A method and system for detecting anomalies in real-time radio field strength 3D waterfall diagram based on multi-view learning

Abstract

The present invention relates to a method and system for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning, comprising: data acquisition and processing: real-time acquisition of spectrum, field strength and time parameters by radio monitoring equipment; preliminary processing and cleaning of the acquired data to remove noise and abnormal data, and then automatically constructing a 3D waterfall diagram according to the three-dimensional information of field strength-time-frequency; classifying, labeling and analyzing a large amount of acquired monitoring data to construct a data set for different abnormal signals, including normal data, malicious electromagnetic interference data, unauthorized signal occurrence data and authorized signal illegal use data; constructing a deep learning anomaly detection model and training it; and realizing 3D waterfall diagram anomaly detection by the trained deep learning anomaly detection model.

Description

Technical Field

[0001] The present invention relates to a method and system for detecting anomalies in a real-time radio field strength 3D waterfall chart based on multi-view learning, which involves data acquisition and visualization of radio monitoring equipment, bidirectional long short-term memory network, multi-view representation fusion learning, cross-view attention fusion mechanism, etc., and belongs to the intersection of the Internet of Things and artificial intelligence. Background Art

[0002] Radio monitoring equipment plays an indispensable role in the field of wireless communications. With the rapid development of wireless communication technology, the demand for the use of various wireless devices and frequency bands is increasing, and the tension in spectrum resources has become more and more obvious. Standardized management of spectrum is not only a necessary means to maintain public resources, but also an important measure to ensure communication security and order. Therefore, radio regulatory agencies around the world are using advanced monitoring equipment to monitor and manage spectrum in real time to ensure the efficient use of legal spectrum and to detect, identify and process illegal or abnormal signals in a timely manner.

[0003] Traditional radio monitoring equipment mainly collects electromagnetic wave parameters such as field strength, frequency, and time in real time to construct a three-dimensional (3D) waterfall flow chart (i.e., a 3D chart of field strength-time-frequency). The data richness of the 3D waterfall chart is reflected in the fact that it can not only display the instantaneous strength of the signal, but also reveal the spectrum distribution of the signal in different frequency bands and time periods. By observing the 3D waterfall flow chart, managers can intuitively monitor the changes of radio signals on the timeline and view the global view of the spectrum and field strength at the same time. This has extremely high practical value for the detection of abnormal signals, especially when monitoring spectrum anomalies such as malicious electromagnetic interference, the appearance of unauthorized signals, and the illegal use of authorized signals. The 3D waterfall chart can provide key clues.

[0004] In the early days of radio spectrum detection, the detection process mainly relied on the staff's experience and professional knowledge reserves to manually judge whether the signal status was normal. However, manual detection has certain limitations. Due to subjective factors and differences in personnel experience, the risk of misjudgment is high, and the operation is cumbersome and time-consuming, which seriously affects the detection efficiency. With the rapid development of radio communication technology, the demand for spectrum anomaly detection technology is increasing. At present, spectrum anomaly detection research is divided into two categories: traditional solutions and intelligent solutions. Traditional solutions are mainly based on signal feature extraction and analysis, while intelligent solutions rely on artificial intelligence and machine learning technologies to improve detection results.

[0005] Traditional spectrum anomaly detection methods judge anomalies by analyzing the characteristic parameters of signals during transmission and operation. These features include modulation mode, center frequency, bandwidth, changes in signal strength, etc. However, these methods often rely on a large amount of historical data and require complex feature engineering, which increases the difficulty of calculation and analysis. With the development of artificial intelligence technology, anomaly detection schemes based on machine learning have gradually become the mainstream in the field of spectrum management. Machine learning algorithms can reduce the reliance on manual feature extraction by automatically learning the temporal and spatial features of signals. At present, algorithms based on 3D deep learning classification have been applied to the processing of radio 3D waterfall flow charts. These algorithms directly extract features from 3D data through 3D convolutional networks and have achieved certain results. However, since such methods need to process a large amount of three-dimensional data, the computational cost is high, which limits their use in applications with high real-time requirements. In contrast, processing methods based on 2D multi-views have begun to receive more and more attention. This method reduces the computational complexity to a large extent by rendering 3D objects as 2D images from different perspectives and then performing convolution processing on each view. Popular view-based methods usually extract 3D visual representations by using view images about a single viewpoint, and perform fusion display of different views based on early or late fusion. However, this simple fusion method at different stages will face problems such as redundancy of information between views and uniqueness of information from different views during the processing, and will eventually lose the connection information between different views. Therefore, how to effectively solve the redundancy and uniqueness of multi-view learning after converting the radio field intensity waterfall flow chart into 2D views has become an important direction of current research. Summary of the invention

[0006] In order to solve the technical problems existing in the above-mentioned background technology, the present invention proposes a multi-view classification method for radio 3D waterfall flow chart. By decomposing the 3D graph into different 2D views, two representation methods are used to handle the heterogeneity of the view. The first is a common representation, which uses a unified encoder to learn and project all views into a shared subspace to achieve view distribution alignment and reduce information redundancy. The second is a private representation, which learns the subspace specific to each view and extracts unique view features. For timeline-related views (such as time-field intensity, time-frequency), LSTM is used to extract time series features, while for field intensity-frequency views, multi-layer perceptron MLP is used for feature extraction. For private space, an attention mechanism is introduced to obtain the final display of each unique space after eliminating redundancy. Finally, the present invention constructs a cross-view attention autoencoder fusion module to achieve information interaction and fusion between different views, and enhances the accuracy and efficiency of classification and anomaly detection. By utilizing this deep learning anomaly detection mechanism, anomaly detection and real-time alarm for unattended radio monitoring tasks are achieved.

[0007] Terminology explanation:

[0008] 1. LSTM model: LSTM (Long Short-Term Memory Network) is used to process and predict problems with long-term dependencies in time series. It is very suitable for processing time-related view data (such as time-field intensity, time-frequency), and is used in the present invention to extract the features of time axis-related views in 3D waterfall flow charts.

[0009] 2. Multilayer Perceptron MLP: It is a feedforward neural network consisting of an input layer, several hidden layers and an output layer. MLP learns the features of input data through nonlinear activation functions and back propagation algorithms. In the present invention, MLP is used for feature extraction of field intensity-frequency view, which is suitable for processing high-dimensional data with nonlinear relationships. It can effectively capture the correlation between field intensity and frequency by processing input information layer by layer and learning complex feature representations.

