A deep learning-based communication interference intelligent identification method

Abstract

The application provides a kind of communication interference intelligent identification method based on deep learning, belongs to communication data processing technical field, to adopt space-time two-way memory resonance network and antagonistic complementary training framework, can deeply mine the space-time coupling characteristics in communication data and accurately identify various kinds of interference in small sample environment.The application realizes dynamic coupling and mutual modulation of time and space characteristics in the extraction stage by constructing a space-time two-way memory resonance network, breaking through the limitations of feature separation processing in traditional methods, so that the model can deeply mine and understand the propagation and evolution law of interference in time and space dimensions, and improve the discovery and identification ability of early micro signs and complex coupled interference.

Description

Technical Field

[0001] This invention relates to the field of communication data processing technology, and in particular to a method for intelligent identification of communication interference based on deep learning. Background Technology

[0002] As the infrastructure of modern information society, the stability and reliability of communication networks are of paramount importance. During communication transmission, signals can be affected by various interferences from internal equipment failures or external environmental factors, leading to a decline in communication quality or even interruption. Quickly and accurately identifying and locating communication interference is a core task for ensuring uninterrupted communication links and network service quality, and represents an important research direction in communication transmission technology.

[0003] In existing technologies, the identification of communication interference mainly relies on human experience or monitoring and alarm systems based on fixed thresholds. Operation and maintenance personnel determine the existence of interference by analyzing network performance indicators such as traffic, latency, packet loss rate, and device event logs. With the development of artificial intelligence technology, some methods have begun to try to use deep learning models, such as using long short-term memory networks to analyze time series data of performance indicators to discover abnormal patterns, or using graph neural networks to analyze network topology to mine the spatial correlation between nodes.

[0004] Existing technologies have obvious shortcomings. Methods based on fixed thresholds are poorly adaptable to dynamically changing interference patterns and are prone to generating a large number of false alarms and false negatives. Existing artificial intelligence methods usually process temporal and spatial features independently and then simply fuse them, failing to effectively capture the complex mutual influence between the two. This results in insufficient ability to identify coupled interference across regions and time, leaving room for improvement. Summary of the Invention

[0005] This invention provides a deep learning-based intelligent identification method for communication interference, which employs a spatiotemporal bidirectional memory resonance network and an adversarial complementary training framework to deeply mine the spatiotemporal coupling features in communication data and accurately identify various types of interference in a small sample environment.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a deep learning-based intelligent identification method for communication interference is provided, applied to an identification terminal. The method includes the following steps: Receive performance index data streams and device log data streams reported by communication links and communication devices, calculate the fluctuation amplitude of the performance index data streams and the device log data streams, dynamically adjust the width of the time alignment window based on the fluctuation amplitude, map the performance index data streams and the device log data streams into a unified feature space for splicing, and generate aligned multi-source time series data; The aligned multi-source time series data is input into a preset spatiotemporal feature fusion network. Temporal features are extracted through the temporal feature extraction branch in the spatiotemporal feature fusion network, and the change rate of temporal features at adjacent time steps is calculated to output a fluctuation trend matrix. The fluctuation trend matrix is ​​converted into a dynamic modulation matrix, and the dynamic modulation matrix is ​​used to perform a Hadamard product operation with the initial coupling weight matrix representing the static network topology to update the topology connection weights and generate a spatiotemporal coupling feature map. The spatiotemporal coupling feature map is input into a multi-branch attention pooling layer. Multiple parallel attention branches within the multi-branch attention pooling layer are controlled to calculate and generate corresponding attention weight vectors. The inner product of each attention weight vector is scored through a scoring network layer containing learnable scoring parameters. The attention weight vector with the largest score is selected as the optimal weight and the spatiotemporal coupling feature map is weighted and summed to generate attention weighted features. Based on the attention-weighted features, the classifier outputs a probability vector of communication interference for each network node.

[0007] Optionally, before executing the step of receiving the performance index data stream and device log data stream reported by the communication link and communication device, a step of training the backbone recognition model composed of the spatiotemporal feature fusion network and the multi-branch attention pooling layer is further included, specifically including: The initial backbone identification model is iteratively trained using historical samples containing real interference labels to obtain the first backbone identification model. The verification sample set is input into the first backbone recognition model for forward inference, and the probability distribution vector of each network node at each time step is output. The maximum value in the probability distribution vector is extracted as the prediction confidence, and a confidence heatmap is generated to identify the low confidence region where the prediction confidence is lower than the benchmark value. By using an adversarial generative network, the corresponding fluctuation trend matrix is ​​extracted from the low confidence region identified by the confidence heatmap, and then concatenated with a random noise vector and input into the generator in the adversarial generative network to output targeted virtual interference samples. The virtual interference samples are mixed with the historical samples to form a hybrid dataset. The hybrid dataset is then used to perform alternating adversarial training on the first backbone recognition model to obtain the trained backbone recognition model.

[0008] Optionally, the step of extracting temporal features through the temporal feature extraction branch within the spatiotemporal feature fusion network, calculating the rate of change of temporal features at adjacent time steps to output a fluctuation trend matrix, and converting the fluctuation trend matrix into a dynamic modulation matrix specifically includes: Extract the temporal hidden state features from the aligned multi-source time series data, and calculate the difference vector of the temporal hidden state features between adjacent time points as the fluctuation rate vector. The fluctuation rate vector is input into a nonlinear activation function layer containing learnable parameters and subjected to numerical normalization to generate the fluctuation trend matrix. The dynamic modulation matrix is ​​constructed based on the fluctuation trend matrix corresponding to each network node, and the dynamic modulation matrix is ​​subjected to the Hadamard product operation with the initial coupling weight matrix to generate an updated coupling weight matrix for guiding the extraction of spatial graph convolutional features.

