A communication interference identification and prediction method based on CNN-LSTM spatio-temporal fusion

Abstract

The application discloses a communication interference recognition and prediction method based on CNN-LSTM space-time fusion, and belongs to the technical field of communication. The application collects original radio frequency data of a communication system, constructs 8-channel full-band PSD images through fast Fourier transform and carries out normalization pretreatment, extracts interference space features through a ResNet18 backbone network, combines an LSTM to build a time sequence model and introduces feature bypass residual connection, adopts a weighted joint loss function to complete double-task end-to-end training, synchronously realizes current recognition and future K-step prediction of three types of interference, namely, no interference, fixed-frequency interference and sweep-frequency interference, generates a frequency hopping sequence based on time attenuation channel risk assessment, forms a closed-loop anti-interference process, and can be adaptively embedded in low-latency deployment. The application effectively breaks through the limitation that traditional methods can only be used for static recognition, solves the pain point that anti-interference decision lags behind interference changes, and improves interference recognition and prediction accuracy in a complex electromagnetic environment.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically a communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion. Background Technology

[0002] Modern wireless communication systems play a crucial role in both civilian and military applications. Especially in missions involving communication in complex electromagnetic environments and high-intensity electronic warfare, the stability and anti-jamming capabilities of wireless communication links directly impact mission success. In complex electromagnetic environments, interference signals exhibit diversity, time-varying characteristics, and suddenness. Common interference types primarily include suppression jamming and deceptive jamming. Since different interference types require differentiated anti-jamming strategies, real-time identification and trend prediction of interference types have become the prerequisite and core for cognitive radio systems to achieve intelligent anti-jamming.

[0003] Traditional interference identification methods mainly rely on threshold decisions, maximum likelihood estimation, and shallow machine learning algorithms. These methods heavily depend on manual feature extraction and suffer from poor generalization ability and insufficient robustness in complex and variable electromagnetic environments, making it difficult to meet the millisecond-level response speed and identification accuracy requirements of modern electronic warfare. Compared to traditional methods, deep learning is better at learning higher-level features and has stronger generalization ability.

[0004] Application number CN2025100670196 proposes a communication interference identification method based on three-dimensional time-frequency graphs and deep transfer learning. By using a multi-modal feature fusion mechanism of LSTM + attention mechanism, the accuracy and robustness of interference identification are improved, which is especially suitable for small sample scenarios.

[0005] Application No. CN2024103719937: A method for identifying communication interference signals based on residual networks. This paper proposes an interference identification method for low signal-to-noise ratio and small sample scenarios. By combining ACGAN data augmentation and residual networks, the method effectively improves the identification capability of interference signals in complex electromagnetic environments.

[0006] However, such methods still have significant limitations:

[0007] First, it can only perform static identification at the current moment, without performing time-series modeling of historical signal sequences, and cannot predict the future evolution trend of interference types;

[0008] Secondly, in scenarios where interference patterns change rapidly, due to limitations in system data acquisition and transmission delays, anti-interference decisions always lag behind the changes in interference, resulting in a passive situation of "slow to fight fast," which makes it difficult to meet the needs of rapid response in practical applications. Summary of the Invention

[0009] To address the above problems, this invention provides a communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion. It aims to provide efficient and intelligent end-to-end anti-interference technology support for cognitive radio systems, ensure the stable operation of wireless communication links, break through the limitations of traditional methods that can only identify statically, solve the pain point that anti-interference decision-making lags behind interference changes, and improve the accuracy of interference identification and prediction in complex electromagnetic environments.

[0010] To address the time-varying and sudden nature of interference signals in complex electromagnetic environments, existing methods often focus on single-frame, single-moment interference type identification, lacking modeling of historical time-series information and making it difficult to reliably predict the future evolution trend of interference types. This invention proposes a communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion. First, a real-time signal processing module performs a Fast Fourier Transform on the original radio frequency signal, calculates the power spectral density of eight channels, and stitches them into a full-band PSD image. Then, a CNN module extracts frequency domain spatial features from a single-frame image, while an LSTM module performs temporal modeling of the feature sequences from multiple consecutive frames, thereby achieving the identification of the current interference type and the prediction of interference trends at several future moments. This technology fully utilizes the rich discriminative features of the power spectral image and effectively overcomes the bottleneck of untimely response during anti-interference through deep spatiotemporal feature fusion and a dual-task learning mechanism. The overall system design framework is as follows: Figure 1 As shown.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0012] A communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion is proposed. This method sequentially executes the steps of data acquisition and preprocessing, spatial feature extraction, temporal modeling and multi-step prediction, dual-task joint training, result output, and closed-loop application. It achieves current identification and future K-step prediction of communication signal interference types, with the interference types fixed as three categories: no interference, fixed-frequency interference, and frequency-sweeping interference. Specifically, it includes:

[0013] S1: Acquisition of raw radio frequency data and construction of PSD image. Raw radio frequency data of communication system is collected in each observation period. Power spectral density of 8 channels is calculated by fast Fourier transform. The PSDs of 8 channels are stitched together in frequency order to generate a full-band PSD image. The full-band PSD image is then horizontally cut into 8 sub-images in 8 equal parts.

[0014] S2: Image normalization preprocessing, which performs uniform scale transformation and normalization on the full-band PSD image or its segmented sub-images so that the processed image meets the input requirements of CNN inference;

[0015] S3: CNN spatial feature extraction and current time identification. Construct a CNN feature extraction network with ResNet18 as the backbone to extract cross-frequency band and cross-channel spatial discrimination features for each frame of PSD image and output the interference type discrimination result at the current time.

[0016] S4: Construct historical sequence samples, using a sliding window of length N to cache the CNN feature vectors of the most recent N frames, forming a temporal input sequence;

[0017] S5: LSTM Temporal Modeling and Multi-Step Prediction. The temporal input sequence is input into the LSTM network for temporal modeling, capturing the temporal evolution of interference types and outputting the interference type prediction results for the next K time steps. At the same time, a feature bypass residual connection mechanism is introduced to prevent the decay of spatial feature information at the current time step.

