Modulation method for microwave communication automatic optimization based on deep learning

Abstract

The invention relates to the technical field of microwave communication, and discloses a microwave communication automatic optimization modulation method based on deep learning, which comprises the following steps of: firstly, transmitting a signal to receiving end equipment by a microwave equipment transmitting end, taking microwave data received by a receiving end as input data, performing initial QAM (Quadrature Amplitude Modulation) grading on the input historical operation data, and then performing data processing on the input historical operation data; the input historical operation data specifically comprises time sequence information of maintenance time, success records and change times; performing data processing and calculation, specifically including classification feature coding, time sequence feature extraction and feature splicing; performing initial probability distribution prediction on the data features; carrying out QAM level selection and feedback updating; and taking the updated historical data and the new operation state as input, repeating the processes of feature engineering, probability prediction, probability adjustment and grade selection, and realizing dynamic adaptive optimization of the QAM modulation grade. According to the method, intelligent switching of modulation levels is realized by automatically and dynamically adjusting the selection probability of a QAM (Quadrature Amplitude Modulation) mode.

Description

A Deep Learning-Based Automatic Optimization Modulation Method for Microwave Communication Technical Field

[0001] This invention relates to the field of microwave communication technology, specifically to a modulation method for automatic optimization of microwave communication based on deep learning. Background Technology

[0002] With the widespread application of microwave communication in backbone networks, metropolitan area networks, and access networks in remote areas, its transmission scenarios are gradually shifting from "static and stable environments" to "dynamic and complex environments." For example, metropolitan microwave links need to cope with sudden factors such as building obstruction and electromagnetic interference, while links in remote areas need to withstand natural interference such as rain attenuation and sandstorms. Such interference can cause a sharp drop in signal-to-noise ratio and a surge in bit error rate. If the QAM modulation method cannot adapt in time, it will directly affect the transmission quality of voice, video, and other services. To solve this problem, the industry generally adopts "adaptive QAM modulation technology based on real-time status." The core idea is to dynamically adjust the modulation level by monitoring key indicators of the communication link (such as duration, bit error rate, and success rate) to seek a balance between "transmission efficiency" and "stability."

[0003] Traditional methods rely on a single static parameter, "distance," which cannot cope with sudden interference that is not related to distance (such as electromagnetic blockage). The distance-level mapping is fixed, resulting in a lack of flexibility in anti-interference. They are based solely on "instantaneous channel estimation" and lack the ability to optimize over a long period of time using historical data.

[0004] Defect 2: Rigid time slot-threshold determination, lacking long-term scenario memory. It relies solely on a short-term rule of "instantaneous channel quality in a single time slot → fixed QAM level," without incorporating a dynamic adjustment strategy based on historical states. This results in an urgent switch to 16QAM in the next time slot, leading to frequent adjustment losses.

[0005] Therefore, to address the above problems, a modulation method for automatic optimization of microwave communication based on deep learning is needed. Summary of the Invention

[0006] The purpose of this invention is to provide a modulation method for automatic optimization in microwave communication based on deep learning. When a transmission error (abnormal mean square error MSE) is detected by the receiving device (such as a microwave relay station or base station receiving unit), this invention achieves intelligent switching of modulation levels by automatically and dynamically adjusting the selection probability of the QAM (quadrature amplitude modulation) mode.

[0007] This invention is implemented as follows:

[0008] This invention provides a modulation method for automatic optimization in microwave communication based on deep learning, which is specifically implemented according to the following steps:

[0009] S1: The microwave equipment transmitter transmits signals to the receiver, using the microwave data received by the receiver as input data. The input historical operation data is then initialized with QAM level. The input historical operation data specifically includes timing information such as duration, success record, and number of changes.

[0010] S2: Perform data processing and calculations, specifically including classification feature encoding, temporal feature extraction, and feature concatenation; this is done in the following steps:

[0011] S2.1: Perform one-hot encoding on historical operational data using QAM levels to generate a 13-dimensional binary vector: if the current level is k, then the k-th bit is 1, and the rest are 0, denoted as .

[0012] S2.2: Perform time-series feature extraction, extracting 5 key features from historical data to form a time-series vector. Specifically, it includes:

[0013] Duration (Current level duration, in seconds);

[0014] Mean duration t － m (average duration of historical levels).