[0009] Optionally, the step of calculating the fluctuation range of the performance indicator data stream and the device log data stream, and dynamically adjusting the width of the time alignment window based on the fluctuation range, to map the performance indicator data stream and the device log data stream into a unified feature space for concatenation, specifically includes: The instantaneous change rate of each indicator in the performance indicator data stream and the device log data stream is calculated in real time, and the instantaneous change rate is weighted and summed to generate a fluctuation index that characterizes the overall volatility of the data stream. The fluctuation index is compared with a preset fluctuation reference threshold, and the adjusted alignment window width is calculated and determined between the preset maximum window length and the minimum window length using a preset nonlinear mapping formula containing an exponential function. Based on the adjusted alignment window width, the data time period is extracted, and the performance index data stream and the device log data stream within the data time period are converted into feature vectors and concatenated. A unified alignment timestamp is then assigned to generate the aligned multi-source time series data.

[0010] Optionally, the step of performing inner product scoring on each attention weight vector through a scoring network layer containing learnable scoring parameters, and selecting the attention weight vector with the largest score as the optimal weight to perform weighted summation on the spatiotemporal coupled feature map, specifically includes: The multiple parallel attention branches use their respective independent query matrices to perform self-attention operations on the spatiotemporal coupled feature map and normalize it using the Softmax function to generate a set of candidate attention weight vectors. The scoring network layer performs dot product operations between the learnable scoring parameters inside it and the set of candidate attention weight vectors to calculate the score of each branch, and selects the attention weight vector corresponding to the branch with the highest score as the winning weight vector. The winning weight vector and the spatiotemporal coupled feature map are subjected to matrix multiplication and pooling dimensionality reduction to generate the attention-weighted features.

[0011] Optionally, the step of extracting the corresponding fluctuation trend matrix from the low-confidence regions identified by the confidence heatmap using an adversarial generative network, concatenating it with a random noise vector, and inputting it into the generator within the adversarial generative network to output targeted virtual interference samples specifically includes: The adversarial generative network extracts the identifiers of target network nodes with confidence values ​​below a preset threshold and the target time intervals from the confidence heatmap. Based on the target network node identifier and the target time interval, the corresponding fluctuation trend matrix is ​​retrieved from the temporal feature extraction branch of the first backbone identification model as the environmental prior matrix. The environment prior matrix is ​​concatenated with a randomly generated noise vector and input into the generator network branch inside the adversarial generative network. The output is a time-series data segment that matches the environment prior matrix in terms of temporal jitter distribution and network topology coupling characteristics. These segments are then combined to form the targeted virtual interference sample.

[0012] Optionally, the step of performing alternating adversarial training on the first backbone recognition model using the hybrid dataset further includes: After each training iteration, the first backbone recognition model with updated parameters is used to infer the fixed verification sample set to update the confidence heatmap, and the updated confidence heatmap is input into the adversarial generative network. The adversarial generative network identifies newly added low-confidence regions based on the updated confidence heatmap and re-retrieves the fluctuation trend matrix to generate a new round of virtual interference samples. The first backbone recognition model updates its network parameters with the goal of minimizing the classification cross-entropy loss of the mixed dataset, while the adversarial generative network updates its network parameters with the goal of maximizing the classification error of the first backbone recognition model on the virtual interference samples. The two are executed alternately and cyclically until the recognition accuracy of the first backbone recognition model on the independent test set reaches a preset convergence threshold.

[0013] Optionally, the step of outputting a probability vector of communication interference for each network node through a classifier based on the attention-weighted features specifically includes: The attention-weighted features are input into a classification network branch consisting of fully connected layers, and the high-dimensional features are mapped into a fault score vector with a dimension equal to the total number of all interference categories. The fault score vector is input into the Softmax function to be converted into a standardized probability distribution vector, and the values ​​corresponding to each category in the probability distribution vector are used as the probability values ​​of the network node experiencing the corresponding type of communication interference.

[0014] Optionally, after outputting the probability vector of communication interference occurring at each network node, the following steps are performed: The probability values ​​of each category are compared with the preset alarm judgment threshold in real time. When the probability value of any category is detected to be greater than the alarm judgment threshold for the first time, an alarm trigger command is generated. In response to the alarm trigger command, an alarm data packet containing the IP address of the abnormal network node, the inferred interference type, and the exact timestamp is generated, and the alarm data packet is sent to the upper-layer network management platform for persistent recording through a preset secure communication channel.