[0010] 3. Shared encoder: A shared encoder is an encoder used to extract common features from multi-view data. It operates in a shared subspace and encodes the common information of different views so that similar or related features of each view can be unified in a shared representation.

[0011] 4. Private encoder: A private encoder is an encoder specifically designed to capture unique information of each view. Each view has its own unique properties and features that may not exist in other views. The private encoder encodes each view in a specific subspace to generate a private representation of the view, thereby preserving the uniqueness of the view.

[0012] 5. Specific subspace: Specific subspace refers to the feature representation space unique to each view. In multi-view learning, different views usually contain their own unique information, and specific subspace is used to capture the unique attributes and features in these views.

[0013] 6. Private representation: Private representation refers to view-specific features learned in a specific subspace. The private representation of each view retains information specific to that view and is different from the representations of other views.

[0014] 7. Shared subspace: Shared subspace refers to the feature space shared by multiple views. In multi-view learning, although different views have their own characteristics, they may also have some common related information. Through shared subspace, the common information of each view is projected into the same subspace, so as to align the view distribution, reduce redundant data, and improve the overall feature fusion effect.

[0015] 8. Common representation: Common representation refers to the common features of various views learned in a shared subspace by a unified encoder.

[0016] 9. Cross-view attention autoencoder fusion module: The cross-view attention autoencoder fusion module is a mechanism proposed by the present invention for information interaction and fusion between different view representations. Through the attention mechanism, the module can weight the features of each view according to its importance and enhance the role of key features.

[0017] In order to achieve the above object, the present invention adopts the following technical solution:

[0018] The first aspect of the present disclosure provides a method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning;

[0019] A method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning, comprising:

[0020] Data collection and processing: The spectrum, field strength and time parameters are collected in real time through radio monitoring equipment. The collected data is preliminarily processed and cleaned to remove noise and abnormal data. Then, a 3D waterfall flow chart is automatically constructed based on the three-dimensional information of field strength-time-frequency.

[0021] Classify, annotate and analyze the large amount of monitoring data collected to construct data sets for different abnormal signals; including: normal data, malicious electromagnetic interference data, unauthorized signal appearance data and authorized signal illegal use data;

[0022] Build a deep learning anomaly detection model and train it;

[0023] 3D waterfall chart anomaly detection is achieved through the trained deep learning anomaly detection model.

[0024] Preferably, according to the present invention, the 3D image is decomposed into different 2D views, and two representation methods are used to handle the heterogeneity of the views: the first is a common representation, which uses a unified encoder to learn and project all views into a shared subspace; the second is a private representation, which learns a subspace specific to each view and extracts specific view features.

[0025] Preferably, according to the present invention, for the time axis related view, LSTM is used to extract the time series features; for the field strength-frequency view, a multi-layer perceptron MLP is used to extract the features.

[0026] Preferably, according to the present invention, an attention mechanism is introduced for private spaces to obtain a final display of each unique space after removing redundancy.

[0027] Preferably, according to the present invention, the 3D graph, ie, the 3D waterfall graph, includes information in three dimensions: field intensity, time, and frequency.

[0028] Preferably, according to the present invention, a time axis related view is constructed to display the changes of different data on the time axis;

[0029] Time axis related views include time-field intensity view and time-frequency view;

[0030] The time-field strength view is used to show the changes in the field strength of the radio signal on the time axis;

[0031] The time-frequency view is used to display spectrum distribution in different time periods and monitor changes in frequency usage.

[0032] Preferably, according to the present invention, the LSTM model is used to extract features from the time-field intensity view and the time-frequency view, including:

[0033] First, the forget gate is used to control the part of the cell state that needs to be forgotten. The calculation formula is:

[0034] Γ f (t) = σ(W f ·[a t-1 ,x t ]+b f )

[0035] Among them, Γ f (t) represents the value of the forget gate output, which controls the cell state C at the previous moment t -1 degree of retention; a t -1 is the hidden state of the previous time step; x t is the current input; W f is the weight matrix; b f is the bias term; σ is the Sigmoid activation function, which ensures that the output value is between U and 1;

[0036] Next, the input gate controls the extent to which new information is written to the cell state. The computational process is divided into three steps; including:

[0037] The first step is to calculate the value of the input gate and decide which information to update at the current moment. The calculation formula is:

[0038] Γ u (t) = σ(W u ·[a t-1 ,x t ]+b u )

[0039] The second step is to generate candidate cell states:

[0040]

[0041] The third step is to update the cell state by combining the output of the forget gate and the input gate to retain and update the cell state:

[0042]

[0043] Among them, Γ u (t) represents the value of the input gate output; it controls the proportion of new information added to the cell state: is the candidate cell state, which transforms the input through the hyperbolic tangent function tanh to ensure that the output is between -1 and 1; W u and W c is the weight matrix of the input gate and candidate state, b u and b c is the corresponding bias term; C t Indicates the cell state at the current moment, C t-1 represents the cell state at the previous time step;

[0044] Finally, the output gate generates the output based on the current cell state and the input information:

[0045] Γ o (t) = σ(W o ·[a t-1 ,x t ]+b o )

[0046] The output values ​​are:

[0047] FT t =Γ o (t)·tanh(C t )

[0048] Among them, Γ o (t) is the value of the output gate, which controls the output strength; FT t is the hidden state at the current moment, which is also the output of the LSTM model; tanh(C t ) represents the nonlinear transformation of the current cell state;

[0049] Based on the time series processing mechanism of the LSTM model, the feature representation FT of the time-field intensity view and the time-frequency view is generated by extracting the features of the time-field intensity view and the time-frequency view. t , the feature representation of the time-field intensity view is defined as FT1, and the feature representation of the time-frequency view is defined as FT2.

