Electromagnetic spectrum occupancy prediction method based on incremental learning
By employing an incremental learning approach and utilizing a prediction model with a bidirectional Bi-LSTM and self-attention mechanism, the problem of performance degradation of traditional static models in dynamic spectrum environments is solved, achieving efficient spectrum occupancy prediction and reducing the complexity and resource consumption of model updates.
Patent Information
- Application Number
- CN202511409029.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-29
- Publication Date
- 2026-01-20
AI Technical Summary
Existing deep learning-based spectrum occupancy prediction methods suffer from performance degradation in dynamic spectrum environments, high training complexity and resource consumption, and are unable to effectively utilize new data to update the model.
We employ an incremental learning approach, using a prediction model with a bidirectional Bi-LSTM and self-attention mechanism structure, combined with incremental training and validation sets, to gradually update the model parameters, avoiding retraining on all data and continuously adapting to changes in the spectral environment.
While ensuring prediction accuracy, the complexity and resource consumption of model updates are significantly reduced, and the prediction performance and adaptability of the model in dynamic spectral environments are improved.
Smart Images
Figure CN121367556A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of intelligent processing of radio signals, and particularly relates to an electromagnetic spectrum occupancy prediction method based on incremental learning. BACKGROUND
[0002] In actual radio spectrum measurement, spectrum data is not static but arrives in the form of continuous data stream. As time goes on, new data points are continuously collected. These data may contain new patterns or trends, reflecting the dynamic changes of the spectrum environment.
[0003] However, most of the current deep learning-based spectrum occupancy prediction methods store all historical data and train the model centrally. That is, after training the model once using a set of historical training data, if you want to integrate the knowledge in the new data, you usually need to retrain after merging all historical data with new data to realize parameter update. This learning method not only increases the training complexity and computational overhead of the model, reduces its operation efficiency, but also does not bring a significant improvement in prediction accuracy in many scenarios.
[0004] Incremental learning can continuously process continuous spectrum data stream in real environment, absorb new spectrum data while retaining and even integrating and optimizing existing spectrum data, without retraining the prediction model from scratch, thereby realizing continuous online update of the prediction model. This learning method provides spectrum data in incremental form, enabling the model to quickly and effectively adapt to the dynamically changing spectrum environment and effectively link the limitation of insufficient device computing power. With the help of incremental learning, the system does not need to store the completed spectrum data set, but stores and uses incremental data step by step, completes the update iteration of the model under the premise of avoiding large-scale retraining.
[0005] The application adopts the method of incremental learning to predict the spectrum occupancy in view of the dynamic evolution of spectrum data, solves the problem of performance decline of traditional static model in dynamic spectrum environment, and significantly reduces the complexity and resource consumption of model update, and the complexity of spectrum occupancy prediction. SUMMARY
[0006] (I) Technical problem to be solved
[0007] The technical problem to be solved by the application is to provide an electromagnetic spectrum occupancy prediction method based on incremental learning to solve the problem of performance decline of traditional static model in dynamic spectrum environment.
[0008] (II) Technical scheme
[0009] In order to solve the above technical problems, the present application provides an electromagnetic spectrum occupancy prediction method based on incremental learning, which comprises the following steps:
[0010] S1, time series data set preprocessing
[0011] The total amount of spectrum data is evaluated and determined, and the historical wireless signal spectrum data set is systematically preprocessed and divided to obtain a basic training set and a basic validation set; according to different incremental time scales, the newly arrived spectrum data collected in time sequence in recent times is taken as an incremental data set, and is divided into an incremental training set, an incremental validation set and an incremental test set;
[0012] S2, basic prediction model training
[0013] The basic prediction model is composed of a bidirectional Bi-LSTM and a self-attention mechanism structure, which is used to predict the spectrum occupancy; the model is first trained on the basic data set to preliminarily learn the basic features and dependent relationships in the data, and is verified by the basic validation set; after this stage, the basic prediction model obtains the initial prediction ability for the basic pattern of the data;
[0014] S3, incremental prediction model training
[0015] The incremental validation set at each time is input into the existing basic prediction model for verification, and then, using the incremental training set, the hyperparameters of the working basic prediction model are adjusted to obtain an updated incremental prediction model; the performance of the updated incremental prediction model is evaluated using the incremental test set;
[0016] S4, adaptive iteration
[0017] Every time span, the spectrum data is collected as an incremental data set to update the working prediction model, the performance of the existing model is evaluated and verified using the latest incremental data set, and the model is ensured to maintain high-performance spectrum occupancy prediction in the continuous spectrum monitoring data; the updated model is deployed in the production environment to continue to serve new requests, and at this time the model should better adapt to the new data distribution.