[0015] Secondly, a deep learning-based intelligent identification device for communication interference is provided, applied to an identification terminal, the device comprising: The data receiving and alignment module is used to receive performance index data streams and device log data streams reported by the communication link and communication device, calculate the fluctuation amplitude of the performance index data streams and the device log data streams, dynamically adjust the width of the time alignment window according to the fluctuation amplitude, map the performance index data streams and device log data streams into a unified feature space for splicing, and generate aligned multi-source time series data. The spatiotemporal feature fusion module is used to input the aligned multi-source time series data into a preset spatiotemporal feature fusion network, extract time series features through the time series feature extraction branch in the spatiotemporal feature fusion network, and calculate the change rate of time series features at adjacent time steps to output a fluctuation trend matrix; convert the fluctuation trend matrix into a dynamic modulation matrix, and use the dynamic modulation matrix to perform a Hadamard product operation with the initial coupling weight matrix representing the static network topology to update the topology connection weights and generate a spatiotemporal coupling feature map; The multi-branch attention weighting module is used to input the spatiotemporal coupled feature map into the multi-branch attention pooling layer, control multiple parallel attention branches in the multi-branch attention pooling layer to calculate and generate corresponding attention weight vectors respectively, and perform inner product scoring on each attention weight vector through a scoring network layer containing learnable scoring parameters. The attention weight vector with the largest score value is selected as the optimal weight to perform weighted summation on the spatiotemporal coupled feature map to generate attention weighted features. The interference classification output module is used to output the probability vector of communication interference occurring at each network node through a classifier based on the attention-weighted features. The model adversarial training module is used to perform training steps, using historical samples of real interference labels and virtual interference samples generated by the adversarial generative network to alternately train the backbone recognition model composed of the spatiotemporal feature fusion network and the multi-branch attention pooling layer.

[0016] Thirdly, an electronic device is provided, comprising: a processor and a memory; the memory is used to store a computer program, which, when executed by the processor, causes the electronic device to perform the deep learning-based intelligent identification method for communication interference described in the first aspect.

[0017] In one possible design, the electronic device described in the third aspect may further include a transceiver. This transceiver may be a transceiver circuit or an interface circuit. The transceiver can be used for communication between the electronic device described in the third aspect and other electronic devices.

[0018] In the embodiments of the present invention, the electronic device described in the third aspect may be a terminal, or a chip (system) or other component or assembly disposed in the terminal, or a system containing the terminal.

[0019] Fourthly, a computer-readable storage medium is provided, comprising: a computer program or instructions; when the computer program or instructions are executed on a computer, the computer causes the computer to perform the deep learning-based intelligent identification method for communication interference described in the first aspect.

[0020] In summary, the above methods and systems have the following technical effects: 1. This invention constructs a spatiotemporal bidirectional memory resonance network, in which temporal and spatial features are dynamically coupled and mutually modulated during the extraction stage. This breaks through the limitations of feature separation processing in traditional methods, enabling the model to deeply mine and understand the propagation and evolution of interference in the temporal and spatial dimensions, and improving the ability to detect and identify early subtle signs and complex coupled interference.

[0021] 2. The adversarial complementary training framework designed in this invention can proactively discover and specifically compensate for the cognitive blind spots of the main model in small sample environments through the collaborative game between the main model and the challenge model. The targeted data augmentation method improves training efficiency and model generalization ability compared with the traditional random generation of virtual samples, and solves the problems of insufficient model training and low recognition rate of rare or new interference caused by the scarcity of real communication interference samples.

[0022] 3. The wave-aware elastic alignment technology and competitive attention pooling layer used in this invention work together. The former ensures the information fidelity and processing efficiency of the input data under different network conditions, while the latter quickly focuses on key features through a winner-takes-all mechanism, effectively reducing the inference latency of the model. The combination of the two ensures high recognition accuracy while taking into account the real-time performance of the model, enabling this method to meet the stringent performance requirements of online communication interference monitoring. Attached Figure Description

[0023] Figure 1A flowchart illustrating the intelligent identification method for communication interference based on deep learning provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of an apparatus for a deep learning-based intelligent identification method for communication interference provided in an embodiment of the present invention. Detailed Implementation

[0024] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0025] In this embodiment of the invention, "instruction" can include direct and indirect instructions, as well as explicit and implicit instructions. The information indicated by a certain piece of information is called the information to be instructed. In specific implementation, there are many ways to instruct the information to be instructed, such as, but not limited to, directly instructing the information to be instructed, such as the information to be instructed itself or its index. It can also indirectly instruct the information to be instructed by instructing other information, where there is a correlation between the other information and the information to be instructed. It can also instruct only a part of the information to be instructed, while the other parts are known or pre-agreed upon. For example, the instruction of specific information can be achieved by using a pre-agreed (e.g., protocol-defined) arrangement of various pieces of information, thereby reducing instruction overhead to some extent. Simultaneously, common parts of various pieces of information can be identified and uniformly indicated to reduce the instruction overhead caused by individually indicating the same information.

[0026] Furthermore, the specific indication method can also be any existing indication method, such as, but not limited to, the above-mentioned indication methods and their various combinations. Specific details of various indication methods can be found in existing technologies, and will not be elaborated upon here. As described above, for example, when multiple pieces of information of the same type need to be indicated, the indication methods for different pieces of information may differ. In specific implementation, the required indication method can be selected according to specific needs. This embodiment of the invention does not limit the selected indication method; therefore, the indication methods involved in this embodiment of the invention should be understood to cover various methods that enable the party to be indicated to obtain the information to be indicated.

[0027] It should be understood that the information to be indicated can be sent as a whole or divided into multiple sub-information messages sent separately, and the sending period and / or timing of these sub-information messages can be the same or different. The specific sending method is not limited in this embodiment of the invention. The sending period and / or timing of these sub-information messages can be predefined, for example, according to a protocol, or configured by the sending device by sending configuration information to the receiving device.