[0015] Success rate rs (number of successful modulations at the current level / total number of times);

[0016] Recent Changes (nc) (Number of level changes in the last 30 seconds);

[0017] Time since startup (ts) (cumulative time since device startup);

[0018] S2.3: Perform feature concatenation by horizontally concatenating the one-hot encoded vector and the temporal vector to form 18-dimensional model input features. ;

[0019] S2.4: Output the feature results of the 18-dimensional standardized feature vector.

[0020] S3: Perform initial probability distribution prediction on the data features; specifically, follow these steps:

[0021] S3.1: Input 18-dimensional feature vector X data, call the 3-layer fully connected network model QAMProbabilityNet, the network structure is configured according to prob_config_dl.py: input layer 18-dimensional → hidden layer 1 (64-dimensional, ReLU activation) → hidden layer 2 (64-dimensional, ReLU activation) → hidden layer 3 (32-dimensional, ReLU activation) → output layer (13-dimensional).

[0022] S3.2: Perform probability transformation, mapping the 13 inactive values ​​z[0], z[1], ..., z

[12] of the output layer to a probability distribution using the Softmax function, as shown in the following formula:

[0023]

[0024] in, For the output layer The number of neurons with inactive values , For the weight vector, For bias terms; Let be the initial predicted probability for the i-th QAM level;

[0025] S3.3: Output the 13-dimensional initial probability distribution; as shown in the following formula:

[0026] .

[0027] Then, adaptive adjustment of the probability distribution is performed:

[0028] S3.4: Input Initial Probability Distribution Duration of the current level ;

[0029] S3.5: Perform data processing and calculations, determine intervals, and based on... Divide the system into four adjustment intervals and match them with different enhancement strategies;

[0030] S3.6: Perform probability enhancement, assigning enhancement coefficients (coeff) to each level according to the interval strategy. i The adjusted probability is calculated as follows:

[0031]

[0032] S3.7: Perform normalization to ensure that the sum of the adjusted probabilities is 1, as shown in the following formula:

[0033]

[0034] S3.8: Output the final 13-dimensional probability distribution result The result satisfies the business logic of the current level having the highest probability, the nearest being larger and the farthest being smaller, and the left being higher and the right being lower.

[0035] S4: QAM Level Selection and Feedback Update;

[0036] Follow these steps:

[0037] S4.1: Input the final probability distribution data ;

[0038] S4.2: Use polynomial sampling torch.multinomial from The next QAM level is randomly selected.

[0039] S4.3: After modulation is performed, success or failure is determined based on the MSE value, where... If successful, otherwise fail; update historical data.

[0040] If the modulation is successful, the duration will be [duration]. Cumulative, number of changes ;

[0041] If it fails, the current level duration and number of changes will be reset. .

[0042] By taking the updated historical data and the new operating status as input, and repeating the feature engineering → probability prediction → probability adjustment → level selection process, dynamic adaptive optimization of QAM modulation level is achieved.

[0043] S4.4: Output the result of the next round of QAM modulation levels.

[0044] S5: Using the updated historical data and the new operating status as input, repeat the feature engineering → probability prediction → probability adjustment → level selection process to achieve dynamic adaptive optimization of QAM modulation level.

[0045] Furthermore, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the deep learning-based microwave communication automatic optimization modulation method described in any one of the above-mentioned methods.

[0046] Furthermore, the present invention provides a computer-storable medium storing a computer program, wherein when the program is executed, it sequentially executes any one of the above-described methods for automatic optimization of microwave communication based on deep learning.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] 1. This invention achieves intelligent switching of modulation levels by automatically and dynamically adjusting the selection probability of the QAM (Quadrature Amplitude Modulation) mode when a transmission error (abnormal mean square error, MSE) is detected by the receiving equipment (such as a microwave relay station or base station receiving unit). By optimizing the "generation logic of modulation mode selection probability", the receiving equipment can more quickly lock onto the optimal QAM level that balances anti-interference capability and transmission efficiency, ensuring stable operation of the link in complex environments.

