Federal transfer learning driven multi-scenario communication parameter optimization system and method

Abstract

This invention relates to the field of wireless communication indicator monitoring technology, and discloses a federated transfer learning-driven multi-scenario communication parameter optimization system and method, including a data acquisition module, a feature extraction and matching module, a federated learning model module, an encrypted transmission module, and a cloud-local collaboration module. The data acquisition module is used to collect radio environment data in different scenarios in real time. The feature extraction and matching module uses a convolutional neural network to perform hierarchical feature extraction on the collected data and calculates the feature similarity between scenarios using a cosine similarity algorithm. The federated learning model module divides the communication parameter model into globally shared parameters and local scenario parameters, and uses a hierarchical aggregation strategy for model updates. The encrypted transmission module uses a homomorphic encryption algorithm to encrypt the transmitted data during the federated learning data interaction process and designs a dynamic update strategy for the encryption key. This invention can meet the security requirements of information transmission in multiple scenarios.

Description

Technical Field

[0001] This invention relates to the field of wireless communication indicator monitoring technology, specifically to a federated transfer learning-driven multi-scenario communication parameter optimization system and method. Background Technology

[0002] In today's digital age, wireless communication networks have become an indispensable critical infrastructure in modern society, and their performance directly affects the operational efficiency and development of various fields. Especially in special scenarios, such as emergency rescue, intelligent manufacturing, and film and television shooting, the reliability and efficiency of wireless networks play a decisive role.

[0003] Currently, common communication network optimization techniques mainly focus on resource allocation optimization (such as spectrum resource allocation and computing resource allocation), network topology optimization (such as base station location and routing selection), and communication protocol optimization (such as modem design and error correction coding design). However, traditional communication parameter optimization schemes reveal many limitations when facing diverse and complex scenarios. On the one hand, the communication environment varies greatly in different scenarios, including the type, intensity, and propagation characteristics of signal interference. Traditional solutions struggle to adapt quickly to these changes, leading to a decline in communication performance. For example, on film and television shooting sets, the presence of numerous lighting devices and wireless microphones generates complex frequency band interference patterns; in densely populated venues such as concerts, multipath fading is pronounced, posing significant challenges to communication. On the other hand, existing optimization methods often require extensive data sharing and centralized processing, which poses serious data security risks in today's increasingly data-privacy-critical environment. Especially in scenarios involving multi-organizational and multi-user collaboration, data privacy issues become a major obstacle to collaborative communication parameter optimization efforts. Summary of the Invention

[0004] The present invention aims to provide a federated transfer learning-driven multi-scenario communication parameter optimization system and method to address the shortcomings of existing communication parameter optimization technologies in terms of multi-scenario adaptability and data privacy protection.

[0005] To solve the above problems, the present invention adopts the following technical solution: Option 1: A multi-scenario communication parameter optimization system driven by federated transfer learning, including a data acquisition module, a feature extraction and matching module, a federated learning model module, an encrypted transmission module, and a cloud-local collaboration module; The data acquisition module is used to collect radio environment data in different scenarios in real time. The feature extraction and matching module uses a convolutional neural network to perform hierarchical feature extraction on the collected data, and calculates the feature similarity between scenes using a cosine similarity algorithm; The federated learning model module divides the communication parameter model into globally shared parameters and local scene parameters, and uses a hierarchical aggregation strategy to update the model; During the federated learning data interaction process, the encrypted transmission module uses a homomorphic encryption algorithm to encrypt the transmitted data and designs a dynamic update strategy for the encryption key. The cloud-local collaboration module enables local devices to collect interference samples in real time for parameter fine-tuning, while the cloud aggregates optimization results from various scenarios to continuously iterate the global model.

[0006] Beneficial effects: This invention's system architecture integrates functions across the entire process, from data acquisition and processing to model optimization and collaborative work, effectively addressing communication parameter optimization needs across multiple scenarios. The data acquisition module ensures comprehensive and accurate acquisition of scenario data; the feature extraction and matching module provides precise scenario association judgment criteria for transfer learning, improving the accuracy and effectiveness of parameter transfer; the federated learning model module's hierarchical aggregation strategy balances model universality and scenario specificity, enhancing model convergence speed and optimization performance; the encrypted transmission module safeguards data privacy and security during transmission and processing; and the cloud-local collaboration module constructs a continuously evolving optimization loop, enabling the system to adapt to new interference patterns and enhancing the robustness of the communication system.