[0028] "Predefined" or "pre-configured" can be achieved by pre-saving corresponding codes, tables, or other means that can be used to indicate relevant information in the device. This embodiment of the invention does not limit the specific implementation method. "Saving" can refer to saving in one or more memories. These memories can be separate installations or integrated into the encoder, decoder, processor, or electronic device. Alternatively, some memories can be separately installed, while others are integrated into the decoder, processor, or electronic device. The type of memory can be any form of storage medium, and this embodiment of the invention does not limit this.

[0029] In the embodiments of this invention, "protocol" may refer to a protocol family in the field of communication, a standard protocol with a similar protocol family frame structure, or a related protocol applied to a future deep learning-based intelligent identification method system for communication interference. The embodiments of this invention do not specifically limit this.

[0030] In this embodiment of the invention, descriptions such as "when," "under the circumstances," "if," and "if" all refer to the device making corresponding processing under certain objective circumstances, and are not limited to a specific time. They do not require the device to make a judgment action during implementation, nor do they imply any other limitations.

[0031] In the description of the embodiments of the present invention, unless otherwise stated, " / " indicates that the objects before and after are in an "or" relationship. For example, A / B can represent A or B. "And / or" in the embodiments of the present invention is merely a description of the relationship between the related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone, where A and B can be singular or plural. Furthermore, in the description of the embodiments of the present invention, unless otherwise stated, "multiple" refers to two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple. Additionally, to facilitate a clear description of the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first," "second," etc., do not limit the quantity or order of execution, and that "first," "second," etc., are not necessarily different. Furthermore, in the embodiments of this invention, words such as "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or description. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this invention should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner for ease of understanding.

[0032] The network architecture and business scenarios described in the embodiments of this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of network architecture and the emergence of new business scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0033] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention. This deep learning-based intelligent identification method for communication interference can be applied to the aforementioned identification terminal. For example, the identification terminal can be a corresponding edge computing module with network transmission capabilities or other terminal devices. The specific process is as follows: A deep learning-based intelligent identification method for communication interference, applied to identification terminals, includes the following steps: Receive performance index data streams and device log data streams reported by communication links and communication devices, calculate the fluctuation range of the performance index data streams and device log data streams, dynamically adjust the width of the time alignment window based on the fluctuation range, map the performance index data streams and device log data streams into a unified feature space for splicing, and generate aligned multi-source time series data; After receiving performance index data streams and device log data streams reported by the communication link and communication equipment, the system calculates the instantaneous change rate of each index in real time to generate a volatility index that characterizes the overall volatility. : In the formula, Represents the instantaneous rate of change of the nth data source at time t; The weights are preset and the sum of all weights is 1. N is the total number of data sources participating in the calculation and is a positive integer greater than or equal to 1.

[0034] Subsequently, the volatility index The adjusted alignment window width is dynamically calculated by comparing it with a preset fluctuation reference threshold T. : In the formula, and These are the system's preset maximum and minimum window lengths, respectively; k is the steepness coefficient, used to control the sensitivity of the window width to changes in data fluctuations, and is used to align the window width. Consolidate the data streams with unified timestamps to generate aligned multi-source time series data. Aligned multi-source time series data are input into a pre-defined spatiotemporal feature fusion network. Temporal features are extracted through the temporal feature extraction branch within the spatiotemporal feature fusion network, and the rate of change of temporal features at adjacent time steps is calculated to output a fluctuation trend matrix. The fluctuation trend matrix is ​​converted into a dynamic modulation matrix, and the dynamic modulation matrix is ​​used to perform a Hadamard product operation with the initial coupling weight matrix representing the static network topology to update the topology connection weights and generate a spatiotemporal coupling feature map. Specifically: Multi-source time series data are input into a spatiotemporal feature fusion network, and temporal latent state features are extracted through its internal temporal feature extraction branch. Calculate the fluctuation rate vector at adjacent time points. : Subsequently, the fluctuation rate vector is input into a nonlinear activation function layer for numerical normalization, and the fluctuation trend matrix is ​​output. : In the formula, and These are the learnable weight matrix and bias vector, respectively; It is a hyperbolic tangent function that stabilizes its value in the interval [-1, 1].

[0035] The system is based on the fluctuation trend matrix of each node. Constructing a dynamic modulation matrix And with the initial coupling weight matrix characterizing the static network topology Perform the Hadamard product operation (element-by-element multiplication) to generate and update the coupling weight matrix. : ; This update weight This is used to guide subsequent spatial graph convolution feature extraction, generating a spatiotemporal coupled feature map that reflects real-time changes in topological connectivity.

[0036] The spatiotemporal coupled feature map is input into a multi-branch attention pooling layer. Multiple parallel attention branches within the multi-branch attention pooling layer are controlled to calculate and generate corresponding attention weight vectors. The inner product of each attention weight vector is scored through a scoring network layer containing learnable scoring parameters. The attention weight vector with the largest score is selected as the optimal weight and the spatiotemporal coupled feature map is weighted and summed to generate attention weighted features. Specifically: multi-branch attention pooling with competitive weighting, by inputting the spatiotemporally coupled feature map G into the multi-branch attention pooling layer, the multiple parallel attention branches inside each utilize independent query matrices. Calculate the corresponding attention weight vector : In the formula, This is the transpose of the feature map. The dimension for querying features.