[0018] S6: Dual-task output and joint loss training. Construct a dual-task learning structure that includes task A for identifying interference at the current time and task B for predicting interference in the next K steps. Use a weighted joint loss function to train the model end-to-end so that the model shares the underlying spatiotemporal features.

[0019] S7: Results Output and Application. Taking the current identification results and the prediction results of the next K steps as input, a channel risk assessment function based on time decay is constructed to calculate the comprehensive risk score of each channel. Based on the risk score, candidate communication channels are selected and frequency hopping sequences are generated. The frequency hopping sequences and parameter control information are sent to the cognitive radio module to complete the radio frequency switching and parameter configuration update, forming a closed-loop processing flow from interference perception to execution.

[0020] The beneficial effects of the above technical solution are as follows: it establishes a complete technical framework from raw radio frequency data processing to current and future interference type identification and prediction, and then to the execution of anti-interference strategies. It integrates the spatial feature extraction capability of CNN and the temporal pattern capture capability of LSTM, breaking through the limitation of traditional methods that can only achieve static identification at a single moment. Through dual-task joint training, the model shares the underlying spatiotemporal features, and the feature bypass residual connection mechanism ensures the sensitivity of current identification. Combined with time-decayed channel risk assessment and closed-loop processing, the model output is directly transformed into an executable anti-interference strategy, solving the problem that anti-interference decision-making lags behind interference changes, and providing full-link intelligent anti-interference support for cognitive radio systems.

[0021] As a further improvement to the above scheme, in step S1, the power spectral density of the i-th channel at the discrete frequency f_k is calculated according to the following formula:

[0022] in Let N represent the original radio frequency sampling sequence in the time domain of the i-th channel, where N represents the number of FFT points and k represents the discrete frequency index. ;

[0023] in, This represents a full-band PSD image after horizontal stitching. concat indicates horizontal stitching operation in frequency order.

[0024] The beneficial effects of the above technical solution are as follows: By standardizing the calculation of power spectral density and the full-band PSD image stitching process through precise mathematical formulas, the conversion process of multi-channel frequency domain data can be quantified and reproduced. It can completely preserve the key features of interference signals in terms of spectrum distribution, energy accumulation, and cross-channel correlation, avoid information loss in the data conversion process, and lay a high-quality data foundation for the subsequent extraction of effective spatial discriminative features by CNN networks.

[0025] As a further improvement to the above scheme, in step S2, the scaling transformation is to adjust the PSD image to the fixed input size required by the ResNet18 backbone network of the CNN, and the normalization process is to map the pixel values ​​of the PSD image to the numerical range of 0-1, so that the size and numerical range of the processed image match the input standard of CNN inference.

[0026] The beneficial effects of the above technical solution are that uniform scaling and normalization of images eliminate the interference of image size differences and pixel value range differences on inference, unify the model input specifications, ensure the consistency and accuracy of the feature extraction process, and effectively improve the model's generalization ability in different data scenarios.

[0027] As a further improvement to the above scheme, in step S3, the core operation of CNN is two-dimensional convolution, which is performed according to the following formula:

[0028] ;

[0029] in, For the first Layer convolution kernel, This is the feature map of the previous layer. For the first Layer The bias of each channel;

[0030] In CNNs, residual blocks are used to mitigate the vanishing gradient problem. The residual blocks are implemented according to the following formula: ;

[0031] After multiple convolutions and global pooling, an intermediate feature vector f_t of dimension D is obtained. The fully connected head of the CNN outputs logits at the current time step. After softmax activation, the predicted probabilities of three classes are obtained: no interference, fixed frequency interference, and frequency sweeping interference.

[0032] The beneficial effects of the above technical solution are as follows: the feature extraction process of CNN is standardized and reproducible through explicit mathematical formulas; the design of residual blocks effectively solves the gradient vanishing problem of deep convolutional networks, ensuring the effective extraction of cross-frequency band and cross-channel spatial discriminative features; the output form after softmax activation can accurately give the predicted probability of various types of interference, rather than a single category label, which improves the accuracy of interference type identification at the current moment and the reference value of the results.

[0033] As a further improvement to the above scheme, in step S4, the time-series input sequence is constructed according to the following formula: ;in, For the first Feature representation of a frame PSD image after CNN encoding for to The temporal input sequence is composed of CNN feature vectors at each time step.

[0034] The beneficial effects of the above technical solution are as follows: the sliding window mechanism and standardization formula realize the effective integration of historical CNN spatial features, transform the static spatial features of a single frame into a continuous temporal feature sequence, which fits the input characteristics of the LSTM network, provides high-quality temporal data for the LSTM network to capture the temporal evolution law of interference type, and ensures the effectiveness and rationality of subsequent temporal modeling.

[0035] As a further improvement to the above scheme, in step S5, the gating update of the LSTM in a single time step sequentially executes the calculation of the forget gate, input gate, candidate cell state, cell state update, output gate, and hidden state, with the specific formula as follows:

[0036] Forgotten Gate: ;

[0037] Input Gate: ;

[0038] Candidate cell status: ;

[0039] Cell status update: ;

[0040] Output gate: ;

[0041] Hidden state: ;

[0042] in, This represents the sigmoid activation function. This represents the hyperbolic tangent activation function. Represents element-wise multiplication. This indicates the hidden state in the previous moment. This indicates the cell state at the previous moment. Indicates the current input number. Step CNN features, , This represents the weights and biases of each learnable element;

[0043] The LSTM's timing prediction header connects the last hidden state to the fully connected layer and outputs the prediction result according to the following formula:

[0044] ;

[0045] The feature bypass residual connection mechanism involves directly adding the residual of the CNN features of the latest frame in the sliding window to the current recognition head output features of the LSTM.

[0046] The beneficial effects of the above technical solution are as follows: the precise gating calculation formula standardizes the LSTM temporal modeling process, effectively capturing the temporal evolution patterns such as frequency drift of swept interference and persistent residence of fixed-frequency interference, and achieving accurate prediction of interference types at the next K time points; the design of the temporal prediction head directly outputs prediction results that meet the K×3 dimension requirements, simplifying the subsequent data processing process; the feature bypass residual connection mechanism effectively prevents the information attenuation of spatial features at the current time point in the LSTM multilayer temporal modeling, ensuring the sensitivity and accuracy of interference identification at the current time point.