[0018] (Three) beneficial effects
[0019] The present application provides an electromagnetic spectrum occupancy prediction method based on incremental learning, and the radio spectrum occupancy prediction method of the present application ensures the prediction accuracy, and the time consumption of incremental learning is much lower than that of traditional batch learning, solving the problem of decreased prediction performance of traditional static models in dynamic electromagnetic spectrum environment. In addition, the complexity and resource consumption of model updating are significantly reduced, and the complexity of spectrum occupancy prediction is reduced. BRIEF DESCRIPTION OF DRAWINGS
[0020] Figure 1 The principle diagram of the spectrum occupancy dynamic prediction based on incremental learning of the application;
[0021] Figure 2 The flow chart of the spectrum occupancy dynamic prediction based on incremental learning. DETAILED DESCRIPTION
[0022] In order to make the purpose, content and advantages of the application more clear, the specific embodiments of the application are described in further detail below in combination with the drawings and examples.
[0023] In order to solve the above problems, the application provides a spectrum occupancy prediction method based on incremental learning, which comprises the following steps:
[0024] S1, time series data set preprocessing
[0025] According to the monitored frequency range, resolution bandwidth and monitoring time, the total size of the spectrum data is evaluated and determined. On this basis, the wireless signal spectrum data set is systematically preprocessed and divided. Specifically, it includes a basic training set and a basic validation set. According to different incremental time scales, the newly arrived spectrum data collected in chronological order in the near future is divided into an incremental data set, an incremental training set, an incremental validation set and an incremental test set, which is used to update the model parameters, so that it can absorb new features and patterns in the data to support continuous learning and evaluation of the model.
[0026] S2, basic prediction model training
[0027] The basic prediction model is composed of a bidirectional Bi-LSTM and a self-attention mechanism structure, which is designed to fully utilize the advantages of different components to efficiently capture complex patterns and dependencies in spectrum occupancy. Among them, the bidirectional Bi-LSTM layer is the core time series feature extractor of the model, and the self-attention mechanism structure dynamically adjusts the weights of the feature sequence output by Bi-LSTM, so that the model can automatically focus on the most critical time steps and features for the current prediction task. The model is first trained on the basic data set to preliminarily learn the basic features and dependencies in the data, and the basic validation set is used for verification; after this stage, the model obtains the initial prediction ability for the basic patterns of the data.
[0028] S3, incremental prediction model training
[0029] On the basis of the preliminary prediction ability of the basic prediction model, an incremental learning mechanism is introduced to continuously adapt to the dynamically changing spectrum data.
[0030] S31, model evaluation and update decision
[0031] The incremental data set of each time is input into the existing model for verification. The main purpose of doing this is not for training, but to establish a key performance benchmark to accurately evaluate the effect of subsequent incremental learning. Guided by the verification results of this time, it helps to decide whether to update the model next.
[0032] S32, incremental prediction model update
[0033] After triggering the model update, the old model's learning rate, regularization coefficient and other hyperparameters need to be adjusted. When updating the model, it is not necessary to start training from scratch, but to use only the newly arrived spectrum data to train from the current model, thereby integrating new knowledge into the existing model. Because only the changing part of the new data is involved, this process is usually more efficient than retraining the entire model. The above process will be repeated periodically. Every set time interval, the system will automatically check the model performance and, if necessary, use the new data in this period to implement incremental updates, so that the model can continuously track the dynamic changes of the data.