[0037] Next, the attention weight vectors of each candidate are applied through a rating network layer containing learnable rating parameters s. Competition based on inner product scoring: Select rating The weight vector corresponding to the largest branch is used as the optimal / winning weight vector. The final attention-weighted feature F is generated: Based on attention-weighted features, a classifier outputs a probability vector of communication interference occurring at each network node. The attention-weighted feature F is input into the classifier branch composed of fully connected layers, and mapped to the fault score vector L: ; In the formula, This is the weight matrix of the fully connected layer. This is the bias vector.

[0038] Subsequently, the fault score vector L is transformed into a standardized probability distribution vector using a Softmax function to output the probability of communication interference occurring at each network node. : In the formula, C represents the total number of interference categories. Fault score corresponding to a specific interference category.

[0039] Step S5: Adversarial Complementary Training of the Backbone Recognition Model In a small-sample environment, the first backbone recognition model is inferred using the validation set, outputting the probability distribution vector P at each time step. The prediction confidence level S is selected as follows: ; Low-confidence regions where S is below the baseline value are identified, and confidence heatmaps are generated. Then, an adversarial generative network is activated to retrieve the corresponding environmental prior matrix (fluctuation trend matrix) from the low-confidence regions of the heatmap. By combining random noise Z with generator G, targeted virtual interference samples are generated. : ; The virtual interference sample is mixed with the historical real sample to form a hybrid dataset, which is then trained alternately by the first backbone recognition model and the adversarial generative network until the model reaches the convergence threshold.

[0040] Optionally, before executing the steps of receiving performance index data streams and device log data streams reported by the communication link and communication device, a step of training the backbone recognition model composed of a spatiotemporal feature fusion network and a multi-branch attention pooling layer is also included, specifically including: The initial backbone identification model is iteratively trained using historical samples containing real interference labels to obtain the first backbone identification model. The validation sample set is input into the first backbone recognition model for forward inference, and the probability distribution vector of each network node at each time step is output. The maximum value in the probability distribution vector is extracted as the prediction confidence, and a confidence heatmap is generated to identify the low confidence region where the prediction confidence is lower than the benchmark value. By using an adversarial generative network, the corresponding fluctuation trend matrix is ​​extracted from the low confidence region marked by the confidence heatmap. This matrix is ​​then concatenated with a random noise vector and input into the generator within the adversarial generative network to output targeted virtual interference samples. Virtual interference samples are mixed with historical samples to form a hybrid dataset. The hybrid dataset is then used to perform alternating adversarial training on the first backbone recognition model to obtain the trained backbone recognition model.

[0041] Optionally, temporal features are extracted through the temporal feature extraction branch within the spatiotemporal feature fusion network, and the rate of change of temporal features at adjacent time steps is calculated to output a fluctuation trend matrix; the step of converting the fluctuation trend matrix into a dynamic modulation matrix specifically includes: Extract the temporal hidden state features from the aligned multi-source time series data, and calculate the difference vector of the temporal hidden state features between adjacent time points as the fluctuation rate vector. The fluctuation rate vector is input into a nonlinear activation function layer containing learnable parameters and then numerically normalized to generate a fluctuation trend matrix. A dynamic modulation matrix is ​​constructed based on the fluctuation trend matrix corresponding to each network node, and the dynamic modulation matrix is ​​then subjected to a Hadamard product operation with the initial coupling weight matrix to generate an updated coupling weight matrix for guiding the extraction of spatial graph convolutional features.

[0042] Optionally, the steps of calculating the fluctuation range of the performance indicator data stream and the device log data stream, dynamically adjusting the width of the time alignment window based on the fluctuation range, and mapping the performance indicator data stream and the device log data stream to a unified feature space for concatenation specifically include: The instantaneous change rate of each indicator in the performance indicator data stream and the device log data stream is calculated in real time, and the instantaneous change rate is weighted and summed to generate a fluctuation index that characterizes the overall fluctuation intensity of the data stream. The volatility index is compared with a preset volatility reference threshold, and the adjusted alignment window width is calculated and determined between the preset maximum window length and the minimum window length using a preset nonlinear mapping formula that includes an exponential function. Based on the adjusted alignment window width, the data time period is extracted, and the performance index data stream and device log data stream within this data time period are converted into feature vectors and concatenated. A unified alignment timestamp is then assigned to generate aligned multi-source time series data.

[0043] Optionally, the step of performing inner product scoring on each attention weight vector through a scoring network layer containing learnable scoring parameters, and selecting the attention weight vector with the largest score as the optimal weight to perform weighted summation on the spatiotemporally coupled feature map, specifically includes: Multiple parallel attention branches use their own independent query matrices to perform self-attention operations on the spatiotemporal coupled feature map and normalize it using the Softmax function to generate a set of candidate attention weight vectors; The scoring network layer performs dot product operations on its internal learnable scoring parameters with a set of candidate attention weight vectors to calculate the score of each branch, and selects the attention weight vector corresponding to the branch with the highest score as the winning weight vector. The winning weight vector and the spatiotemporally coupled feature map are subjected to matrix multiplication and pooling dimensionality reduction to generate attention-weighted features.

[0044] Optionally, the steps of extracting the corresponding fluctuation trend matrix from the low-confidence regions identified by the confidence heatmap using an adversarial generative network, concatenating it with a random noise vector, and inputting it into the generator within the adversarial generative network to output targeted virtual interference samples specifically include: Generative adversarial networks extract the identifiers of target network nodes with confidence values ​​below a preset threshold and the target time intervals from the confidence heatmap. Based on the target network node identifier and the target time interval, the corresponding fluctuation trend matrix is ​​retrieved from the temporal feature extraction branch of the first backbone identification model as the environmental prior matrix. The environmental prior matrix is ​​concatenated with a randomly generated noise vector and input into the generator network branch inside the generative adversarial network. The output is a time-series data fragment that matches the environmental prior matrix in terms of temporal jitter distribution and network topology coupling characteristics. These fragments are then combined to form targeted virtual interference samples.