[0050] Preferably, according to the present invention, a multi-layer perceptron is used to extract features from the field intensity-frequency view; comprising:

[0051] The multi-layer perceptron includes several fully connected layers, each of which includes a weight matrix and a bias term;

[0052] First, assume that the input data of the field strength-frequency view is a vector S = [s1, s2, ..., s m ...,s M ], where s m Represents the field strength value corresponding to the mth frequency, and M is the total number of frequencies; the multi-layer perceptron MLP gradually extracts the high-dimensional features of the input data through multi-layer nonlinear mapping. For the first fully connected layer, the calculation formula for the feature representation is:

[0053] h (1) =σ(W (1) ·S+b (1) )

[0054] Among them, W (1) is the weight matrix of the first fully connected layer, b (1) is the bias vector, σ is the activation function, which maps the input to the hidden layer feature h (1) middle;

[0055] Next, enter the second fully connected layer, and the calculation formula is:

[0056] h (2) =σ(W (2) ·h (1) +b (2) )

[0057] Here, W (2) and b (2) are the weight matrix and bias vector of the second fully connected layer, and the output feature h (2) further mapped to higher dimensions;

[0058] Finally, the output feature representation is:

[0059] FS=W (3) ·h (2) +b (3)

[0060] Among them, W (3) and b (3) is the weight matrix and bias term of the output layer, and FS is the feature representation of the extracted field strength-frequency view.

[0061] Preferably, the specific subspace and the shared subspace process the shared information between views and the specific information of each view respectively through common representation and private representation according to the present invention; given a certain view m and its extracted feature vector u m , using a shared encoder C(;θ c ) maps it to a view-invariant shared subspace, while using a view-specific private encoder Embed a specific view into a unique subspace;

[0062] The common representation handles the shared information between views and uses a shared encoder C(·; θ c ), shared encoder C(·;θ c ) parameters θ are shared between all views c ; The formula is as follows:

[0063] C m =C(u m θ c )

[0064] Among them, C m represents the feature representation of view m in the shared subspace, u m is the input feature of view m, θ c are the parameters of the shared encoder;

[0065] The private representation is responsible for capturing the specific information of each view; through the view-specific private encoder The features of each view are embedded into its own specific subspace; the formula is as follows:

[0066]

[0067] in, represents the feature representation of view m in the unique subspace, are the private encoder parameters of view m, and different views have their own independent parameters;

[0068] The three views generate the following corresponding feature representations:

[0069] For visual Figure 1 That is, the time-field view: the shared subspace output is C1, and the private subspace output is

[0070] For visual Figure 2 That is, the time-frequency view: the shared subspace output is C2, and the private subspace output is

[0071] For visual Figure 3 That is, the field strength-frequency view: the shared subspace output is C3, and the private subspace output is

[0072] Preferably, according to the present invention, similarity constraints are added to the common representation, which is enhanced by cosine similarity loss to align common cross-view features in the view-shared subspace and reduce the differences between them;

[0073] Cosine Similarity Loss Function The definition is as follows:

[0074]

[0075] Preferably, according to the present invention, when extracting private representations, the attention mechanism of shared space and vector is used to further optimize the feature extraction process, for each view:

[0076] First, the weight of each modal feature is calculated through the attention mechanism. The formula is as follows

[0077]

[0078] Among them, α m is the attention weight of view m, is the feature representation of view m, is the attention vector, is the weight matrix, is the bias term; all views share parameters W4 and b4.

[0079] Use attention weights to calculate redundant features between views:

[0080]

[0081] Among them, E s Represents redundant features of all views;

[0082] Then, the feature representation in the unique subspace is obtained by subtracting the redundant features from the original features of each view, as follows:

[0083]

[0084] P m It refers to the final unique features obtained from the corresponding view m, that is, the final display of each unique space after eliminating redundancy.

[0085] Preferably, according to the present invention, the deep learning anomaly detection model is a cross-view attention autoencoder fusion model, which uses the cross-view attention autoencoder fusion model to fuse the common representation and the private representation to form the final feature; including:

[0086] Cross-view interaction stage;

[0087] The private representation of each view is calculated; three views P1, P2, P3 and their shared representations C1, C2, C3 have been obtained; the private representation of each view P m will be shared with all the m Perform attention calculation to generate updated private representation; the formula is as follows:

[0088] P m′=Attention(P m ,[C1,C2,C3])

[0089] Among them, P' m is the updated private representation, Attention is the standard self-attention mechanism, and the calculation method is:

[0090]

[0091] Among them, Q, K, and V are respectively combined with different weight matrices W through the input matrix Q , W K , W V Linear transformation yields, d k It is the dimension of Key, used for normalization;

[0092] Encoder fusion stage:

[0093] In the encoder fusion stage, the updated private representation is finally fused using the autoencoder; each view P' m As input, the autoencoder generates the final vector representation:

[0094] Z m =Autoencoder(P m ′)

[0095] Finally, the outputs of the three views Z1, Z2, and Z3 are concatenated to form the final feature vector for prediction:

[0096] Z final =[Z1,Z2,Z3]

[0097] This final feature vector passes through the subsequent classification layer to generate the final prediction result.

[0098] A second aspect of the present disclosure provides a system for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning;

[0099] A multi-view learning-based anomaly detection system for real-time radio field strength 3D waterfall diagram, comprising:

[0100] The data acquisition and processing module is configured to: collect spectrum, field strength and time parameters in real time through radio monitoring equipment; the collected data is preliminarily processed and cleaned to remove noise and abnormal data, and then automatically build a 3D waterfall flow chart based on the three-dimensional information of field strength-time-frequency; classify, label and analyze the large amount of collected monitoring data to build a data set for different abnormal signals; including: normal data, malicious electromagnetic interference data, unauthorized signal occurrence data and authorized signal illegal use data;

[0101] The anomaly detection model building module is configured to: build a deep learning anomaly detection model and perform training;

[0102] The anomaly detection module is configured to: implement 3D waterfall chart anomaly detection through the trained deep learning anomaly detection model.

[0103] In order to enhance the readability and real-time nature of monitoring data, the system also integrates a real-time communication module based on WebSocket. This module ensures that the collected data and the generated 3D waterfall flow chart can be displayed in real time on the front-end visualization screen.

[0104] The third aspect of the present disclosure provides a medium on which a program is stored. When the program is executed by a processor, the steps in the method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning as described in the first aspect of the present disclosure are implemented.

[0105] The fourth aspect of the present disclosure provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for anomaly detection in a real-time radio field strength 3D waterfall diagram based on multi-view learning as described in the first aspect of the present disclosure are implemented.

[0106] The beneficial effects of the present invention are:

[0107] 1. Improve anomaly detection accuracy: The present invention adopts a 3D waterfall chart anomaly detection method based on multi-view learning, which can make full use of multi-dimensional information such as the timing, spectrum and intensity of the signal. Through multi-view fusion learning, the model can capture the key features of the signal from different perspectives and improve the detection accuracy of complex abnormal signals. Compared with the traditional detection method based on a single view, it can better identify abnormal signals with multiple changing characteristics, especially in the monitoring of interference signals and illegal signals.