[0034] S33, incremental prediction model evaluation
[0035] In the spectrum occupancy prediction system based on incremental learning, the model will be updated using new data at a preset time period (such as every day or every week) to continuously adapt to changes in the spectrum environment. Part of the incremental data set is used as an incremental test set to estimate the performance of the updated incremental model. The system performance is measured by the three core evaluation indicators: mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R2) to measure the prediction effect during the entire period from the last model update to the current update. These indicators together provide a comprehensive view of model performance for evaluating the prediction effect during the entire period from the last model update to the current update. Ideally, MSE and MAE are as small as possible (accurate prediction), and R2 is as close to 1 as possible (the model explains the data fluctuations well). If the above performance indicators decrease continuously for N iterations and exceed the set threshold, the system will automatically roll back to the version before the current performance decreased, which is a stable model that has been verified and performs reliably, to ensure the robustness and reliability of the prediction system.
[0036] S4, adaptive iteration
[0037] Every time span, collect spectrum data as an incremental data set, update the prediction model in use, and use the latest incremental data set to evaluate and verify the performance of the existing model, ensuring that the model maintains high performance in predicting spectrum occupancy in continuous spectrum monitoring data. The updated model is deployed in the production environment to continue to serve new requests. At this time, the model should better adapt to the new data distribution and improve the accuracy of prediction.
[0038] Example 1:
[0039] The specific implementation process of the present application is as follows.
[0040] (1) Time series data set preprocessing
[0041] According to the parameters such as the frequency range to be monitored, the resolution bandwidth and the monitoring time, the total size of the spectrum data is evaluated and determined, and on this basis, the wireless signal spectrum data set is systematically preprocessed and divided, including the basic training set and the basic prediction model test set. According to the fixed incremental time span, the spectrum data collected in the recent period of time is divided into incremental verification set, incremental training set and incremental test set, which is used to update the existing model parameters, so that it can absorb the new features and patterns of signals in the latest spectrum data, to support the continuous learning and evaluation of the prediction model.
[0042] In specific implementation, for the 5G wireless signal frequency band 703-4900MHz, the resolution bandwidth is taken as 30kHz, and the environmental electromagnetic signal spectrum data is continuously monitored for 30 days. The data is divided into a basic training set and a basic verification set.
[0043] The fixed incremental time span is set to every 10 days, so from the 31st day of signal monitoring, every 10 days, i.e. 31-40 days, 41-50 days…, the newly collected spectrum data in this period is taken as an incremental data set, and is divided into incremental verification set, incremental training set and incremental test set according to a certain proportion.
[0044] (2) Design of basic prediction model
[0045] The basic prediction model adopts a bidirectional Bi-LSTM and self-attention mechanism structure, which aims to fully exploit the advantages of different components to efficiently capture complex patterns and dependencies in spectrum occupancy. The bidirectional Bi-LSTM layer is the core time series feature extractor of the model, and the self-attention mechanism structure dynamically adjusts the weights of the feature sequence output by Bi-LSTM, so that the model can automatically focus on the most critical features for the current spectrum occupancy prediction task.
[0046] The model is first trained on the basic data set to learn the basic features and dependencies in the spectrum data; after this stage, the model obtains the initial prediction ability for the basic patterns of spectrum data.
[0047] In specific implementation,
[0048] The base training set is input into the Bi-LSTM layer. The forward LSTM of the Bi-LSTM layer processes the input sequence in time order, capturing the spectrum occupancy dependency from history to future; the backward LSTM processes the sequence in reverse order, extracting the spectrum occupancy information from future to history, which helps to identify the influence of potential patterns or periodic features in the future on the current state. For each spectrum amplitude, its forward and backward hidden states are spliced to form a hidden state vector that integrates the complete bidirectional context information. This enables the model to understand any point in the spectrum data based on its entire historical and future information, greatly enhancing the model's ability to model the spectrum time series dynamics, which is crucial for accurate spectrum occupancy prediction.
[0049] The hidden state vector processed by the Bi-LSTM is sent to the self-attention mechanism layer. This mechanism dynamically allocates weights by calculating the correlation between elements within the sequence. The core steps include: the output of the Bi-LSTM is processed through a series of changes to obtain a set of attention weights. These weights quantify the importance of each spectrum amplitude in the spectrum sequence to all other spectrum amplitudes, allowing the prediction model to dynamically focus on the most relevant historical or future spectrum amplitudes for the current prediction, such as sudden high occupancy periods or silent periods, and suppress noise interference, thereby more accurately capturing long-term dependencies and key patterns and improving the accuracy of spectrum occupancy prediction. The feature representation weighted by self-attention is then passed to the fully connected layer to perform the regression prediction task of spectrum occupancy. Finally, the output layer produces continuous prediction values representing the estimate of the future spectrum occupancy state. Finally, the base validation set is used to verify the model effect and confirm whether it meets the expected training goal.