[0047] As a further improvement to the above scheme, in step S6, the weighted joint loss function is calculated according to the following formula:

[0048] ;

[0049] in, , For fixed weighting coefficients, The cross-entropy loss is the current time step used to identify the interference. The average cross-entropy loss for the K-step interference prediction task;

[0050] The cross-entropy loss for interference identification at the current moment is calculated using the following formula:

[0051] ;

[0052] The average cross-entropy loss of the future K-step interference prediction is calculated using the following formula:

[0053] ;

[0054] in, Indicates category index, This represents the one-hot encoding of the current real label. This represents the model's predicted probability for that category at the current time. Indicates the future prediction step size. This represents the index at the k-th step in the future. Indicates the first Time of the first The true label of the class, The model represents the first Time of the first The predicted probability of a class;

[0055] The model adopts a shared LSTM temporal modeling layer and a dual-output branch structure. The two output branches are the current identification head and the future K-step prediction head, respectively. During training, the gradient of the total joint loss is backpropagated to the dual-output branch and the shared LSTM layer.

[0056] The beneficial effects of the above technical solution are as follows: the weighted joint loss function incorporates the current interference identification and the future K-step interference prediction into the same training framework, realizes the sharing of underlying spatiotemporal features, avoids the target bias problem caused by single-task training, and makes the feature representation learned by the model more comprehensive; the clear loss calculation formula makes the model training process quantifiable and controllable, and the gradient backpropagation to the whole network realizes the collaborative optimization of dual output branches and shared layers, effectively improving the overall performance of the model in completing both identification and prediction tasks at the same time.

[0057] As a further improvement to the above scheme, in step S7, the channel risk assessment function based on time attenuation is calculated according to the following formula:

[0058] ;

[0059] in, For channel Comprehensive risk score, Indicates channel The result of the disturbance determination at the current moment, Indicates channel In the future The result of the disturbance determination step A fixed future risk attenuation coefficient;

[0060] The decision-making module first selects channels that are not currently or in the future subject to interference as candidate communication channels. If no such channel exists, then it considers all channels... Sort the channels in ascending order and extract the top-K channels with the lowest risk scores to generate a frequency hopping sequence.

[0061] The beneficial effects of the above technical solutions are as follows: the channel risk assessment function based on time decay fully considers the uncertainty of future predictions, gives higher weight to recent predictions, and makes the comprehensive channel risk score more in line with the actual interference evolution, resulting in a more scientific assessment result; the hierarchical candidate channel screening strategy prioritizes absolutely safe channels and selects low-risk channels when there are no safe channels, and the generated frequency hopping sequence can accurately avoid interference, effectively improving the continuity and stability of the communication link.

[0062] As a further improvement to the above solution, this method is deployed on the Cambricon MLU220-SOM embedded platform, and the deployment process includes:

[0063] In the offline phase, the training and solidification of CNN and LSTM networks are completed on a general computing platform, and the trained model parameters are converted into an executable model format that is compatible with the Cambrian inference framework.

[0064] During the deployment phase, model loading and inference calls are completed on the Cambricon MLU220-SOM side through the Cambricon runtime. At the same time, end-to-end processing latency is reduced by using board-side inference operator adaptation, data type optimization, tensor input organization, and parallel inference.

[0065] The method has an average single inference time of 37ms on a general computing platform and an average single inference time of 27ms on the Cambricon MLU220-SOM embedded platform.

[0066] The beneficial effects of the above technical solution are as follows: A dedicated end-to-end deployment process was designed for the Cambricon MLU220-SOM embedded platform. Through offline model conversion and board-side multi-dimensional inference optimization, the model was able to run efficiently on the embedded platform. Compared with the inference latency of general computing platforms, the inference latency was significantly reduced, achieving the millisecond-level real-time inference requirements at the edge. This transformed the method from a theoretical model into a technical solution with practical engineering implementation capabilities, meeting the application requirements of real-time anti-interference in the field.

[0067] As a further improvement to the above scheme, this method supports multi-channel expansion, and the number of channels can be adjusted to 4 or 16. The multi-channel PSD data is organized into network input by frequency domain stitching, channel stacking or matrix mapping. The CNN backbone network can be replaced with a convolutional network with image feature extraction capabilities, and the LSTM temporal modeling unit can be replaced with GRU, bidirectional recurrent network or temporal convolutional network. The feature bypass residual connection mechanism can be adjusted to feature stitching, weighted fusion or gated fusion. The weight coefficients α and β of the dual-task joint loss can be adjusted according to the interference jump rate.

[0068] The beneficial effects of the above technical solution are as follows: it gives the method a strong degree of flexibility and scalability. It can flexibly adjust the network structure, feature fusion method and loss function weight according to the actual needs of the actual communication system such as channel configuration, hardware computing power and interference scenarios. It can adapt to different engineering application scenarios, break through the limitations of fixed model structure on application scenarios, and greatly improve the practical engineering applicability and generalization ability of the method.

[0069] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0070] Firstly, this invention addresses the practical problems of time-varying, sudden, and diverse interference signals in complex electromagnetic environments. It establishes a complete technical process from raw radio frequency data acquisition, power spectral density construction, interference identification, future trend prediction to strategy generation and frequency adjustment execution. This process elevates traditional solutions from static judgment at the current moment to dynamic perception and advance decision-making several moments into the future. In this way, the system can provide both the current interference state and predict subsequent interference evolution trends within each observation cycle, thus providing cognitive radio equipment with earlier, more continuous, and more targeted decision-making basis. This significantly improves the time misalignment problem between interference perception and anti-interference decision-making in complex adversarial scenarios.

[0071] Secondly, in terms of input representation, this invention uses the original radio frequency signal as a basis, calculates the power spectral density of eight channels using Fast Fourier Transform, and then stitches the PSDs of each channel together in frequency order to form a full-band PSD image. This image is then fed into the network after scaling and normalization. This data organization method can uniformly map multi-channel frequency domain information onto the same spectrum, preserving key characteristics of interference in spectral distribution, energy accumulation, frequency band occupancy, and cross-channel correlation. This facilitates the model's more complete learning of the spatial distribution patterns of complex interference.