[0007] This invention innovatively integrates federated learning and transfer learning technologies to construct a federated transfer learning-driven multi-scenario communication parameter optimization system and method. This enables rapid adaptive tuning of communication parameters under different radio environments, while ensuring data privacy and security during transmission and processing. It also improves network communication performance in diverse scenarios and meets the urgent needs of various industries for high-quality communication.

[0008] Preferably, in the feature extraction and matching module, when the feature similarity between the source scene and the target scene exceeds a threshold of 0.7, the transfer learning mechanism is automatically triggered.

[0009] Beneficial effects: A clearly defined similarity threshold provides a quantitative standard for initiating transfer learning, avoiding blind transfer and improving its accuracy and effectiveness. This allows the system to quickly transfer effective anti-interference parameters between suitable scenarios, reducing repetitive training in new scenarios and significantly improving parameter adaptation efficiency. For example, in film and television production, if the similarity of interference modes between different studios exceeds 0.7, parameter transfer can improve anti-interference efficiency by 60%, significantly reducing shooting interruptions caused by frequency band conflicts.

[0010] Preferably, when the federated learning model module performs weighted aggregation of global parameters, it assigns 1.5-1.8 times the weight to complex interference scenarios and 1.0 times the weight to ordinary scenarios; wherein, a complex interference scenario is defined as a scenario in which 20 or more wireless devices are used simultaneously, or the multipath fading coefficient is ≥1.8, or the 2.4G band interference duty cycle is ≥30%.

[0011] Beneficial effects: This weighted strategy fully considers the complexity differences of different scenarios, paying more attention to scenarios with complex interference, and ensuring that the core algorithm has good universality even in complex environments. In complex interference scenarios such as film and television bases with dense multi-device environments, assigning higher weights can better aggregate model updates for these scenarios, making the global model more adaptable to complex interference environments and improving the performance of communication systems in such scenarios.

[0012] Option 2: A federated transfer learning-driven multi-scenario communication parameter optimization method, employing the multi-scenario communication parameter optimization system described in Option 1, including the following steps: S1: Real-time acquisition of radio environment data in different scenarios; S2: A convolutional neural network is used to extract hierarchical features from the collected data, focusing on capturing key dimensions including signal strength fluctuations, spectrum occupancy patterns, and interference duration. S3: Calculate the feature similarity between the source scene and the target scene using the cosine similarity algorithm. When the similarity exceeds the 0.7 threshold, the transfer learning mechanism is automatically triggered to transfer the effective anti-interference parameters in the source scene. S4: The communication parameter model is divided into global shared parameters and local scene parameters. Global parameters are aggregated in the cloud through federated learning to update the models of each scene. When weighted fusion, more weight is given to scenes with complex interference. Local parameters are allowed to be adaptively adjusted by each scene node according to local interference characteristics and the encrypted parameter gradient is uploaded. S5: In federated learning data exchange, homomorphic encryption algorithm is used to encrypt the transmitted parameter gradients, and the encryption key is automatically changed every 10 model iterations. S6: Local devices collect interference samples in real time, fine-tune parameters based on the transferred model, and summarize the optimization results of various scenarios in the cloud. The global model is continuously iterated through federated transfer learning.

[0013] Beneficial effects: This invention details the complete process from data acquisition to model optimization and continuous iteration, demonstrating strong operability and systematicity. By extracting features from key dimensions, it accurately captures scene characteristics; strict similarity thresholds and transfer mechanisms ensure efficient parameter transfer; hierarchical model parameter processing strategies improve model convergence speed and optimization performance; encryption and key update strategies protect data privacy; and cloud-local collaborative iteration enables the system to continuously evolve and adapt to new interference environments.

[0014] Preferably, in step S3, the anti-interference parameters for migration include a frequency hopping sequence and a power control strategy; wherein, the frequency hopping sequence includes a frequency band of 2.4GHz-2.5GHz and a frequency hopping rate of 35-65 hops / second, and the power control strategy includes a dynamic power adjustment range of -10-20dBm and a power control threshold accuracy of ±0.1dBm.

[0015] Beneficial effects: Clearly defining the specific types of anti-interference parameters for transfer makes transfer learning more targeted and practical. Frequency hopping sequences and power control strategies are common and effective anti-interference methods. Transferring these parameters between similar scenarios can quickly improve the anti-interference capability of the target scenario.