[0050] (3) Incremental prediction model design
[0051] (31) Model evaluation and update decision
[0052] The incremental validation set is input into the existing model to verify the model performance. The main purpose of this is not to train the model, but to establish a key performance benchmark to accurately evaluate the effect of subsequent incremental learning. Based on the results of this validation, it helps to decide whether to update the existing model in the next step.
[0053] In specific implementation, such as Figure 1, take 10-15% of the incremental data set as an incremental validation set, input to the basic prediction model training to verify the effect of spectrum occupation degree prediction, monitor the performance of the existing model on the incremental spectrum data. If the spectrum occupation degree prediction accuracy is high and stable, the existing model is retained and not updated, and the next cycle of monitoring is continued. If the performance of the basic prediction model on the first incremental data is significantly worse than its performance on the basic validation set, it often indicates that the spectrum sampling environment has changed greatly, and the data distribution of the newly arrived spectrum data may have data drift from the old spectrum data, triggering the model update process.
[0054] (32) Incremental prediction model update
[0055] After triggering the model update, the adjustment learning rate, regularization coefficient and other hyperparameters of the existing model in use need to be debugged using the incremental training set. When updating the model, it is not necessary to start from scratch, but to use the current existing model as the starting point and only use the newly arrived spectrum data in the latest time interval for training, so as to integrate new knowledge into the existing model. Because only the changing part of the new data is involved, this process is usually more efficient than retraining the entire model. The above process will be repeated periodically. Every time a set time interval passes, the system will automatically check the performance of the existing model and, if necessary, use the new spectrum data in this period to implement incremental update, so that the model can continuously track the dynamic changes of spectrum data.
[0056] In specific implementation, as shown in Figure 1 , taking the first incremental learning update as an example, after deciding to update the basic prediction model, 60-70% of the first incremental data set is input to the basic prediction model as an incremental training set to update the hyperparameters of the basic prediction model. The basic prediction model tries different hyperparameter configurations on this data subset and selects the best combination according to the spectrum occupation degree estimation accuracy. This process allows the model to better adapt to newly collected spectrum data. After the first incremental learning is completed, the basic prediction model is updated to incremental model 1.
[0057] (33) Incremental prediction model evaluation
[0058] In the spectrum occupancy prediction system based on incremental learning, the prediction model is updated using new data every preset time period to continuously adapt to changes in the spectrum environment. Part of the incremental data set is used as an incremental test set to estimate the performance of the updated incremental model. The model performance is measured by three core evaluation indicators in the current incremental period to evaluate the spectrum occupancy prediction effect in the entire time period from the last prediction model update to the current update. The three core evaluation indicators are mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R2). MSE focuses on the high sensitivity to large prediction errors such as sudden spectrum occupancy peaks or steep drops in spectrum occupancy prediction. The smaller the value, the more accurate the model prediction. MAE focuses on the average level of all prediction errors in spectrum occupancy prediction, reflecting the overall deviation of the model prediction. The smaller the value, the smaller the overall deviation of the model prediction. R2 focuses on the model's ability to explain the fluctuations (variations) in spectrum data in spectrum occupancy prediction. The closer to 1, the better the model can explain the changes in spectrum data. The closer to 0, the weaker the explanation ability, and it may even be worse than directly predicting the mean. These indicators together provide a comprehensive view of model performance for evaluating the prediction effect in the entire time period from the last model update to the current update. Ideally, MSE and MAE are as small as possible (accurate prediction), and R2 is as close to 1 as possible (the model explains the data fluctuations well). If the above performance indicators decrease continuously for N times and exceed the set threshold, the system will automatically roll back to the version before the current performance decrease, which is a stable model that has been verified and has reliable performance, to ensure the robustness and reliability of the prediction system.