[0045] Optionally, the step of performing alternating adversarial training on the first backbone recognition model using a mixed dataset further includes: After each training iteration, the first backbone recognition model with updated parameters is used to infer the confidence heatmap on a fixed set of verification samples, and the updated confidence heatmap is then input into the generative adversarial network. The Generative Adversarial Network identifies newly added low-confidence regions based on the updated confidence heatmap and retrieves the fluctuation trend matrix again to generate a new round of virtual interference samples. The first backbone recognition model updates its network parameters with the goal of minimizing the classification cross-entropy loss of the mixed dataset. At the same time, the adversarial generative network updates its network parameters with the goal of maximizing the classification error of the first backbone recognition model on virtual interference samples. The two are executed alternately and cyclically until the recognition accuracy of the first backbone recognition model on the independent test set reaches the preset convergence threshold.

[0046] Optionally, the step of outputting the probability vector of communication interference for each network node through a classifier based on attention-weighted features specifically includes: The attention-weighted features are input into a classification network branch consisting of fully connected layers, and the high-dimensional features are mapped into a fault score vector with a dimension equal to the total number of all interference categories. The fault score vector is input into the Softmax function to be converted into a standardized probability distribution vector. The values ​​corresponding to each category in the probability distribution vector are used as the probability values ​​of the corresponding type of communication interference occurring in the network node.

[0047] Optionally, after outputting the probability vector of communication interference occurring at each network node, the following steps are performed: The probability values ​​of each category are compared with the preset alarm judgment threshold in real time. When the probability value of any category is detected to be greater than the alarm judgment threshold for the first time, an alarm trigger command is generated. In response to an alarm trigger command, an alarm data packet containing the IP address of the abnormal network node, the inferred type of interference, and the exact timestamp is generated. The alarm data packet is then sent to the upper-layer network management platform for persistent recording through a preset secure communication channel.

[0048] Secondly, a deep learning-based intelligent communication interference identification device is provided for use in identification terminals. The device includes: The data receiving and alignment module is used to receive performance index data streams and device log data streams reported by communication links and communication devices, calculate the fluctuation range of the performance index data streams and device log data streams, dynamically adjust the width of the time alignment window according to the fluctuation range, map the performance index data streams and device log data streams into a unified feature space for splicing, and generate aligned multi-source time series data. The spatiotemporal feature fusion module is used to input aligned multi-source time series data into a preset spatiotemporal feature fusion network, extract time series features through the time series feature extraction branch in the spatiotemporal feature fusion network, and calculate the rate of change of time series features at adjacent time steps to output a fluctuation trend matrix; the fluctuation trend matrix is ​​converted into a dynamic modulation matrix, and the dynamic modulation matrix is ​​used to perform a Hadamard product operation with the initial coupling weight matrix representing the static network topology to update the topology connection weights and generate a spatiotemporal coupling feature map; The multi-branch attention weighting module is used to input the spatiotemporal coupled feature map into the multi-branch attention pooling layer, control multiple parallel attention branches in the multi-branch attention pooling layer to calculate and generate corresponding attention weight vectors, and perform inner product scoring on each attention weight vector through a scoring network layer containing learnable scoring parameters. The attention weight vector with the largest score value is selected as the optimal weight to perform weighted summation on the spatiotemporal coupled feature map to generate attention weighted features. The interference classification output module is used to output the probability vector of communication interference occurring at each network node through a classifier based on attention-weighted features. The model adversarial training module is used to perform training steps, using historical samples with real interference labels and virtual interference samples generated by the adversarial generative network to alternately train the backbone recognition model, which consists of a spatiotemporal feature fusion network and a multi-branch attention pooling layer.

[0049] The electronic device provided in this embodiment of the invention, exemplarily, can be a network device, or a chip (system) or other component or assembly that can be disposed in a network device. The electronic device may include a processor. Optionally, the electronic device may also include a memory and / or a transceiver. The processor is coupled to the memory and transceiver, for example, by means of a communication bus connection.

[0050] The following is a detailed introduction to the various components of the electronic device: In this context, the processor is the control center of the electronic device. It can be a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0051] Alternatively, the processor can perform various functions of the electronic device, such as the methods described above, by running or executing software programs stored in memory and by calling data stored in memory.

[0052] In a specific implementation, as one example, the processor may include one or more CPUs, such as CPU0 and CPU1.

[0053] In a specific implementation, as one example, the electronic device may also include multiple processors. Each of these processors may be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0054] The memory is used to store the software program that executes the solution of the present invention, and the execution is controlled by the processor. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0055] Optionally, the memory can be read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through the interface circuit of the electronic device; the embodiments of the present invention do not specifically limit this.

[0056] A transceiver is used for communication with other electronic devices. For example, if the electronic device is a terminal, the transceiver can be used to communicate with a network device or with another terminal device. Similarly, if the electronic device is a network device, the transceiver can be used to communicate with a terminal or with another network device.

[0057] Optionally, the transceiver may include a receiver and a transmitter. The receiver is used to implement the receiving function, and the transmitter is used to implement the sending function.

