of samples; K represents the number of categories; Indicates the probability of the model predicting the i-th sample; Represents the one-hot encoding of the true label of the i-th sample.

Images