[0049] 2. Breaking through the limitations of traditional algorithms that rely solely on maintenance time, this invention combines five time-series features—maintenance time, average maintenance time, success rate, number of recent changes, and time since startup—with QAM-level One-Hot encoding. Through a four-layer neural network learning the non-linear relationships between features, the model can automatically identify scenarios of "initial startup (time < 100 seconds) + recent changes > 5 times," outputting a more conservative probability distribution (increasing the probability of low-level interference by 15%-20%). This invention, through "early stopping on the validation set (avoiding overfitting) + hyperparameter polling (selecting optimal parameters such as sigma_low=3.5, boost_1=2.5)," accurately captures the pattern of "short maintenance time + low success rate → high probability of low-level interference" in strong interference scenarios, improving accuracy to 45.13%, meeting the core requirement of microwave communication for "rapid degradation of interference scenarios." Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0051] Figure 1 is a flowchart of the method of the present invention;

[0052] Figure 2 is a flowchart illustrating the execution of a specific embodiment of the present invention;

[0053] Figure 3 is a screenshot of the test code of the present invention, which narrows the range of the last traversal based on the changing trend after several rounds of traversal.

[0054] Figure 4 is a screenshot of the test code of the present invention, which narrows the range of the last traversal based on the changing trend after several rounds of traversal.

[0055] Figure 5 is a screenshot of the test code of the present invention, which narrows the range of the last traversal based on the changing trend after several rounds of traversal.

[0056] Figure 6 is a comparison chart of the test results of the conventional method of the present invention and the present invention;

[0057] Figure 7 is a comparison chart of the offline test results of the conventional method of the present invention and the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to describe selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Please refer to Figures 1 and 2. This invention provides a modulation method for automatic optimization of microwave communication based on deep learning, which is specifically implemented according to the following steps:

[0060] S1: The microwave equipment transmitter transmits signals to the receiver, using the microwave data received by the receiver as input data. The input historical operation data is then initialized with QAM level. The input historical operation data specifically includes timing information such as duration, success record, and number of changes.

[0061] S2: Perform data processing and calculations, specifically including classification feature encoding, temporal feature extraction, and feature concatenation; this is done in the following steps:

[0062] S2.1: Perform one-hot encoding on historical operational data using QAM levels to generate a 13-dimensional binary vector: if the current level is k, then the k-th bit is 1, and the rest are 0, denoted as . ;

[0063] S2.2: Perform time-series feature extraction, extracting 5 key features from historical data to form a time-series vector. Specifically, it includes:

[0064] Duration t m (Current level duration, in seconds);

[0065] Mean duration t － m (average duration of historical levels).

[0066] Success rate rs (number of successful modulations at the current level / total number of times);

[0067] Recent Changes (nc) (Number of level changes in the last 30 seconds);

[0068] Time since startup (ts) (cumulative time since device startup);

[0069] S2.3: Perform feature concatenation by horizontally concatenating the one-hot encoded vector and the temporal vector to form 18-dimensional model input features. ;

[0070] S2.4: Output the feature results of the 18-dimensional standardized feature vector.

[0071] S3: Perform initial probability distribution prediction on the data features; specifically, follow these steps:

[0072] S3.1: Input 18-dimensional feature vector X data, call the 3-layer fully connected network model QAMProbabilityNet, the network structure is configured according to prob_config_dl.py: input layer 18-dimensional → hidden layer 1 (64-dimensional, ReLU activation) → hidden layer 2 (64-dimensional, ReLU activation) → hidden layer 3 (32-dimensional, ReLU activation) → output layer (13-dimensional).

[0073] S3.2: Perform probability transformation, mapping the 13 inactive values ​​z[0], z[1], ..., z

[12] of the output layer to a probability distribution using the Softmax function, as shown in the following formula:

[0074]

[0075] in, The inactive value of the i-th neuron in the output layer , For the weight vector, For bias terms; Let be the initial predicted probability for the i-th QAM level;

[0076] S3.3: Output the 13-dimensional initial probability distribution; as shown in the following formula:

[0077] .

[0078] Then, adaptive adjustment of the probability distribution is performed:

[0079] S3.4: Input Initial Probability Distribution Duration of the current level ;

[0080] S3.5: Perform data processing and calculations, determine intervals, and base decisions on t. m Divide the system into four adjustment intervals and match them with different enhancement strategies;

[0081] S3.6: Perform probability enhancement, assigning enhancement coefficients (coeff) to each level according to the interval strategy. i The adjusted probability is calculated as follows:

[0082]

[0083] S3.7: Perform normalization to ensure that the sum of the adjusted probabilities is 1, as shown in the following formula:

[0084]

[0085] S3.8: Output the final 13-dimensional probability distribution result The model aims to achieve the business logic of maximizing the probability of the current level, prioritizing closer levels over farther ones, and prioritizing left-hand higher levels over right-hand lower levels. During model training, 20,000 simulated training samples (covering all QAM levels and scenarios) were generated using QAMDataset, divided into a 70% training set, a 15% validation set, and a 15% test set. The model employed the KL divergence loss function (measuring the difference between predicted and target probabilities), the Adam optimizer (learning rate 0.001), and an early stopping mechanism (patience value 15) to prevent overfitting.