[0108] 2. Reduce manual intervention and improve the level of automation: Traditional spectrum anomaly detection relies on manual experience, is easily affected by subjective factors, and has low efficiency. The present invention greatly reduces the need for manual intervention through automated machine learning methods. Using the bidirectional long short-term memory network (Bi-LSTM) and cross-view attention mechanism, the system can autonomously learn the temporal and spatial characteristics of the signal, detect and mark abnormal signals in a timely manner, and significantly improve monitoring efficiency and accuracy.

[0109] 3. Multi-view information fusion to enhance robustness: Multi-view learning can mine key information from data from different perspectives and levels, and fuse data of different dimensions (such as time, frequency, field strength, etc.), effectively improving the ability to perceive signal changes. By introducing a cross-view attention fusion mechanism, the present invention can dynamically adjust the weight of each view to ensure that key signal features are fully learned, thereby enhancing the robustness of the anomaly detection model.

[0110] 4. Improve system adaptability and handle complex signal patterns: Signal fluctuations in the radio environment are complex and changeable, and traditional methods are prone to misjudgment or missed judgment when processing these complex signals. The multi-view learning method of the present invention, combined with the time series modeling capability of Bi-LSTM, enables the model to flexibly adapt to different signal patterns, especially when processing high-noise, rapidly changing or irregularly fluctuating signals, and can provide more accurate detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0112] Figure 1 A schematic diagram of using the LSTM model to extract features from the time-field intensity view and the time-frequency view;

[0113] Figure 2 Schematic diagram of view-specific subspace and shared subspace in multi-view feature learning;

[0114] Figure 3 Schematic diagram of the cross-view attention autoencoder fusion model;

[0115] Figure 4 A 3D waterfall diagram of time-frequency-field strength obtained for real-time monitoring tasks of radio monitoring equipment;

[0116] Figure 5 It is a two-dimensional broken line diagram corresponding to the frequency-field strength view. DETAILED DESCRIPTION

[0117] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0118] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0119] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0120] In the absence of conflict, the embodiments in the present disclosure and the features in the embodiments may be combined with each other.

[0121] Example 1

[0122] A method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning, comprising:

[0123] Data collection and processing: The spectrum, field strength and time parameters are collected in real time through radio monitoring equipment. The collected data is preliminarily processed and cleaned to remove noise and abnormal data. Then, a 3D waterfall flow chart is automatically constructed based on the three-dimensional information of field strength-time-frequency.

[0124] In the monitoring tasks of radio monitoring equipment, each task sets a frequency range and a corresponding scanning step. When scanning the frequency range, the equipment obtains data in "frames", and each frame represents a complete scanning process. In actual operation, the time required to complete a frame scan may vary due to factors such as signal interference and transmission interference. In order to achieve real-time monitoring and display, this system uses the start time of the frame as the time axis, adds the data of each frame to the queue according to the first-in-first-out (FIFO) principle, and constructs a 3D waterfall flow chart. The 3D waterfall flow chart is automatically reconstructed at regular intervals, adding the data of the latest frame and replacing the data of the earliest frame, thereby dynamically displaying the changing relationship between frequency, time and field strength. This design can efficiently reflect the dynamic changes of the spectrum and realize continuous and real-time radio monitoring visualization.

[0125] The 3D waterfall flow chart can intuitively display the changes of radio signals in different time periods and frequency ranges. By comparing the field strength distribution in multiple time periods, monitoring personnel can quickly identify potential abnormal signals or interference sources.

[0126] All monitored data will be stored in real time in the back-end database. The database not only records the original monitoring data, but also contains the intermediate data and final display results of the 3D waterfall flow chart. The system supports the query and retrieval of historical data, which is convenient for subsequent abnormal analysis and trend monitoring.

[0127] During the monitoring task, the WebSocket module will continue to push the latest monitoring data to the front end. The user does not need to manually refresh the page. The data changes and analysis results will be automatically synchronized to the large screen. The large screen display adopts a highly customizable view layout. Users can adjust the displayed spectrum range, time axis span, field strength range and other parameters according to their needs to ensure that the information can be displayed in the most intuitive way. In addition, the system supports multiple visualization modes. Users can choose to display a static 3D waterfall flow chart or a dynamic mode to observe the real-time changes of the signal over time.

[0128] On the basis of the aforementioned system development module, based on the stored data, further combined with the analysis and processing of professionals, by classifying, labeling and analyzing the large amount of collected monitoring data, the system can classify abnormal situations according to different radio signal characteristics, and construct data sets for different abnormal signals; including: normal data, malicious electromagnetic interference data, unauthorized signal occurrence data and authorized signal illegal use data; this data set not only records normal radio signal samples, but also includes abnormal situations that may occur in actual applications. The final data set is classified into the following categories:

[0129] Normal data: Normal radio signal operation without interference or unauthorized operation. This data is used to build a baseline model to help identify abnormal signals.

[0130] Malicious electromagnetic interference data: Contains electromagnetic interference signals that may be caused by external devices or environmental factors. Such signals often manifest as a short-term sudden increase in intensity or abnormal frequency usage, which may cause communication failures. Professional analysis helps the model identify potential malicious interference sources by extracting the characteristics of interference signals.

[0131] Unauthorized signal occurrence data: contains unauthorized spectrum signals. Such signals may appear in unallocated frequency bands, which may interfere with or cover legitimate signals in serious cases. By analyzing the spectrum distribution and field strength changes of these signals, illegal occupation of frequency bands can be effectively detected.

[0132] Authorized signal illegal use data: records signal samples of legal devices that do not comply with spectrum allocation rules, such as exceeding the authorized frequency range or using excessive power. The system can identify illegal use behaviors by comparing with normal signal samples and issue warnings in a timely manner.

[0133] A deep learning anomaly detection model is constructed and trained; after the above steps are completed, based on the deep learning model created by the present invention, the system can classify different types of radio signals. The deep learning model is trained on a large number of annotated data sets, including different categories such as normal data, malicious electromagnetic interference, unauthorized signals, and illegal use of authorized signals. The model extracts features from these classified data and constructs a multi-view learning framework, combining information from different views such as time-field strength, time-frequency, and field strength-frequency to comprehensively analyze the signal.