[0072] Third, this invention employs a spatiotemporal fusion modeling approach combining CNN and LSTM, which can simultaneously consider spatial discrimination information in a single frame of spectral image and temporal evolution information in a continuous observation sequence. The document explicitly uses ResNet18 as the CNN backbone network to extract cross-frequency and cross-channel spatial features from a single frame of PSD image, outputting three probabilities: no interference, fixed-frequency interference, and swept-frequency interference. Subsequently, LSTM is used to perform temporal modeling on the CNN features across multiple consecutive frames, further capturing the frequency drift characteristics of swept-frequency interference, the persistent characteristics of fixed-frequency interference, and the evolution of interference states over time. Through this network organization, the system can simultaneously consider current state perception and future trend inference in the same modeling process, thus possessing a stronger ability to characterize complex interference phenomena such as rapid jumps, continuous persistence, and cross-temporal diffusion, which is beneficial for improving the stability of identification and prediction results in complex electromagnetic environments.

[0073] Fourth, this invention further introduces feature bypass residual connections and dual-task joint loss design in its model structure and training mechanism, which can enhance the sensitivity of current-moment recognition while taking into account the effectiveness of future multi-step prediction. To avoid information decay of spatial features at the current moment after LSTM multi-layer temporal modeling, the CNN features of the latest frame in the sliding window are directly added to the current recognition head output as residuals, which helps maintain the sensitivity of current interference state discrimination. At the same time, during the model training phase, the current recognition task and the future K-step prediction task are unified into the same framework, and weighted joint loss is used for end-to-end training, so that the shared spatiotemporal features serve both current category discrimination and future trend modeling. This structural arrangement helps to reduce the target bias problem common in single-task training, making the model learn a more comprehensive representation, and is more suitable for the business needs of "being able to see both the present and subsequent changes" in actual anti-interference scenarios.

[0074] Fifth, this invention does not stop at simply "providing a category label" at the output end. Instead, it further uses the current identification results and the prediction results for the next K steps for candidate channel selection, risk quantification assessment, frequency hopping sequence generation, and parameter control information distribution, forming a closed-loop processing flow from perception, judgment, prediction to execution. In particular, it designs a channel risk assessment approach based on time decay, which can comprehensively consider the current disturbance state and the disturbance trend in the next few steps to form a comparable risk score for each channel. Based on this, it prioritizes the channels that are safer now and in the future, or generates suggested frequency hopping sequences from the channels with the lowest risk. This output method is closer to the actual usage needs of the equipment side, and can directly transform the results of the network model into an executable anti-interference strategy, improving link continuity, timeliness of frequency adjustment, and communication reliability.

[0075] Sixth, the present invention possesses superior multi-channel online processing capabilities. The inference script employs a one-time inference method using eight subgraphs to form a batch tensor, outputting an 8×3 shape. After inference, an 8-dimensional prediction result vector is generated, and the results directly identify whether each channel is interference-free, fixed-frequency interference, or frequency-sweeping interference, while simultaneously providing a description of the disturbed channel number. In other words, this scheme already possesses the engineering foundation for parallel identification of eight channels, naturally organizing single-channel classification results into a global channel state description, providing fine-grained input for subsequent risk assessment, frequency point selection, and frequency hopping control. Compared to schemes that only output a single overall conclusion, this multi-channel, vectorized output format, which directly serves the decision-making module, is more suitable for online deployment and coordinated control in real wireless communication systems.

[0076] Seventh, this invention has good board-end deployment value and real-time application prospects. The final implementation scheme has been clarified, and the algorithm can be ported and accelerated for the Cambricon MLU220-SOM embedded platform. End-to-end processing latency is reduced through board-end inference operator adaptation, data type optimization, tensor input organization, and parallel inference. The results presented in the material show that the average single inference time on a general computing platform is approximately 37ms. After porting the algorithm to the MLU220-SOM, the average single inference time is further reduced to approximately 27ms. This demonstrates that this invention not only possesses methodological advancements but also strong engineering implementation capabilities. It can meet the millisecond-level online inference requirements at the edge while ensuring the integrity of recognition and prediction functions, providing reliable support for future real-time active anti-interference applications in cognitive radio, unmanned platform communication links, and complex electromagnetic environments. Attached Figure Description

[0077] Figure 1 Design the overall framework for the system.

[0078] Figure 2 The four stages are closely connected and integrated through seven specific steps from S1 to S7.

[0079] Figure 3 The algorithm architecture diagram for incorporating the feature bypass residual connection mechanism.

[0080] Figure 4 The flowchart shows the training process for dual-task output and joint loss. Detailed Implementation

[0081] To enable those skilled in the art to better understand the technical solution, the present invention will be described in detail below with reference to embodiments. The description in this part is only exemplary and explanatory, and should not be used to limit the scope of protection of the present invention in any way.

[0082] like Figure 1As shown, the overall system design framework of the present invention includes a raw radio frequency data acquisition module, a signal preprocessing and PSD image construction module, a CNN-LSTM spatiotemporal fusion deep learning model, an active anti-interference decision and strategy generation module, and a cognitive radio execution module. The raw RF data acquisition module acquires multi-channel RF sampling data in each observation period; the signal preprocessing and PSD image construction module calculates the power spectral density of each channel through fast Fourier transform and stitches the 8 channel PSDs into a full-band PSD image in frequency order; the image normalization preprocessing module performs size unification and normalization processing on the PSD image to meet the CNN input requirements; the CNN-LSTM spatiotemporal fusion deep learning model extracts cross-band and cross-channel spatial discrimination features from a single frame PSD image and outputs the current interference type. At the same time, the feature sequence sliding window in the CNN-LSTM spatiotemporal fusion deep learning model caches the CNN features of the most recent N frames and forms a temporal input sequence. The LSTM temporal modeling prediction network outputs the interference type prediction result for the next K steps based on the historical feature sequence. Finally, through dual-task joint training, the current interference state and the future predicted interference state are output simultaneously; the active anti-interference decision and strategy generation module calculates the channel risk score based on the current identification result and the future prediction result, generates candidate communication channels and frequency hopping sequences, and sends the parameter control information to the cognitive radio execution module to realize closed-loop processing from interference perception, trend prediction to strategy execution.