[0058] Optionally, the transceiver can be integrated with the processor or exist independently and coupled to the processor through the interface circuit of the electronic device. This embodiment of the invention does not specifically limit this.

[0059] It is understood that the structure of the electronic device in this embodiment does not constitute a limitation on the electronic device. The actual electronic device may include more or fewer components, or combine certain components, or have different component arrangements.

[0060] Furthermore, the technical effects of the electronic device can be referred to the technical effects of the method described in the above method embodiments, and will not be repeated here.

[0061] It should be understood that the processor in the embodiments of the present invention can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0062] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0063] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0064] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0065] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0066] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0067] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0068] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0069] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0070] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0071] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0072] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0073] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning-based intelligent identification method for communication interference, applied to identification terminals, characterized in that, The method includes the following steps: Receive performance index data streams and device log data streams reported by communication links and communication devices, calculate the fluctuation amplitude of the performance index data streams and the device log data streams, dynamically adjust the width of the time alignment window based on the fluctuation amplitude, map the performance index data streams and the device log data streams into a unified feature space for splicing, and generate aligned multi-source time series data; The aligned multi-source time series data is input into a preset spatiotemporal feature fusion network. Temporal features are extracted through the temporal feature extraction branch in the spatiotemporal feature fusion network, and the change rate of temporal features at adjacent time steps is calculated to output a fluctuation trend matrix. The fluctuation trend matrix is ​​converted into a dynamic modulation matrix, and the dynamic modulation matrix is ​​used to perform a Hadamard product operation with the initial coupling weight matrix representing the static network topology to update the topology connection weights and generate a spatiotemporal coupling feature map. The spatiotemporal coupling feature map is input into a multi-branch attention pooling layer. Multiple parallel attention branches within the multi-branch attention pooling layer are controlled to calculate and generate corresponding attention weight vectors. The inner product of each attention weight vector is scored through a scoring network layer containing learnable scoring parameters. The attention weight vector with the largest score is selected as the optimal weight and the spatiotemporal coupling feature map is weighted and summed to generate attention weighted features. Based on the attention-weighted features, the classifier outputs a probability vector of communication interference for each network node.

2. The deep learning-based intelligent identification method for communication interference according to claim 1, characterized in that, Before executing the step of receiving the performance index data stream and device log data stream reported by the communication link and communication device, the method further includes a step of training the backbone recognition model composed of the spatiotemporal feature fusion network and the multi-branch attention pooling layer, specifically including: The initial backbone identification model is iteratively trained using historical samples containing real interference labels to obtain the first backbone identification model. The verification sample set is input into the first backbone recognition model for forward inference, and the probability distribution vector of each network node at each time step is output. The maximum value in the probability distribution vector is extracted as the prediction confidence, and a confidence heatmap is generated to identify the low confidence region where the prediction confidence is lower than the benchmark value. By using an adversarial generative network, the corresponding fluctuation trend matrix is ​​extracted from the low confidence region identified by the confidence heatmap, and then concatenated with a random noise vector and input into the generator in the adversarial generative network to output targeted virtual interference samples. The virtual interference samples are mixed with the historical samples to form a hybrid dataset. The hybrid dataset is then used to perform alternating adversarial training on the first backbone recognition model to obtain the trained backbone recognition model.

3. The intelligent identification method for communication interference based on deep learning according to claim 1, characterized in that, The process involves extracting temporal features through the temporal feature extraction branch within the spatiotemporal feature fusion network, and calculating the rate of change of temporal features at adjacent time steps to output a fluctuation trend matrix. The step of converting the fluctuation trend matrix into a dynamic modulation matrix specifically includes: Extract the temporal hidden state features from the aligned multi-source time series data, and calculate the difference vector of the temporal hidden state features between adjacent time points as the fluctuation rate vector. The fluctuation rate vector is input into a nonlinear activation function layer containing learnable parameters and subjected to numerical normalization to generate the fluctuation trend matrix. The dynamic modulation matrix is ​​constructed based on the fluctuation trend matrix corresponding to each network node, and the dynamic modulation matrix is ​​subjected to the Hadamard product operation with the initial coupling weight matrix to generate an updated coupling weight matrix for guiding the extraction of spatial graph convolutional features.

4. The intelligent identification method for communication interference based on deep learning according to claim 1, characterized in that, The steps of calculating the fluctuation range of the performance indicator data stream and the device log data stream, dynamically adjusting the width of the time alignment window based on the fluctuation range, and mapping the performance indicator data stream and the device log data stream to a unified feature space for concatenation specifically include: The instantaneous change rate of each indicator in the performance indicator data stream and the device log data stream is calculated in real time, and the instantaneous change rate is weighted and summed to generate a fluctuation index that characterizes the overall volatility of the data stream. The fluctuation index is compared with a preset fluctuation reference threshold, and the adjusted alignment window width is calculated and determined between the preset maximum window length and the minimum window length using a preset nonlinear mapping formula containing an exponential function. Based on the adjusted alignment window width, the data time period is extracted, and the performance index data stream and the device log data stream within the data time period are converted into feature vectors and concatenated. A unified alignment timestamp is then assigned to generate the aligned multi-source time series data.