[0134] 3D waterfall chart anomaly detection is achieved through the trained deep learning anomaly detection model. By integrating the feature representations of different views, abnormal signals can be accurately identified and potential abnormal situations can be alarmed. The model has the ability of adaptive learning and can continuously optimize the detection performance through the continuous collection of data sets in practical applications to ensure the accuracy and real-time performance of radio signal monitoring.

[0135] Example 2

[0136] The difference between the method for detecting anomalies in a real-time radio field strength 3D waterfall chart based on multi-view learning described in Example 1 is that:

[0137] The 3D image is decomposed into different 2D views, and two representation methods are used to handle the heterogeneity of views: the first is a common representation, which uses a unified encoder to learn and project all views into a shared subspace to achieve view distribution alignment and reduce information redundancy. The second is a private representation, which learns the subspace specific to each view and extracts unique view features.

[0138] For the timeline-related view, LSTM is used to extract time series features; for the field strength-frequency view, multi-layer perceptron MLP is used for feature extraction.

[0139] For private spaces, an attention mechanism is introduced to obtain the final display of each unique space after removing redundancy.

[0140] Construct three views, the 3D map, namely the 3D waterfall map, includes rich information in three dimensions: field strength, time, and frequency. In order to improve processing efficiency and feature extraction accuracy, the present invention decomposes the 3D map into multiple 2D views, and characterizes and analyzes the radio signal from different angles.

[0141] Build timeline-related views to show the changes of different data on the timeline;

[0142] Time axis related views include time-field intensity view and time-frequency view;

[0143] The time-field strength view is used to display the changes in the field strength of the radio signal on the time axis; by recording the field strength value at each moment, it can effectively capture short-term signal fluctuations and help identify sudden interference signals.

[0144] The time-frequency view is used to display spectrum distribution in different time periods and monitor changes in frequency usage. Similar to the time-intensity view, this view also contains time series information, so both use the same technology, namely LSTM, to extract time series features.

[0145] Use the LSTM model to extract features from the time-field intensity view and the time-frequency view: Figure 1 As shown, including:

[0146] In order to overcome the common gradient vanishing and gradient exploding problems of traditional RNN when dealing with long-term dependency problems, the long short-term memory (LSTM) network came into being. By introducing three gate mechanisms, LSTM can effectively selectively memorize and forget information in time series, ensuring that the model can flexibly process data with long time spans.

[0147] In order to overcome the gradient vanishing and gradient exploding problems encountered by RNN in long-term dependency problems, the LSTM model introduces three gate mechanisms: forget gate, input gate, and output gate to control the storage and update process of information in the time series. The core of the LSTM model lies in its cell state (C t ) and the three gates work together to enable the LSTM model to selectively retain or discard information.

[0148] First, the forget gate is used to control the part of the cell state that needs to be forgotten. The calculation formula is:

[0149] Γ f (t) = σ(W f ·[a t-1 ,x t ]+b f )

[0150] Among them, Γ f (t) represents the value of the forget gate output, which controls the cell state C at the previous moment t -1 degree of retention; a t -1 is the hidden state of the previous time step; x t is the current input; W f is the weight matrix; b f is the bias term; σ is the Sigmoid activation function, which ensures that the output value is between U and 1;

[0151] Next, the input gate controls the extent to which new information is written to the cell state. The computational process is divided into three steps; including:

[0152] The first step is to calculate the value of the input gate and decide which information to update at the current moment. The calculation formula is:

[0153] Γ u (t) = σ(W u ·[a t-1 ,x t ]+b u )

[0154] The second step is to generate candidate cell states:

[0155]

[0156] The third step is to update the cell state by combining the output of the forget gate and the input gate to retain and update the cell state:

[0157]

[0158] Among them, Γ u (t) represents the value of the input gate output; it controls the proportion of new information added to the cell state: is the candidate cell state, which transforms the input through the hyperbolic tangent function tanh to ensure that the output is between -1 and 1; W u and W c is the weight matrix of the input gate and candidate state, b u and b c is the corresponding bias term; C t Represents the cell state at the current moment and is one of the core variables used by the LSTM model to store long-term information. At each time step, LSTM updates this cell state in order to retain or discard information based on the current input and previous state. t-1 Represents the cell state at the previous time step; contains the long-term memory accumulated up to the previous time step. At the current moment, LSTM decides how to use this state through the forget gate.

[0159] Finally, the output gate determines the output of the LSTM at the current moment. The output gate generates output based on the current cell state and input information:

[0160] Γ o (t) = σ(W o ·[a t-1 ,x t ]+b o )

[0161] The output values ​​are:

[0162] FT t =Γ o (t)·tanh(C t )

[0163] Among them, Γ o (t) is the value of the output gate, which controls the output strength; FT t is the hidden state at the current moment, which is also the output of the LSTM model; tanh(C t ) represents the nonlinear transformation of the current cell state; through the synergistic effect of these three gate mechanisms, the LSTM model can effectively control the flow of information, avoid the common gradient problems in traditional RNNs, and achieve effective modeling of long time series data.

[0164] Based on the time series processing mechanism of the LSTM model, the feature representation FT of the time-field intensity view and the time-frequency view is generated by extracting the features of the time-field intensity view and the time-frequency view. t , the feature representation of the time-field intensity view is defined as FT1, and the feature representation of the time-frequency view is defined as FT2.

[0165] Use a multi-layer perceptron to extract features from the field strength-frequency view; including:

[0166] The field strength-frequency view shows the signal strength at different frequencies. The multi-layer perceptron (MLP) can effectively capture the complex nonlinear relationship between field strength and frequency and generate more expressive feature representations.