[0083] like Figure 2 As shown, the technical solution of the present invention includes four stages: data acquisition and preprocessing, spatial feature extraction, temporal modeling and multi-step prediction, and result output and closed-loop application.

[0084] The first two stages complete static spatial modeling, while the latter two stages achieve dynamic prediction and decision-making closed loop based on spatiotemporal fusion, ensuring that the system completes end-to-end millisecond-level inference within the observation period.

[0085] The data acquisition and preprocessing stages correspond to steps S1 and S2, which are responsible for acquiring the original radio frequency signals, generating PSD images through FFT transformation, and normalizing them, providing standardized, high-information-density inputs for subsequent models.

[0086] The spatial feature extraction step corresponds to step S3, which uses a CNN network to extract cross-frequency band and cross-channel spatial discrimination features from a single frame PSD image to achieve accurate identification of the interference type at the current moment.

[0087] The temporal modeling and multi-step prediction process corresponds to steps S4 and S5. By constructing historical CNN feature sequences and inputting them into an LSTM network, the temporal evolution patterns of interference types are captured, enabling reliable predictions for the next K time points.

[0088] The output corresponds to steps S6 and S7 in the closed-loop application. A dual-task joint loss is used to train the model, enabling it to simultaneously provide the current interference type identification result and the interference trend estimate for the next few steps during a single inference process. Based on the above output, the system performs availability screening and risk quantification assessment on each candidate channel. Under the premise of meeting communication constraints, it generates frequency selection and frequency hopping suggestion sequences, and sends these sequences to the cognitive radio module to complete radio frequency switching and parameter configuration updates. This achieves a continuous workflow from interference perception and trend estimation to frequency adjustment, reducing the impact of policy delivery lag on communication reliability.

[0089] The flowcharts for each stage are implemented through seven specific steps, S1 to S7, such as... Figure 2 As shown, the specific steps are as follows:

[0090] S1: Raw RF data acquisition and PSD image construction;

[0091] Raw radio frequency data of the communication system is collected in each observation period. The power spectral density of each channel is calculated by fast Fourier transform. Then, the PSDs of the eight channels are stitched together in frequency order to generate a full-band PSD image.

[0092] ;

[0093] in Indicates the first Each channel at discrete frequency The power spectral density value at that location, Indicates the first The original radio frequency sampling sequence in the time domain of each channel, where N represents the number of FFT points; This represents a discrete frequency index.

[0094] The PSD images of the 8 channels are horizontally stitched together in frequency order to generate a full-band PSD image:

[0095] ;

[0096] in This represents a full-band PSD image after horizontal stitching. This indicates a horizontal splicing operation in frequency order.

[0097] To support channel-level parallel inference and output an 8-dimensional result vector, The image is divided horizontally into 8 equal parts, forming 8 sub-images, each corresponding to an independent channel.

[0098] S2: Image normalization preprocessing;

[0099] The full-band PSD image or its segmented sub-images are subjected to uniform scaling and normalization to meet the input requirements for CNN inference.

[0100] S3: CNN spatial feature extraction and current time-of-flight identification;

[0101] A CNN feature extraction network with ResNet18 as the backbone is constructed to extract spatial discriminative features from each frame of the PSD image and output the interference type discrimination result at the current time. The core operation of the CNN is two-dimensional convolution:

[0102] ;

[0103] in For the first Layer convolution kernel, For the feature map of the previous layer, ReLU activation is followed by max pooling to gradually reduce the spatial dimension. Residual blocks effectively alleviate gradient vanishing.

[0104] ;

[0105] After multiple layers of convolutional global pooling, an intermediate feature vector of dimension D is obtained. At the same time, the full connection header outputs the current time. ,go through Three types of probabilities were obtained: no interference, fixed-frequency interference, and frequency-sweeping interference.

[0106] S4: Construct historical sequence samples;

[0107] A sliding window of length N is used to cache the CNN feature vectors of the most recent N frames, forming a temporal input sequence:

[0108] ;

[0109] in Let be the feature representation of the PSD image of frame t after CNN encoding.

[0110] S5: LSTM Temporal Modeling and Multi-Step Prediction;

[0111] The above feature sequences are input into an LSTM network for time series modeling to capture the dynamic pattern of interference type evolution over time, thereby enabling the prediction of interference type at the next K time points.

[0112] ;

[0113] The gating update formula for a single time step of LSTM is as follows:

[0114] (The Gate of Oblivion);

[0115] (Input Gate);

[0116] (Candidate cell status);

[0117] (Cell status update);

[0118] (Output gate);

[0119] (Hidden state);

[0120] in Indicates the activation function; Indicates hyperbolic tangent; Represents element-wise multiplication; This indicates the hidden state in the previous moment; This indicates the cell state at the previous moment; This represents the current input, i.e., the CNN feature at step t; This represents the weights and biases of each learnable element; H represents the number of hidden units in the LSTM. To achieve multi-step prediction in the next K steps, the hidden state of the last layer of the LSTM is connected to the time series prediction head.

[0121] ;

[0122] The output can be either a sequence of overall interference types or an 8-channel vectorized sequence. This step captures the time-varying characteristics of swept-frequency interference, such as frequency drift and persistent presence of fixed-frequency interference, achieving a leap from static identification to dynamic trend prediction.

[0123] Algorithm architecture such as Figure 3 As shown, to further prevent information attenuation of spatial features at the current moment after multi-layer temporal modeling by LSTM, this invention introduces a feature bypass residual connection mechanism when outputting the interference type identification result at the current moment. Specifically, the latest frame CNN features in the sliding window sequence are directly added to the residual of the current recognition head output features of the LSTM. This mechanism ensures the sensitivity of interference state discrimination at the current moment, while allowing the LSTM network to focus more on learning the evolutionary patterns of future time steps.

[0124] S6: Dual-task output and joint loss training;

[0125] Construct a dual-task learning structure: Task A outputs the current interference type identification, and Task B outputs the interference type sequence prediction for the next K steps. During training, a weighted joint loss function is used to enable the model to share underlying spatiotemporal features while balancing identification and prediction accuracy.