5. The intelligent identification method for communication interference based on deep learning according to claim 1, characterized in that, The step of performing inner product scoring on each attention weight vector through a scoring network layer containing learnable scoring parameters, and selecting the attention weight vector with the largest score as the optimal weight to perform weighted summation on the spatiotemporal coupled feature map specifically includes: The multiple parallel attention branches use their respective independent query matrices to perform self-attention operations on the spatiotemporal coupled feature map and normalize it using the Softmax function to generate a set of candidate attention weight vectors. The scoring network layer performs dot product operations between the learnable scoring parameters inside it and the set of candidate attention weight vectors to calculate the score of each branch, and selects the attention weight vector corresponding to the branch with the highest score as the winning weight vector. The winning weight vector and the spatiotemporal coupled feature map are subjected to matrix multiplication and pooling dimensionality reduction to generate the attention-weighted features.

6. The intelligent identification method for communication interference based on deep learning according to claim 2, characterized in that, The step of extracting the corresponding fluctuation trend matrix from the low-confidence regions identified by the confidence heatmap using an adversarial generative network, concatenating it with a random noise vector, and inputting it into the generator within the adversarial generative network to output targeted virtual interference samples specifically includes: The adversarial generative network extracts the identifiers of target network nodes with confidence values ​​below a preset threshold and the target time intervals from the confidence heatmap. Based on the target network node identifier and the target time interval, the corresponding fluctuation trend matrix is ​​retrieved from the temporal feature extraction branch of the first backbone identification model as the environmental prior matrix. The environment prior matrix is ​​concatenated with a randomly generated noise vector and input into the generator network branch inside the adversarial generative network. The output is a time-series data segment that matches the environment prior matrix in terms of temporal jitter distribution and network topology coupling characteristics. These segments are then combined to form the targeted virtual interference sample.

7. The deep learning-based intelligent identification method for communication interference according to claim 2, characterized in that, The step of performing alternating adversarial training on the first backbone identification model using the hybrid dataset further includes: After each training iteration, the first backbone recognition model with updated parameters is used to infer the fixed verification sample set to update the confidence heatmap, and the updated confidence heatmap is input into the adversarial generative network. The adversarial generative network identifies newly added low-confidence regions based on the updated confidence heatmap and re-retrieves the fluctuation trend matrix to generate a new round of virtual interference samples. The first backbone recognition model updates its network parameters with the goal of minimizing the classification cross-entropy loss of the mixed dataset, while the adversarial generative network updates its network parameters with the goal of maximizing the classification error of the first backbone recognition model on the virtual interference samples. The two are executed alternately and cyclically until the recognition accuracy of the first backbone recognition model on the independent test set reaches a preset convergence threshold.

8. The intelligent identification method for communication interference based on deep learning according to claim 1, characterized in that, The step of outputting the probability vector of communication interference for each network node through a classifier based on the attention-weighted features specifically includes: The attention-weighted features are input into a classification network branch consisting of fully connected layers, and the high-dimensional features are mapped into a fault score vector with a dimension equal to the total number of all interference categories. The fault score vector is input into the Softmax function to be converted into a standardized probability distribution vector, and the values ​​corresponding to each category in the probability distribution vector are used as the probability values ​​of the network node experiencing the corresponding type of communication interference.

9. The intelligent identification method for communication interference based on deep learning according to claim 1, characterized in that, It also includes performing the following steps after outputting the probability vector of communication interference occurring at each network node: The probability values ​​of each category are compared with the preset alarm judgment threshold in real time. When the probability value of any category is detected to be greater than the alarm judgment threshold for the first time, an alarm trigger command is generated. In response to the alarm trigger command, an alarm data packet containing the IP address of the abnormal network node, the inferred interference type, and the exact timestamp is generated, and the alarm data packet is sent to the upper-layer network management platform for persistent recording through a preset secure communication channel.

10. A communication interference intelligent identification device based on deep learning, characterized in that, The device, used in an identification terminal, includes: The data receiving and alignment module is used to receive performance index data streams and device log data streams reported by the communication link and communication device, calculate the fluctuation amplitude of the performance index data streams and the device log data streams, dynamically adjust the width of the time alignment window according to the fluctuation amplitude, map the performance index data streams and device log data streams into a unified feature space for splicing, and generate aligned multi-source time series data. The spatiotemporal feature fusion module is used to input the aligned multi-source time series data into a preset spatiotemporal feature fusion network, extract time series features through the time series feature extraction branch in the spatiotemporal feature fusion network, and calculate the change rate of time series features at adjacent time steps to output a fluctuation trend matrix; convert the fluctuation trend matrix into a dynamic modulation matrix, and use the dynamic modulation matrix to perform a Hadamard product operation with the initial coupling weight matrix representing the static network topology to update the topology connection weights and generate a spatiotemporal coupling feature map; The multi-branch attention weighting module is used to input the spatiotemporal coupled feature map into the multi-branch attention pooling layer, control multiple parallel attention branches in the multi-branch attention pooling layer to calculate and generate corresponding attention weight vectors respectively, and perform inner product scoring on each attention weight vector through a scoring network layer containing learnable scoring parameters. The attention weight vector with the largest score value is selected as the optimal weight to perform weighted summation on the spatiotemporal coupled feature map to generate attention weighted features. The interference classification output module is used to output the probability vector of communication interference occurring at each network node through a classifier based on the attention-weighted features. The model adversarial training module is used to perform training steps, using historical samples of real interference labels and virtual interference samples generated by the adversarial generative network to alternately train the backbone recognition model composed of the spatiotemporal feature fusion network and the multi-branch attention pooling layer.

Images