[0167] The multi-layer perceptron consists of several fully connected layers, each of which includes a weight matrix and a bias term; its feature extraction process can be described by the following formula:

[0168] First, assume that the input data of the field strength-frequency view is a vector S = [s1, s2, ..., s m ...,s M ], where s m Represents the field strength value corresponding to the mth frequency, and M is the total number of frequencies; the multi-layer perceptron MLP gradually extracts the high-dimensional features of the input data through multi-layer nonlinear mapping. For the first fully connected layer, the calculation formula for the feature representation is:

[0169] h (1) =σ(W (1) ·S+b (1) )

[0170] Among them, W (1) is the weight matrix of the first fully connected layer, b (1) is the bias vector, and σ is the activation function (usually ReLU) that maps the input to the hidden layer feature h (1) middle;

[0171] Next, enter the second fully connected layer, and the calculation formula is:

[0172] h(2) =σ(W (2) ·h (1) +b (2) )

[0173] Here, W (2) and b (2) are the weight matrix and bias vector of the second fully connected layer, and the output feature h (2) further mapped to higher dimensions;

[0174] Finally, the output feature representation is:

[0175] FS=W (3) ·h (2) +b (3)

[0176] Among them, W (3) and b (3) is the weight matrix and bias term of the output layer, and FS is the feature representation of the extracted field strength-frequency view.

[0177] Through the above-mentioned multi-layer mapping and nonlinear transformation, MLP can gradually extract important features in the field strength-frequency view, capture the complex frequency distribution patterns in the signal, and effectively improve the accuracy and expression ability of feature extraction.

[0178] like Figure 2 As shown in the figure, the specific subspace and shared subspace process the shared information between views and the specific information of each view respectively through common representation and private representation; given a view m and its extracted feature vector u m , using a shared encoder C(;θ c ) maps it to a view-invariant shared subspace, while using a view-specific private encoder Embed a specific view into a unique subspace;

[0179] The common representation handles the shared information between views and uses a shared encoder C(·; θ c ), shared encoder C(·;θ c ) parameters θ are shared between all views c ; The formula is as follows:

[0180] C m =C(u m θ c )

[0181] Among them, C m represents the feature representation of view m in the shared subspace, u m is the input feature of view m, θ c are the parameters of the shared encoder;

[0182] The private representation is responsible for capturing the specific information of each view; through the view-specific private encoder The features of each view are embedded into its own specific subspace; the formula is as follows:

[0183]

[0184] in, represents the feature representation of view m in the unique subspace, are the private encoder parameters of view m, and different views have their own independent parameters;

[0185] The three views generate the following corresponding feature representations:

[0186] For visual Figure 1 That is, the time-field view: the shared subspace output is C1, and the private subspace output is

[0187] For visual Figure 2 That is, the time-frequency view: the shared subspace output is C2, and the private subspace output is

[0188] For visual Figure 3 That is, the field strength-frequency view: the shared subspace output is C3, and the private subspace output is

[0189] Adding similarity constraints to the common representation, enhanced by cosine similarity loss, aligns common cross-view features in the view-shared subspace and reduces the differences between them;

[0190] Cosine Similarity Loss Function The definition is as follows:

[0191]

[0192] Here, the loss The closer it is to 0, the more similar the two vectors are.

[0193] This loss is finally added to the classification loss and used as the final loss function of the model to guide model training.

[0194] When extracting private representations, the shared space and vector attention mechanism is used to further optimize the feature extraction process to extract redundant features of different views. For each view:

[0195] First, the weight of each modal feature is calculated through the attention mechanism. The formula is as follows

[0196]

[0197] Among them, α m is the attention weight of view m, is the feature representation of view m, is the attention vector, is the weight matrix, is the bias term; in order to better extract redundant features of different views, all views share parameters W4 and b4.

[0198] Use attention weights to calculate redundant features between views:

[0199]

[0200] Among them, E s Represents redundant features of all views;

[0201] Then, the feature representation in the unique subspace is obtained by subtracting the redundant features from the original features of each view, as follows:

[0202]

[0203] P m It refers to the final unique features obtained from the corresponding view m, that is, the final display of each unique space after eliminating redundancy.

[0204] The deep learning anomaly detection model is a cross-view attention autoencoder fusion model, which realizes information interaction and fusion between different views, and enhances the accuracy and efficiency of classification and anomaly detection. By using the deep learning anomaly detection model, anomaly detection and real-time alarm are realized for unattended radio monitoring tasks; the cross-view attention autoencoder fusion model is used to fuse the common representation and private representation to form the final feature; including:

[0205] In different views, the importance of each feature in the final prediction varies. If these features are simply spliced ​​together, the interaction between the modalities is easily ignored, and redundant information may be introduced, resulting in suboptimal problems. To this end, the present invention proposes a cross-view attention autoencoder fusion module to fully realize information interaction and knowledge exchange between different feature representations. In this module, the shared representations of different views are no longer merged, but the attention calculation is performed between the private representation and all the shared representations to generate a new private representation, which is finally fused through the autoencoder for the final prediction.

[0206] Figure 3 The figure is a schematic diagram of the specific architecture of the cross-view attention autoencoder fusion model. The architecture design of the cross-view attention autoencoder fusion model is mainly divided into two stages: the cross-view interaction stage and the encoder fusion stage, so as to effectively fuse the data representations of different views, eliminate redundant information, and improve the accuracy of prediction.

[0207] Cross-view interaction stage: Generate an independent private representation for each view, and also calculate a shared global representation; in the private representation of each view, interact with the shared representation through the attention mechanism to update the representation of the view so that it can obtain relevant information from other views; this interactive calculation method not only retains the characteristics of each view, but also captures the correlation between views, ensuring that the potential connections between different views are not ignored when fusing them.

[0208] Encoder fusion stage: The updated view representations in the cross-view interaction stage are input into the autoencoder; the autoencoder is responsible for further fusing the updated private representations to generate a more concise and representative global representation. The autoencoder includes an encoder and a decoder. The encoder compresses the private representations of different views into hidden layer representations to extract key information; the decoder reconstructs the hidden layer representation to verify and optimize the validity of the information; through the compression and reconstruction of the autoencoder, the model finally obtains a global representation that eliminates redundancy.

[0209] Final fusion and prediction: After the encoder fusion stage, multiple views are output and concatenated to form the final feature vector; the feature vector is input to the classification layer to generate the final prediction result. This design ensures that the model will not accumulate redundant information due to simple concatenation while fusing information from different views, making the final feature representation more accurate and efficient.

[0210] The specific steps are as follows:

[0211] Cross-view interaction stage;

[0212] The private representation of each view is calculated; three views p1, p2, p3 and their shared representations C1, C2, C3 have been obtained; the private representation of each view P m will be shared with all the m Perform attention calculation to generate updated private representation; the formula is as follows:

[0213] P m ′=Attention(P m ,[C1,C2,C3])

[0214] Among them, P' m is the updated private representation, Attention is the standard self-attention mechanism, and the calculation method is:

[0215]

[0216] Among them, Q, K, and V are respectively combined with different weight matrices W through the input matrix Q , WK , W V Linear transformation yields, d k It is the dimension of Key, used for normalization;

[0217] Through this cross-view interactive computation, the private representation of each view can be updated according to the shared representation, capturing the correlation between views.