[0086] Sample generation logic: Each sample contains "18-dimensional input features + 13-dimensional target probability distribution":

[0087] Input features: Randomly generated current QAM level (0-12), duration (exponential distribution, average 20 seconds), success rate (30%-90% uniform distribution), number of recent changes (0-9 times), and time since startup (0-1000 seconds);

[0088] Target probability distribution: Generated based on a "conservative strategy + multi-scenario adaptation"—using an asymmetric distribution skewed towards lower levels, combined with enhancements based on maintenance time and success rate tiers.

[0089] Duration > 30 seconds (stable period): Current level probability × 2.2, current-1 level × 1.8;

[0090] Duration 10-30 seconds (interference attenuation period): Current level × 1.5, current-1 × 1.3;

[0091] Duration of interference: 2-10 seconds: current-1 level increases linearly (from 1.2 to 2.4).

[0092] Duration 0-2 seconds (initial stage of an outbreak): current-1×1.2, current-2×1.08;

[0093] Success rate >70%: current-1×1.8, current-2×1.6, further enhancing the probability of low-level selection.

[0094] S4: QAM Level Selection and Feedback Update; Follow these steps for details:

[0095] S4.1: Input the final probability distribution data ;

[0096] S4.2: Use polynomial sampling torch.multinomial from The next QAM level is randomly selected.

[0097] S4.3: After modulation is performed, success or failure is determined based on the MSE value, where... If successful, otherwise fail; update historical data.

[0098] If the modulation is successful, the duration will be [duration]. Cumulative, number of changes ;

[0099] If it fails, the current level duration and number of changes will be reset. .

[0100] By taking the updated historical data and the new operating status as input, and repeating the feature engineering → probability prediction → probability adjustment → level selection process, dynamic adaptive optimization of QAM modulation level is achieved.

[0101] S4.4: Output the result of the next round of QAM modulation levels.

[0102] S5: Using the updated historical data and the new operating status as input, repeat the feature engineering → probability prediction → probability adjustment → level selection process to achieve dynamic adaptive optimization of QAM modulation level.

[0103] In this embodiment, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the deep learning-based microwave communication automatic optimization modulation method described in any one of the above-mentioned embodiments.

[0104] In this embodiment, the present invention provides a computer-storable medium storing a computer program, wherein when the program is executed, it sequentially executes any one of the above-described methods for automatic optimization of microwave communication based on deep learning.

[0105] In this embodiment, the present invention takes the current QAM level as the center and, according to the laws of the actual scenario, uses a "normal distribution with differentiated decay on the left and right sides" to generate the selection probability. A large standard deviation is set for lower levels (sigma_left=2.0), and the probability decays slowly; a small standard deviation is set for higher levels (sigma_right=1.0), and the probability decays quickly. The probability of the current level is simply extracted by ratio=1.0 / (maintain_time+1.0), without any other dynamic optimization. The code does not involve model training or multi-feature fusion, and only relies on hard-coded normal distribution parameters.

[0106] Figures 6-7: As shown in Figure 6, Method 3: Quick Test (Fixed Scenario: Initial Level 5, Optimal Level 3, 100 Iterations)

[0107] Conclusion: This invention can hit the optimal level faster, reduce invalid exploration, and significantly improve the speed of sudden response.

[0108] As shown in Figure 7, Method 2: Statistical significance test (30 independent trials, 50 iterations per iteration)

[0109] Conclusion: The paired t-test results show that the present invention is superior, and the p-value is <0.05, indicating that the performance difference is statistically significant. At a 95% confidence level, the modulation accuracy of the present invention is significantly higher than that of the traditional algorithm, and the stability is better (the standard deviation is reduced).