[0016] Preferably, in step S5, Gaussian random noise is added to the gradient using differential privacy technology. The noise parameters satisfy the following: mean 0, variance 0.01-0.02, and the original information leakage rate of the gradient data after adding noise is ≤0.1%.

[0017] Beneficial effects: The combination of differential privacy technology and encrypted transmission further enhances data privacy protection capabilities. Even if some information is leaked during data transmission, the addition of random noise makes it difficult for attackers to reverse-engineer the original data through model updates. This provides more comprehensive and reliable protection for data privacy in multi-organization, multi-user collaboration scenarios and promotes collaborative cooperation among different entities in optimizing communication parameters.

[0018] Preferably, when the local device performs parameter fine-tuning, it does so based on real-time interference samples of the signal-to-noise ratio and packet loss rate of the current frequency band; wherein, the signal-to-noise ratio includes: monitoring accuracy ±1dB and adjustment response time ≤10 seconds, and the packet loss rate includes: monitoring accuracy ±0.1% and adjustment response time ≤5 seconds; and after parameter fine-tuning, the signal-to-noise ratio increases by ≥3dB and the packet loss rate decreases by ≥2.5 percentage points.

[0019] Beneficial effects: Local equipment fine-tunes parameters based on real-time key interference indicators, enabling rapid response to real-time changes in the wireless environment, such as interference from temporarily intruding drones. This real-time adjustment mechanism ensures that the communication system maintains good communication performance when facing dynamically changing interference, improving the system's real-time adaptability and reliability.

[0020] Preferably, when the cloud summarizes the optimization results of each scenario, it forms a closed loop of "scenario data accumulation - model optimization - parameter distribution", with a closed loop period of ≤30 minutes: the scenario data accumulation frequency is ≥1 time / 5 seconds, the model iteration update time is ≤15 minutes, the parameter distribution delay is ≤10 seconds, and the accuracy of identifying new interference is improved by ≥32 percentage points after the closed loop has been running for 6 months.

[0021] Beneficial effects: The closed-loop mechanism enables the system to self-improve and continuously evolve. During filming, a production crew discovered a new type of microwave interference. Their locally optimized anti-interference parameters were encrypted and uploaded. The cloud-based model learned this characteristic and pushed it to all similar scenarios, achieving collective evolution of the anti-interference algorithm. This continuously improves the entire system's anti-interference capabilities in various scenarios, enhancing the robustness and stability of the communication system.

[0022] Preferably, the system provides scene feature similarity heatmaps and parameter optimization suggestions. The heatmap similarity display accuracy is ±0.01. The heatmap can display the similarity between the current environment and historical scenes in real time. The similarity range is 0-1. After the parameter set recommended based on similarity is reused, the on-site deployment time is ≤10 minutes.

[0023] Beneficial effects: The visual presentation and parameter optimization suggestions greatly improve the system's usability and on-site deployment efficiency. In critical shooting scenarios with high communication reliability requirements, such as drone tracking and synchronous sound recording, technicians can quickly determine the correlation between the current environment and historical scenes based on scene feature similarity heatmaps, and perform efficient parameter configuration with reference to parameter optimization suggestions, ensuring communication quality and reducing shooting errors caused by improper parameter configuration.

[0024] Preferably, the local processing of real-time interference response is: processing latency ≤ 10 seconds, hardware computing power requirement: GPU utilization rate < 40% or CPU utilization rate < 50%; the cloud processing of large-scale model training is: single training time ≤ 15 minutes, hardware configuration: ≥ 8 cores 32G server, and the computing power requirement of on-site equipment is reduced by ≥ 50%.

[0025] Beneficial effects: This flexible allocation strategy of computing resources is adaptable to hardware platforms with different cost levels. Small film crews can use lightweight local models and rely on cloud computing power to obtain the same anti-interference capabilities as large film crews, which reduces the technical threshold and deployment costs, enabling more scenarios to benefit from the optimization capabilities of this system and promoting the widespread application and promotion of the technology.