[0126] ;

[0127] in , These are weighting coefficients, which can be dynamically adjusted according to the disturbance transition rate; and Using cross-entropy loss:

[0128] ;

[0129] ;

[0130] in This represents the cross-entropy loss value for the interference type identification task at the current moment; Indicates a category index; The one-hot encoding representing the actual label; This represents the model's predicted probability for that category at the current time. This represents the average cross-entropy loss value for the K-step interference type prediction task. Indicates the future prediction step size; Indicates the index of the k-th step in the future; Indicates the k-th time in the future; This represents the true label of class c at time t+k. This represents the model's predicted probability for class c at time t+k; This means averaging the loss over K steps to ensure consistent loss scale across different prediction step sizes.

[0131] The training process for dual-task output and joint loss is as follows: Figure 4 As shown, before the training data enters the model, the CNN first extracts features from the continuous spectrum, forming a time-ordered feature sequence as input. The supervision signal is divided into two parts: one is the real interference label at the current moment, and the other is the sequence of real labels for the next K steps. The main body of the model uses a shared LSTM temporal modeling layer to encode the input sequence. Based on the shared representation, two output branches are generated: the current moment recognition head outputs the current moment's logits, used to obtain the current category; the next K step prediction head outputs the next K step's logits sequence, used to obtain the prediction results for each future step. During training, the task losses of the two branches are calculated separately: the current recognition corresponds to task A loss, and the future prediction corresponds to task B loss. Then, they are weighted and summed according to weights α and β to obtain the total joint loss. Finally, the gradient of the total loss is backpropagated back to the two branches and the shared LSTM layer, so that the shared layer simultaneously considers both the current discrimination and future prediction targets, thereby completing dual-task learning under the same set of parameters.

[0132] S7: Results Output and Application;

[0133] Current recognition result And future K-step prediction results As input for strategy generation, the availability screening and risk assessment of candidate channels are performed under communication constraints. The output includes a set of communicable frequencies, frequency hopping sequences, and corresponding parameter control information. The results are then sent to the cognitive radio station for radio frequency switching and parameter updates, enabling anti-interference actions to be implemented on the equipment side in a timely manner, thereby improving link continuity and communication reliability in time-varying interference environments.

[0134] At the same time, the model outputs the disturbed channel at the current moment. And future K-step perturbation channel prediction As input, a time-decrease-based channel risk assessment function is constructed to calculate a comprehensive risk score for each channel:

[0135] ;

[0136] in, and Representing channels respectively At the present moment and in the future The result of the disturbance determination step; This represents the future risk attenuation coefficient, indicating that the later the prediction time, the higher the uncertainty, and the smaller its weight in the current decision. The decision module prioritizes channels that are safe both now and in the future as candidate communication channels; if no absolutely safe channel is available, then... The system sorts the channels in ascending order, extracts the top-K channels with the lowest risk scores to generate suggested frequency hopping sequences, and sends these sequences to the cognitive radio module to perform radio frequency switching and parameter updates. This enables the system to complete a closed-loop process consisting of identification result output, prediction of future interference trends, calculation of channel risk scores, and execution of frequency adjustment.

[0137] The interference type identification and prediction algorithm proposed in this invention is deployed on an embedded system to achieve real-time identification of communication signal interference and proactive anti-interference decision-making for the future. The overall algorithm process includes four stages: power spectral density map construction, spatial feature extraction, temporal modeling and prediction, and strategy output. First, the raw data stream of the communication system is acquired. The signal processing module is called on the board to perform fast Fourier transform on the raw data and calculate the power spectral density. The power spectral densities of the eight channels are spliced ​​in frequency order to form a full-band PSD image. The image is then resized and normalized to meet the input requirements of the deep network. Subsequently, the single-frame full-band PSD image is fed into the CNN backbone network to extract frequency domain spatial discriminative features and output the interference type identification result at the current moment. At the same time, the intermediate feature vectors of the CNN are retained for subsequent temporal modeling. Based on this, the CNN features of the most recent N frames are arranged in time order to form a sliding window sequence and input into the LSTM network to capture the dynamic evolution of interference over time and output the interference type prediction sequence for the next K moments.

[0138] At the level of proactive anti-interference methods, the system generates a set of communicable frequencies and a frequency hopping suggestion sequence based on the current identification results and the prediction results of the next K steps. This is used to avoid frequency bands and channels where interference is predicted to occur in advance, achieving forward-looking frequency planning. Specifically, the CNN output is used to determine the current set of disturbed channels and the type of interference, while the LSTM output is used to characterize the evolution trend of interference in the future time domain. When the prediction results show that a certain type of interference will evolve from no interference to fixed-frequency interference or frequency-sweeping interference in the future, the strategy module removes the potential disturbed channels from the candidate set in advance and selects the communication channel and frequency hopping order from the remaining channels. This allows the communication link to complete frequency switching and resource preparation before the interference arrives, reducing the risk of link interruption and decision lag caused by the traditional "switching after interference occurs" approach.

[0139] This implementation scheme allows for configuration of the prediction step size K, sliding window length N, and output granularity according to engineering requirements: the prediction output can be either a sequence of overall interference types or a channel-level vectorized result, so as to drive frequency set selection and frequency hopping sequence planning more precisely; at the same time, the inference results and policy output can be recorded with timestamps for subsequent online statistics, threshold adaptation, and model iteration updates, thereby continuously improving the ability to characterize interference evolution trends and the effectiveness of active anti-interference in complex electromagnetic environments.

[0140] To meet the real-time requirements of the embedded platform MLU220-SOM, this implementation scheme performs on-board porting and accelerated deployment of the model: In the offline phase, CNN and LSTM networks are trained and solidified in a general training environment, and the trained model parameters are converted into an executable model format adapted to the Cambricon inference framework. In the deployment phase, the model is loaded and inference is invoked on the MLU220-SOM side through the Cambricon runtime, enabling the CNN+LSTM joint model to perform end-to-end inference on the board. During the porting process, the network is adapted and programmed using on-board inference operators and data type optimization strategies to ensure stable operation on the MLU. End-to-end latency is reduced through parallel inference and tensor input organization, thus meeting the online requirements for real-time recognition and prediction. The algorithm's average inference time on a general computing platform is approximately 37ms; after porting and deployment to the Cambricon MLU220-SOM, the average inference time is reduced to approximately 27ms. The interference identification and prediction algorithm proposed in this invention has good deployability and real-time performance on edge AI acceleration boards. It can reduce inference latency while ensuring the integrity of identification and prediction functions, and provide timely decision input for proactive anti-interference strategies for future evolutionary trends.