[0059] In specific implementation, 15-30% of the incremental data set is input into the trained incremental model as an incremental test set, and the spectrum occupancy prediction result is evaluated according to the predicted value and the true value of each frequency point occupancy.
[0060]
[0061]
[0062] wherein n is the number of samples, y i is the true value, is the predicted value, is the average value of the true value.
[0063] Ideally, MSE and MAE are as small as possible (accurate prediction), and R2 is as close to 1 as possible, and the model explains the data fluctuations well. If MSE is very high but MAE is relatively low, it may mean that the model is mostly correct, but there are a few points with very large prediction errors, such as sudden occupancy that cannot be predicted. If both MSE and MAE are very high, it means that the overall prediction deviation of the model is large.
[0064] (4) Adaptive iteration
[0065] Every time span, collect spectrum data as an incremental data set, update the prediction model in use, evaluate and verify the performance of the existing model using the latest incremental data set, ensure that the model maintains high-performance spectrum occupancy prediction in continuous spectrum monitoring data. The updated model is deployed to the production environment to continue to serve new requests, at this time the model should better adapt to the new data distribution and improve the prediction accuracy.
[0066] Embodiment 2:
[0067] An electromagnetic spectrum occupancy prediction method based on incremental learning, characterized in that the method comprises the following steps:
[0068] S1, time series data preprocessing
[0069] The total size of the spectrum data is evaluated and determined; on this basis, the historical wireless signal spectrum data set is systematically preprocessed and divided to obtain a basic training set and a basic validation set; according to different incremental time scales, the newly arrived spectrum data collected in chronological order in recent times is taken as an incremental data set, and is divided into an incremental training set, an incremental validation set, and an incremental test set;
[0070] S2, basic prediction model training
[0071] The basic prediction model is composed of a bidirectional Bi-LSTM and a self-attention mechanism structure, which is used to predict the spectrum occupancy; the model is first trained on the basic data set to preliminarily learn the basic features and dependency relationships in the data, and is verified using the basic validation set; after this stage, the basic prediction model obtains the initial prediction ability for the basic pattern of the data;
[0072] S3, incremental prediction model training
[0073] The incremental validation set at each time is input into the existing basic prediction model for verification, and then, using the incremental training set, the working basic prediction model is adjusted in super parameters, and an updated incremental prediction model is obtained; the performance of the updated incremental prediction model is evaluated using the incremental test set;
[0074] S4, adaptive iteration
[0075] Every time span, collect spectrum data as an incremental data set, update the prediction model in use, evaluate and verify the performance of the existing model using the latest incremental data set, ensure that the model maintains high-performance spectrum occupancy prediction in continuous spectrum monitoring data. The updated model is deployed to the production environment to continue to serve new requests, at this time the model should better adapt to the new data distribution.
[0076] Step S3 includes
[0077] (31) Model evaluation and update decision
[0078] The incremental validation set at each time is input into the existing model to verify the model performance. The main purpose of doing this is not to train the model, but to establish a key performance benchmark to accurately evaluate the effect of subsequent incremental learning. Guided by the validation results, it helps to decide whether the existing model needs to be updated next.
[0079] (32) Incremental prediction model update
[0080] After triggering the model update, the learning rate, regularization coefficient and other hyperparameters of the existing model being worked on need to be adjusted. When updating the model, it is not necessary to start from scratch, but to use only the newly arrived spectrum data in the latest time interval for training, thereby integrating new knowledge into the existing model. Since only the changing part of the new data is involved, this process is usually more efficient than retraining the entire model. The above process will be repeated periodically. Every time interval, the system will automatically check the performance of the existing model and, if necessary, use the new spectrum data in this period to implement incremental updates, so that the model can continue to track the dynamic changes of the spectrum data.