[0218] Encoder fusion stage:

[0219] In the encoder fusion stage, the updated private representation is finally fused using the autoencoder; each view P' m As input, the autoencoder generates the final vector representation:

[0220] Z m =Autoencoder(P m ′)

[0221] The structure of the autoencoder is a symmetrical neural network, consisting of an encoder and a decoder. The specific formula is as follows:

[0222] Encoder:

[0223] h m =σ(W e ·P m ′+b e )

[0224] Among them, W e and b e are the weight and bias of the encoder respectively, σ is the activation function;

[0225] Decoder:

[0226]

[0227] Among them, W d and b d are the weight and bias of the decoder respectively, It is the reconstructed representation; through the role of the autoencoder, the private representations of different views can be fused into a global representation and information redundancy can be eliminated.

[0228] Finally, the outputs of the three views Z1, Z2, and Z3 are concatenated to form the final feature vector for prediction:

[0229] Z final =[Z1,Z2,Z3]

[0230] This final feature vector passes through the subsequent classification layer to generate the final prediction result.

[0231] Figure 4 A 3D waterfall diagram of time-frequency-field strength obtained for real-time monitoring tasks of radio monitoring equipment; Figure 4 In the figure, the x-axis is time, the y-axis is frequency, and the z-axis is field strength; Figure 5 It is a two-dimensional line graph corresponding to the frequency-field strength view; Figure 5 In the figure, the x-axis is frequency and the y-axis is field strength.

[0232] Example 3

[0233] A multi-view learning-based anomaly detection system for real-time radio field strength 3D waterfall diagram, comprising:

[0234] The data acquisition and processing module is configured to: collect spectrum, field strength and time parameters in real time through radio monitoring equipment; the collected data is preliminarily processed and cleaned to remove noise and abnormal data, and then automatically build a 3D waterfall flow chart based on the three-dimensional information of field strength-time-frequency; classify, label and analyze the large amount of collected monitoring data to build a data set for different abnormal signals; including: normal data, malicious electromagnetic interference data, unauthorized signal occurrence data and authorized signal illegal use data;

[0235] The anomaly detection model building module is configured to: build a deep learning anomaly detection model and perform training;

[0236] The anomaly detection module is configured to: implement 3D waterfall chart anomaly detection through the trained deep learning anomaly detection model.

[0237] In order to enhance the readability and real-time nature of monitoring data, the system also integrates a real-time communication module based on WebSocket. This module ensures that the collected data and the generated 3D waterfall flow chart can be displayed in real time on the front-end visualization screen.

[0238] Example 4

[0239] A medium having a program stored thereon, which, when executed by a processor, implements the steps in a method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning as described in Embodiment 1 or 2.

[0240] Example 5

[0241] An electronic device comprises a memory, a processor and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in a method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning as described in Example 1 or 2 are implemented.

Claims

1. A method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning, characterized in that: include: Data collection and processing: Real-time collection of spectrum, field strength and time parameters through radio monitoring equipment; The collected data is initially processed and cleaned to remove noise and abnormal data. Then, a 3D waterfall flow chart is automatically constructed based on the three-dimensional information of field strength-time-frequency. Classify, annotate and analyze the large amount of monitoring data collected to construct data sets for different abnormal signals; including: normal data, malicious electromagnetic interference data, unauthorized signal appearance data and authorized signal illegal use data; Build a deep learning anomaly detection model and train it; 3D waterfall chart anomaly detection is achieved through the trained deep learning anomaly detection model.

2. According to the multi-view learning-based anomaly detection method for real-time radio field strength 3D waterfall diagram according to claim 1, it is characterized in that: The 3D image is decomposed into different 2D views, and two representations are used to handle the heterogeneity of views: the first is a common representation, which uses a unified encoder to learn and project all views into a shared subspace; the second is a private representation, which learns a subspace specific to each view and extracts unique view features; For the time axis related view, LSTM is used to extract the time series features; for the field strength-frequency view, multi-layer perceptron MLP is used for feature extraction; For private spaces, an attention mechanism is introduced to obtain the final display of each unique space after removing redundancy; The 3D graph, i.e. the 3D waterfall flow graph, includes information in three dimensions: field strength, time, and frequency; Build timeline-related views to show the changes of different data on the timeline; Time axis related views include time-field intensity view and time-frequency view; The time-field strength view is used to show the changes in the field strength of the radio signal on the time axis; The time-frequency view is used to display spectrum distribution in different time periods and monitor changes in frequency usage.

3. According to the multi-view learning-based anomaly detection method for real-time radio field strength 3D waterfall diagram according to claim 1, it is characterized in that: Use the LSTM model to extract features from the time-field intensity view and the time-frequency view: including: First, the forget gate is used to control the part of the cell state that needs to be forgotten. The calculation formula is: C f (t)=σ(W f ·[a t-1 ,x t ]+b f ) Among them, Γ f (t) represents the value of the forget gate output, which controls the cell state C at the previous moment t -1 degree of retention; a t -1 is the hidden state of the previous time step; x t is the current input; W f is the weight matrix; b f is the bias term; σ is the Sigmoid activation function, which ensures that the output value is between U and 1; Next, the input gate controls the extent to which new information is written to the cell state. The computational process is divided into three steps; including: The first step is to calculate the value of the input gate and decide which information to update at the current moment. The calculation formula is: C u (t)=σ(W u ·[a t-1 ,x t ]+b u ) The second step is to generate candidate cell states: The third step is to update the cell state by combining the output of the forget gate and the input gate to retain and update the cell state: Among them, Γ u (t) represents the value of the input gate output; it controls the proportion of new information added to the cell state: is the candidate cell state, which transforms the input through the hyperbolic tangent function tanh to ensure that the output is between -1 and 1; W u and W c is the weight matrix of the input gate and candidate state, b u and b c is the corresponding bias term; C t Indicates the cell state at the current moment, C t-1 represents the cell state at the previous time step; Finally, the output gate generates the output based on the current cell state and the input information: C o (t)=σ(W o ·[a t-1 ,x t ]+b o ) The output values ​​are: FT t =Γ o (t)·tanh(C t ) Among them, Γ o (t) is the value of the output gate, which controls the output strength; FT t is the hidden state at the current moment, which is also the output of the LSTM model; tanh(C t ) represents the nonlinear transformation of the current cell state; Based on the time series processing mechanism of the LSTM model, the feature representation FT of the time-field intensity view and the time-frequency view is generated by extracting the features of the time-field intensity view and the time-frequency view. t , the feature representation of the time-field intensity view is defined as FT1, and the feature representation of the time-frequency view is defined as FT2.