[0110] As shown in Figure 7, Method 1: Offline evaluation (50 random scenarios, covering different initial levels, optimal levels, and success thresholds)

[0111] Conclusion: This invention still possesses significant advantages in complex and ever-changing scenarios. Compared to traditional algorithms, the average convergence speed is significantly improved, enabling faster adaptation to matching requirements of different initial and optimal levels; the average accuracy across all scenarios is improved, fully validating its robustness against sudden interference. A balance between stability and efficiency is achieved.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations will be apparent to those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A modulation method for automatic optimization of microwave communication based on deep learning, characterized in that: The specific steps are as follows: S1: The microwave equipment transmitter transmits signals to the receiver, using the received microwave data as input data. The historical operational data is used to establish the initial QAM level. This historical operational data includes time-series information such as duration, successful records, and number of changes. S2: Data processing and calculation are performed, including classification feature encoding, time-series feature extraction, and feature concatenation. S3: Initial probability distribution prediction is performed on the data features. S4: QAM level selection and feedback update. S5: The updated historical data and new operational status are used as input, and the feature engineering → probability prediction → probability adjustment → level selection process is repeated to achieve dynamic adaptive optimization of the QAM modulation level.

2. The modulation method for automatic optimization of microwave communication based on deep learning according to claim 1, characterized in that: In step S2, the following steps are specifically performed: S2.1: Perform one-hot encoding on the historical running data using QAM levels to generate a 13-dimensional binary vector: If the current level is k, then the k-th bit is 1, and the rest are 0, denoted as ； S2.2: Perform time-series feature extraction, extracting 5 key features from historical data to form a time-series vector. Specifically, this includes: duration t m Mean duration t － m; success rate rs; number of recent changes nc; time since startup ts; S2.3: Perform feature concatenation by horizontally concatenating the one-hot encoded vector and the temporal vector to form 18-dimensional model input features. S2.4: Output the feature results of the 18-dimensional standardized feature vector.

3. The modulation method for automatic optimization of microwave communication based on deep learning according to claim 1, characterized in that: In step S3, the following steps are performed: S3.1: Input 18-dimensional feature vector X data, call the 3-layer fully connected network model QAMProbabilityNet, the network structure is configured according to prob_config_dl.py: 18-dimensional input layer → hidden layer 1 → hidden layer 2 → hidden layer 3 → output layer; S3.2: Perform probability transformation, map the 13 inactive values ​​z[0], z[1], ..., z[12] of the output layer to a probability distribution through the Softmax function, as shown in the following formula: ;in, The inactive value of the i-th neuron in the output layer ， For the weight vector, For bias terms; S3.3: Output the 13-dimensional initial probability distribution result; as shown in the following formula: 。 4. The modulation method for automatic optimization of microwave communication based on deep learning according to claim 3, characterized in that: Then perform adaptive adjustment of the probability distribution: S3.4: Input the initial probability distribution Duration of the current level S3.5: Perform data processing and calculations, determine intervals, and based on... Divide the system into four adjustment intervals and match different enhancement strategies; S3.6: Perform probability enhancement, and assign enhancement coefficients (coeff) to each level according to the interval strategy. i The adjusted probability is calculated as follows: S3.7: Perform normalization to ensure that the sum of the adjusted probabilities is 1, as shown in the following formula: S3.8: Output the final 13-dimensional probability distribution result. The result satisfies the business logic of the current level having the highest probability, the nearest being larger and the farthest being smaller, and the left being higher and the right being lower.

5. The modulation method for automatic optimization of microwave communication based on deep learning according to claim 1, characterized in that: In step S4, the following steps are performed: S4.1: Input the final probability distribution data. ； S4.2: Employing polynomial sampling from Randomly select the next QAM level; S4.3: After modulation is performed, success or failure is determined based on the MSE value, where... If successful, otherwise fail; update historical data; S4.4: Output the result of the next round of QAM modulation levels.

6. The modulation method for automatic optimization of microwave communication based on deep learning according to claim 5, characterized in that: In step S4.3, after modulation is performed, if successful, the duration is... Cumulative, number of changes If it fails, the current level duration and number of changes will be reset. 。 7. The modulation method for automatic optimization of microwave communication based on deep learning according to claim 6, characterized in that: By taking the updated historical data and the new operating status as input, and repeating the feature engineering → probability prediction → probability adjustment → level selection process, dynamic adaptive optimization of QAM modulation level is achieved.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the modulation method for automatic optimization of microwave communication based on deep learning, as described in any one of claims 1-6.

9. A computer-storable medium storing a computer program therein, characterized in that: When the program is executed, it sequentially executes any one of the deep learning-based microwave communication automatic optimization modulation methods described in claims 1-6 above.

Images