[0081] (33) Incremental prediction model evaluation
[0082] In the spectrum occupancy prediction system based on incremental learning, the prediction model is updated using new data at a preset time period to continuously adapt to changes in the spectrum environment. Part of the incremental data set is used as an incremental test set to estimate the performance of the updated incremental model. The model performance is measured by three core evaluation indicators in the current incremental period to evaluate the spectrum occupancy prediction effect in the entire time period from the last prediction model update to the current update. The three core evaluation indicators are mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R2). MSE focuses on the high sensitivity to large prediction errors such as sudden spectrum occupancy peaks or steep drops in spectrum occupancy prediction. The smaller the value, the more accurate the model prediction. MAE focuses on the average level of all prediction errors in spectrum occupancy prediction, reflecting the overall deviation of the model prediction. The smaller the value, the smaller the overall deviation of the model prediction. R2 focuses on the model's ability to explain the fluctuations (variations) in spectrum data in spectrum occupancy prediction. The closer to 1, the better the model can explain the changes in spectrum data; the closer to 0, the weaker the explanation ability, and it may even be worse than directly predicting the mean. These indicators provide a comprehensive view of model performance for evaluating the prediction effect in the entire time period from the last model update to the current update. Ideally, MSE and MAE are as small as possible (accurate prediction), and R2 is as close to 1 as possible (the model explains the data fluctuations well). If the performance indicators decrease continuously for N iterations and exceed the set threshold, the system will automatically roll back to the previous version before the current performance decreased, which is a stable model that has been verified and has reliable performance, to ensure the robustness and reliability of the prediction system.
[0083] The radio spectrum occupancy prediction method of the present application significantly reduces the time consumption of incremental learning compared to traditional batch learning while ensuring prediction accuracy, solving the problem of decreased prediction performance of traditional static models in dynamic electromagnetic spectrum environments. In addition, the complexity and resource consumption of model updating are significantly reduced, and the complexity of spectrum occupancy prediction is reduced.
[0084] The above description is only the preferred embodiments of the present application. It should be noted that for those skilled in the art, without departing from the technical principles of the present application, several improvements and modifications can be made, and these improvements and modifications should also be considered within the scope of protection of the present application.
Claims
1. A method for electromagnetic spectrum occupancy prediction based on incremental learning, characterized in that, The method comprises the following steps: S1, time series dataset preprocessing The total amount of spectrum data is evaluated and determined, and the historical wireless signal spectrum dataset is systematically preprocessed and divided to obtain a basic training set and a basic validation set; according to different incremental time scales, the newly arrived spectrum data collected in time sequence in recent times is taken as an incremental dataset, and is divided into an incremental training set, an incremental validation set and an incremental test set; S2, basic prediction model training The basic prediction model is composed of a bidirectional Bi-LSTM and a self-attention mechanism structure, and is used for predicting spectrum occupancy; the model is first trained on the basic dataset to preliminarily learn the basic features and dependency relationships in the data, and is verified by the basic validation set; after the end of this stage, the basic prediction model obtains the initial prediction ability for the basic mode of the data; S3, incremental prediction model training The incremental validation set at each time is input into the existing basic prediction model for verification, and then, using the incremental training set, the hyperparameters of the working basic prediction model are adjusted to obtain an updated incremental prediction model; the performance of the updated incremental prediction model is evaluated using the incremental test set; S4, adaptive iteration Every time span, the spectrum data is collected as an incremental dataset, the working prediction model is updated, the performance of the existing model is evaluated and verified using the latest incremental dataset, and the model is ensured to maintain high-performance spectrum occupancy prediction in the continuous spectrum monitoring data; The updated model is deployed into a production environment to continue to serve new requests, and at this time the model should better adapt to the new data distribution.
2. The incremental learning based electromagnetic spectrum occupancy prediction method of claim 1, wherein, In S1, the total amount of spectrum data is evaluated and determined according to the frequency range, resolution bandwidth and monitoring time parameters to be monitored.
3. The incremental learning based electromagnetic spectrum occupancy prediction method of claim 1, wherein, In S2, the bidirectional Bi-LSTM layer is the core time sequence feature extractor of the model, and the self-attention mechanism structure dynamically adjusts the weight of the feature sequence output by the Bi-LSTM, so that the model can automatically focus on the most critical features for the current spectrum occupancy prediction task.
4. The incremental learning based electromagnetic spectrum occupancy prediction method of claim 3, wherein, The basic training set is input into the Bi-LSTM layer; the forward LSTM of the Bi-LSTM layer processes the input sequence in time sequence to capture the spectrum occupancy dependency relationship from history to future; The backward LSTM processes the sequence in reverse order to extract spectrum occupancy information from future to history, which is helpful to identify the influence of potential patterns or periodic features in the future on the current state; for each spectrum amplitude, the forward and backward hidden states are spliced to form a hidden state vector that integrates complete bidirectional context information.