[0141] Without departing from the core technical concept of this invention, various equivalent substitutions can be made to the specific embodiments of this invention. Regarding the input data construction method, the current embodiment uses a fast Fourier transform to calculate the power spectral density of the original radio frequency data, splices the power spectral densities of the eight channels in frequency order to form a full-band PSD image, and then feeds it into a CNN network for spatial feature extraction. In alternative implementations, depending on the actual communication system, hardware interface form, and data acquisition capabilities, short-time Fourier transform spectra, time-frequency energy maps, normalized spectrum maps, or directly based on channel quality index sequences can be used to construct the network input. As long as it can characterize the disturbance state of each channel, the spectral energy distribution, and its temporal evolution relationship, it can be used to achieve the interference identification and future trend prediction functions described in this invention. Correspondingly, the current embodiment uses eight channels as an example for explanation. In practical applications, the number of channels can be expanded to four, sixteen, or more channels depending on the system configuration. Multi-channel data can be organized using frequency domain splicing, or it can be fed into the subsequent network processing module using channel stacking, matrix mapping, or sequence expansion. Essentially, it still falls within the technical scope of this invention, which uses multi-channel interference observation information to drive spatiotemporal joint modeling.

[0142] Regarding model composition, the current embodiment employs a spatiotemporal fusion network combining CNN and LSTM. A single-frame full-band PSD image first undergoes spatial discriminative feature extraction via a CNN backbone network. Then, the features from the most recent N frames are sequentially input into an LSTM for temporal modeling, simultaneously outputting the current interference identification result and the predicted interference type for the next K steps. In an alternative embodiment, the CNN backbone network can be replaced with other convolutional networks capable of image feature extraction, and the temporal modeling unit can be replaced with GRU, bidirectional recurrent networks, temporal convolutional networks, or other sequence models capable of characterizing the temporal evolution of interference. As long as the overall approach of extracting spatial features from current observation information, modeling future trends from historical sequences, and further completing the dual tasks of current identification and future prediction is retained, it falls within the scope of this invention. Furthermore, the current embodiment employs feature bypass residual connections to enhance the responsiveness of the current identification result to the spatial features of the latest frame. In alternative embodiments, this bypass connection can also be adjusted to feature splicing, weighted fusion, or gated fusion depending on the model size and on-board computing power. The current embodiment uses a weighted joint loss function to train the current identification task and the future prediction task in a unified manner. In the alternative, the weight coefficients of the two tasks can also be adaptively adjusted according to the interference jump rate, sample distribution and the business side’s emphasis on real-time identification or forward prediction. The purpose is still to enable the shared spatiotemporal features to serve both current discrimination and future prediction.

[0143] Regarding the output format and engineering deployment method, in the current embodiment, the strategy module generates a set of communicable frequencies and a frequency hopping suggestion sequence based on the current identification results and the prediction results of the next K steps, in order to avoid potential interference channels in advance. In an alternative embodiment, the output results can be further extended to interference risk scores, candidate channel priority ranking, frequency point switching suggestion tables, or parameter issuance instructions for the device control interface. The strategy layer can also complete frequency resource scheduling by using threshold decision, risk-weighted screening, rule matching, or heuristic frequency hopping planning according to different system requirements. Meanwhile, the current embodiment has shown that this method can be deployed on the MLU220-SOM embedded platform and meets the real-time online operation requirements through model conversion, board-side adaptation, and inference optimization. In an alternative embodiment, it can also be deployed on GPU platforms, other NPU platforms, FPGA coprocessing platforms, or embedded computing devices with edge inference capabilities. As long as real-time processing of raw observation data, online identification and prediction of interference states, and timely generation and execution of anti-interference strategies are achieved, they can all be considered equivalent replacements for the embodiments of the present invention.

[0144] It should be noted that, in this document, the terms "comprising," "including," and any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Specific examples have been used in this document to illustrate the principles and implementation methods of the present invention. These examples are merely for the purpose of helping to understand the method and core ideas of the present invention. The above descriptions are only preferred embodiments of the present invention. It should be pointed out that, due to the limitations of written expression and the objective existence of infinite specific structures, those skilled in the art can make several improvements, modifications, or variations without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner. These improvements, modifications, variations, or combinations, or the direct application of the concept and technical solution of the present invention to other situations without modification, should all be considered within the scope of protection of the present invention.