4. According to the multi-view learning-based anomaly detection method for real-time radio field strength 3D waterfall chart, the method is characterized in that: Use a multi-layer perceptron to extract features from the field strength-frequency view; including: The multi-layer perceptron includes several fully connected layers, each of which includes a weight matrix and a bias term; First, assume that the input data of the field strength-frequency view is a vector S = [s1, s2, ..., s , ...,s M ], where s m Represents the field strength value corresponding to the mth frequency, and M is the total number of frequencies; the multi-layer perceptron MLP gradually extracts the high-dimensional features of the input data through multi-layer nonlinear mapping. For the first fully connected layer, the calculation formula for the feature representation is: h (1) =σ(W (1) ·S+b (1) ) Among them, W (1) is the weight matrix of the first fully connected layer, b (1) is the bias vector, σ is the activation function, which maps the input to the hidden layer feature h (1) middle; Next, enter the second fully connected layer, and the calculation formula is: h (2) =σ(W (2) ·h (1) +b (2) ) Here, W (2) and b (2) are the weight matrix and bias vector of the second fully connected layer, and the output feature h (2) further mapped to higher dimensions; Finally, the output feature representation is: FS=W (3) ·h (2) +b (3) Among them, W (3) and b (3) is the weight matrix and bias term of the output layer, and FS is the feature representation of the extracted field strength-frequency view.

5. According to the multi-view learning-based anomaly detection method for real-time radio field strength 3D waterfall diagram, the method is characterized in that: The specific subspace and shared subspace process the shared information between views and the specific information of each view respectively through common representation and private representation; given a view m and its extracted feature vector u m , using a shared encoder C(;θ c ) maps it to a view-invariant shared subspace, while using a view-specific private encoder Embed a specific view into a unique subspace; The common representation handles the shared information between views and uses a shared encoder C(·; θ c ), shared encoder C(·;θ c ) parameters θ are shared between all views c ; The formula is as follows: C m =C(u m ;θ c ) Among them, C m represents the feature representation of view m in the shared subspace, u m is the input feature of view m, θ c are the parameters of the shared encoder; The private representation is responsible for capturing the specific information of each view; through the view-specific private encoder The features of each view are embedded into its own specific subspace; the formula is as follows: in, represents the feature representation of view m in the unique subspace, are the private encoder parameters of view m, and different views have their own independent parameters; The three views generate the following corresponding feature representations: For view 1, i.e., the time-field intensity view: the shared subspace output is C1, and the private subspace output is For view 2, i.e., the time-frequency view: the shared subspace output is C2, and the private subspace output is For view 3, i.e., the field strength-frequency view: the shared subspace output is C , , the private subspace output is Further preferably, similarity constraints are added to the common representation, which is enhanced by cosine similarity loss to align common cross-view features in the view-shared subspace and reduce the differences between them; Cosine Similarity Loss Function The definition is as follows:

6. The method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning according to claim 1, characterized in that: When extracting private representations, we use the shared space and vector attention mechanism to further optimize the feature extraction process. For each view: First, the weight of each modal feature is calculated through the attention mechanism. The formula is as follows Among them, α m is the attention weight of view m, is the feature representation of view m, is the attention vector, is the weight matrix, is the bias term; all views share parameters W4 and b4; Use attention weights to calculate redundant features between views: Among them, E s Represents redundant features of all views; Then, the feature representation in the unique subspace is obtained by subtracting the redundant features from the original features of each view, as follows: P m It refers to the final unique features obtained from the corresponding view m, that is, the final display of each unique space after eliminating redundancy.

7. The method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning according to claim 1, characterized in that: The deep learning anomaly detection model is a cross-view attention autoencoder fusion model, which uses the cross-view attention autoencoder fusion model to fuse the common representation and the private representation to form the final feature; including: Cross-view interaction stage; The private representation of each view is calculated; three views P1, P2, P3 and their shared representations C1, C2, C3 have been obtained; the private representation of each view P m will be shared with all the m Perform attention calculation to generate updated private representation; the formula is as follows: P m ′=Attention(P m ,[C1,C2,C3]) Among them, P′ m is the updated private representation, Attention is the standard self-attention mechanism, and the calculation method is: Among them, Q, K, and V are respectively combined with different weight matrices W through the input matrix Q , W K , W V Linear transformation yields, d k It is the dimension of Key, used for normalization; Encoder fusion stage: In the encoder fusion stage, the updated private representation is finally fused using the autoencoder; each view P′ m As input, the autoencoder generates the final vector representation: Z m =Autoencoder(P m ′) Finally, the outputs of the three views Z1, Z2, and Z3 are concatenated to form the final feature vector for prediction: WITH final =[Z1,Z2,Z3] This final feature vector passes through the subsequent classification layer to generate the final prediction result.

8. A multi-view learning-based anomaly detection system for real-time radio field strength 3D waterfall diagram, characterized in that: include: The data acquisition and processing module is configured to: collect spectrum, field strength and time parameters in real time through radio monitoring equipment; The collected data is initially processed and cleaned to remove noise and abnormal data. Then, a 3D waterfall flow chart is automatically constructed based on the three-dimensional information of field strength-time-frequency. Classify, annotate and analyze the large amount of monitoring data collected to construct data sets for different abnormal signals; including: normal data, malicious electromagnetic interference data, unauthorized signal appearance data and authorized signal illegal use data; The anomaly detection model building module is configured to: build a deep learning anomaly detection model and perform training; The anomaly detection module is configured to: implement 3D waterfall chart anomaly detection through the trained deep learning anomaly detection model.

9. A medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in a method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning as described in any one of claims 1 to 7 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the method for detecting anomalies in a real-time radio field strength 3D waterfall diagram based on multi-view learning as described in any one of claims 1-7 are implemented.

Images