5. The incremental learning based electromagnetic spectrum occupancy prediction method of claim 4, wherein, The hidden state vector processed by Bi-LSTM is sent to the self-attention mechanism layer, which dynamically allocates weights by calculating the correlation between elements within the sequence. The output of Bi-LSTM is transformed through a series of changes to obtain a set of attention weights, which quantify the importance of each spectral amplitude in the spectral sequence to all other spectral amplitudes. This allows the prediction model to dynamically focus on the most relevant historical or future spectral amplitudes for the current prediction, suppress noise interference, and more accurately capture long-term dependencies and key patterns, thereby improving the accuracy of spectral occupancy prediction. The feature representation weighted by self-attention is then passed to the fully connected layer to perform the regression prediction task of spectral occupancy. Finally, the output layer produces continuous prediction values representing the estimate of future spectral occupancy. The model effect is verified using the basic validation set to confirm whether it meets the expected training target.
6. The incremental learning based electromagnetic spectrum occupancy prediction method of claim 1, wherein, In S3, 10-15% of the incremental data set is taken as the incremental validation set, which is input to the basic prediction model for spectral occupancy prediction effect verification to monitor the performance of the existing model on incremental spectral data. If the spectral occupancy prediction accuracy is high and stable, the existing model is retained and not updated, and the next cycle of monitoring continues. If the performance of the basic prediction model on the first incremental data is significantly worse than its performance on the basic validation set, there is a data drift between the data distribution of the newly arrived spectral data and the old spectral data, triggering the model update process.
7. The incremental learning based electromagnetic spectrum occupancy prediction method of claim 6, wherein, In S3, after triggering the model update, the hyperparameters of the existing model in operation are adjusted using the incremental training set. When updating the model, only the newly arrived spectral data in the current time interval is used for training, so that new knowledge can be integrated into the existing model. Because only the changing part of the new data is involved, the above process is repeated periodically. Every time a set time interval passes, the system automatically checks the performance of the existing model and uses the new spectral data in this period to implement incremental updates when necessary, so that the model can continuously track the dynamic changes in spectral data.
8. The incremental learning based electromagnetic spectrum occupancy prediction method of claim 7, wherein, In S3, part of the incremental data set is used as an incremental test set to estimate the performance of the updated incremental model. The performance of the model is measured by three core evaluation indicators to evaluate the spectral occupancy prediction effect from the last prediction model update to the current update. The three core evaluation indicators are mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R2). Ideally, MSE and MAE are as small as possible, and R2 is as close to 1 as possible. If the above performance indicators have decreased continuously for N times and exceed the set threshold, the system will automatically roll back to the version before the current performance decline, which is a stable model that has been verified and has reliable performance, to ensure the robustness and reliability of the prediction system.
9. The incremental learning based electromagnetic spectrum occupancy prediction method of claim 8, wherein, 15-30% of the incremental data set is input to the trained incremental model as an incremental test set to evaluate the spectral occupancy prediction results based on the predicted values and true values of each frequency point occupancy. where n is the number of samples, y i is the true value, is the predicted value, is the mean of the true values.
10. The incremental learning based electromagnetic spectrum occupancy prediction method of claim 8, wherein, MSE focuses on the high sensitivity to large prediction errors in spectrum occupancy prediction, the smaller the value, the more accurate the model prediction; MAE focuses on the average level of all prediction errors in spectrum occupancy prediction, reflecting the overall deviation of the model prediction, the smaller the value, the smaller the overall deviation of the model prediction; R 2 In spectrum occupancy prediction, the model focuses on the ability to explain the fluctuations in spectrum data, the closer to 1, the better the model can explain the changes in spectrum data; the closer to 0, the weaker the ability to explain, and it may even be worse than directly predicting the mean; if MSE is very high, but MAE is relatively low, it means that the model is mostly correct, but there are a few points with very large prediction errors; if MSE and MAE are both very high, it means that the overall prediction deviation of the model is large.