Claims

1. A communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion, applied to anti-interference processing of communication signals in cognitive radio systems, used for current identification and multi-step prediction of interference types in communication signals and generation of anti-interference execution strategies, characterized in that... This method sequentially executes the steps of data acquisition and preprocessing, spatial feature extraction, temporal modeling and multi-step prediction, dual-task joint training, result output and closed-loop application, to achieve current identification and future K-step prediction of communication signal interference types, where K is a positive integer, and the interference types are fixed as three categories: no interference, fixed-frequency interference, and frequency-sweeping interference, specifically including: S1: Acquisition of raw radio frequency data and construction of PSD image. Raw radio frequency data of communication system is collected in each observation period. Power spectral density of 8 channels is calculated by fast Fourier transform. PSD of 8 channels is stitched together in frequency order to generate full-band PSD image. S2: Image normalization preprocessing, which performs uniform scaling and normalization on the full-band PSD image so that the processed image meets the input requirements for CNN inference; S3: CNN spatial feature extraction and current time identification. Construct a CNN feature extraction network with ResNet18 as the backbone to extract cross-frequency band and cross-channel spatial discrimination features for each frame of PSD image and output the interference type discrimination result at the current time. S4: Construct historical sequence samples, using a sliding window of length N to cache the CNN feature vectors of the most recent N frames, where N is a positive integer, to form a temporal input sequence; S5: LSTM Temporal Modeling and Multi-Step Prediction. The temporal input sequence is input into the LSTM network for temporal modeling, capturing the temporal evolution of interference types and outputting the interference type prediction results for the next K time steps. At the same time, a feature bypass residual connection mechanism is introduced to prevent the decay of spatial feature information at the current time step. S6: Dual-task output and joint loss training. Construct a dual-task learning structure that includes task A for identifying interference at the current time and task B for predicting interference in the next K steps. Use a weighted joint loss function to train the model end-to-end so that the model shares the underlying spatiotemporal features. The weighted joint loss function is calculated using the following formula: ; in, , For fixed weighting coefficients, satisfying 0 < α < 1 and 0 < β < 1, The cross-entropy loss is the current time step used to identify the interference. The average cross-entropy loss for the K-step interference prediction task; The cross-entropy loss for interference identification at the current moment is calculated using the following formula: ; The average cross-entropy loss of the future K-step interference prediction is calculated using the following formula: ； Where 'c' represents the category index. This represents the one-hot encoding of the current real label. This represents the model's predicted probability of this category at the current moment, where k represents the index at the k-th future step. This represents the true label of class c at time t+k. This represents the model's predicted probability for class c at time t+k; The model adopts a structure with a shared LSTM temporal modeling layer and two output branches. The two output branches are the current identification head and the future K-step prediction head, respectively. During training, the gradient of the total joint loss is backpropagated to the two output branches and the shared LSTM layer. S7: Results Output and Application. The current identification results and the prediction results of the next K steps are used as inputs to construct a channel risk assessment function based on time decay to calculate the comprehensive risk score of each channel. Based on the risk score, candidate communication channels are selected and frequency hopping sequences are generated. The frequency hopping sequences and parameter control information are sent to the cognitive radio module to complete the radio frequency switching and parameter configuration update, forming a closed-loop processing flow from interference perception to execution. In step S7, the channel risk assessment function based on time decay is calculated according to the following formula: ； in, For channel Comprehensive risk score, Indicates channel The result of the disturbance determination at the current moment, Indicates channel In the future The result of the disturbance determination step A fixed future risk attenuation coefficient; The decision-making module first selects channels that are neither currently nor in the future subject to interference as candidate communication channels. If no candidate communication channel is currently or in the future subject to interference, then all channels are... Sort the channels in ascending order and extract the top-K channels with the lowest risk scores to generate a frequency hopping sequence.

2. The communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion according to claim 1, characterized in that, In step S1, the i-th channel at discrete frequency The power spectral density at a given point is calculated using the following formula: ； in Let N represent the original radio frequency sampling sequence in the time domain of the i-th channel, where N represents the number of FFT points and k represents the discrete frequency index. Full-band PSD images are constructed using the following formula: ； in, This represents a full-band PSD image after horizontal stitching, while concat indicates a horizontal stitching operation in frequency order.

3. The communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion according to claim 2, characterized in that, In step S2, scaling is used to adjust the PSD image to the fixed input size required by the ResNet18 backbone network of the CNN, and normalization is used to map the pixel values ​​of the PSD image to the numerical range of 0-1, so that the size and numerical range of the processed image match the input standard of CNN inference.

4. The communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion according to claim 3, characterized in that, In step S3, the core operation of CNN is two-dimensional convolution, which is performed according to the following formula: ； in, For the l-th layer convolution kernel, For the feature map of layer l-1, This is the bias of the c-th channel in the l-th layer; In CNNs, residual blocks are used to mitigate the vanishing gradient problem. The residual blocks are implemented according to the following formula: ; After multiple convolutions and global pooling, an intermediate feature vector of dimension D is obtained. The fully connected head of the CNN outputs the current time-time logits, which, after being activated by softmax, yield the prediction probabilities for three classes: no interference, fixed-frequency interference, and frequency-sweeping interference.

5. The communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion according to claim 4, characterized in that, In step S4, the time-series input sequence is constructed according to the following formula: ; in, Let t be the feature representation of the PSD image after CNN encoding. It is the time-series input sequence composed of CNN feature vectors from time t-N+1 to time t.

6. The communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion according to claim 5, characterized in that, In step S5, the LSTM sequentially passes through the forget gate, input gate, candidate cell state, cell state update, and output gate to complete the gating update and obtain the hidden state in a single time step. The calculation formulas for each step are as follows: Forgotten Gate: ; Input Gate: ; Candidate cell status: ; Cell status update: ; Output gate: ; Hidden state: ; in, This represents the sigmoid activation function. This represents the hyperbolic tangent activation function. Represents element-wise multiplication. This represents the hidden state at time t-1. This indicates the cell state at time t-1. This represents the CNN feature at step t of the current input. , This represents the weights and biases of each learnable element; The LSTM's timing prediction header connects the last hidden state to the fully connected layer and outputs the prediction result according to the following formula: ; The feature bypass residual connection mechanism involves performing dimensionality matching between the CNN features of the latest frame in the sliding window and the current recognition head output features of the LSTM, and then adding the element-wise residuals.

7. The communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion according to claim 1, characterized in that, This method was deployed on the Cambricon MLU220-SOM embedded platform, and the deployment process included: In the offline phase, the training and solidification of CNN and LSTM networks are completed on a general computing platform, and the trained model parameters are converted into an executable model format that is compatible with the Cambrian inference framework. During the deployment phase, model loading and inference calls are completed on the Cambricon MLU220-SOM side through the Cambricon runtime. At the same time, board-side inference operator adaptation, data type optimization, tensor input organization, and parallel inference are adopted to reduce end-to-end processing latency.

8. The communication interference identification and prediction method based on CNN-LSTM spatiotemporal fusion according to claim 1, characterized in that, This method supports multi-channel expansion, with the number of channels adjustable to 4 or 16. Multi-channel PSD data is organized as network input using frequency domain stitching, channel stacking, or matrix mapping. The CNN backbone network can be replaced with a convolutional network capable of image feature extraction, and the LSTM temporal modeling unit can be replaced with GRU, bidirectional recurrent network, or temporal convolutional network. The feature bypass residual connection mechanism can be adjusted to feature fusion forms such as feature stitching, weighted fusion, or gated fusion. The weight coefficients α and β of the dual-task joint loss can be adjusted according to the interference jump rate.

Images