[0026] Working principle of the invention: The federated transfer learning-driven multi-scenario communication parameter optimization system and method of this invention is based on the organic combination of federated learning and transfer learning. First, the data acquisition module collects radio environment data in real time under different scenarios, covering multi-dimensional information such as signal strength fluctuations, spectrum occupancy patterns, and interference duration. Next, the feature extraction and matching module uses a convolutional neural network to perform deep hierarchical feature extraction on the collected data, transforming the complex radio environment data into representative feature vectors. The cosine similarity algorithm is used to calculate the feature similarity between the source and target scenarios. When the similarity exceeds a pre-set threshold of 0.7, the two scenarios are considered to have high similarity, automatically triggering the transfer learning mechanism. At this point, validated anti-interference parameters from the source scenario, such as frequency hopping sequences and power control strategies, are transferred to the target scenario, avoiding the tedious process of optimizing parameters from scratch in the target scenario and greatly improving parameter adaptation efficiency.

[0027] In the federated learning model module, the communication parameter model is divided into globally shared parameters and local scenario parameters. For globally shared parameters, model updates from various scenarios are aggregated in the cloud using federated learning. During the weighted fusion process, the complexity of interference in different scenarios is fully considered, and scenarios with complex interference are given higher weights to ensure that the core algorithm has good universality across various scenarios. For local scenario parameters, each scenario node is allowed to adaptively adjust based on its unique local interference characteristics, uploading only encrypted parameter gradients. This ensures that the model can adapt to local special conditions while protecting local data privacy.

[0028] In the data exchange phase of federated learning, the encrypted transmission module plays a crucial role. Homomorphic encryption is used to encrypt the transmitted parameter gradients, allowing model aggregation calculations to be performed on the data in its encrypted state without decryption, effectively preventing data leakage. Furthermore, a dynamic update strategy is designed to automatically change the encryption key every 10 model iterations, and differential privacy technology is combined to add random noise to the gradients, further enhancing the security of data privacy protection.

[0029] Finally, the cloud-local collaborative module constructs a continuously optimized closed loop. Local devices utilize real-time collected interference samples, such as the signal-to-noise ratio and packet loss rate of the current frequency band, to fine-tune parameters based on the transferred model, enabling rapid response to real-time changes in the wireless environment. The cloud aggregates the optimization results from various scenarios and continuously iterates the global model through federated transfer learning. When a new interference pattern is detected in a scenario, its locally optimized anti-interference parameters are encrypted and uploaded. The cloud model learns this feature and pushes it to all similar scenarios, achieving collective evolution of the anti-interference algorithm. This allows the entire system to continuously adapt to new interference environments and continuously improve communication performance.

[0030] Advantages of this invention: 1. Rapid Transfer of Anti-interference Capabilities Across Multiple Scenarios: Through precise scene feature matching and an efficient transfer learning mechanism, the optimization time for communication parameters in new scenarios is significantly reduced from several hours to within 10 minutes. This advantage is particularly pronounced in scenarios such as film and television shooting where locations are frequently changed and equipment is added or removed temporarily. When the similarity of interference patterns between different studios within the same film and television base exceeds 0.7, parameter transfer can improve anti-interference efficiency by 60%, greatly reducing shooting interruptions caused by frequency band conflicts and ensuring shooting progress and quality.

[0031] 2. Balancing Privacy Protection and Data Sharing: The application of homomorphic encryption and dynamic key mechanisms ensures that data is "usable but not visible" in various scenarios. This effectively addresses data privacy concerns when multiple production teams and organizations collaborate. Different film and television companies can jointly participate in anti-interference model training without disclosing their own shooting frequency strategies, forming an industry-level anti-interference knowledge base. Overall spectrum utilization is improved by 35%, achieving a perfect balance between data sharing and privacy protection, and promoting cross-institutional cooperation and technological innovation.

[0032] 3. A continuously evolving intelligent anti-interference system: The cloud-local collaborative mechanism provides strong support for the continuous iteration of anti-interference algorithms. New interference patterns, such as interference from novel wireless devices, can be quickly integrated into the global model through federated learning and then promptly distributed to all scenarios. According to actual tests, after six months of continuous operation, the system's accuracy in identifying novel interference significantly improved from the initial 60% to 92%, significantly enhancing the robustness and adaptability of the communication system, enabling it to cope with ever-changing and complex communication environments.

[0033] 4. Flexible allocation of computing resources: The local system focuses on real-time interference response, such as millisecond-level frequency switching, enabling rapid reactions to changes in the local wireless environment; the cloud handles large-scale model training, fully utilizing its powerful computing resources. This division of labor reduces the computing power requirements of on-site equipment and is compatible with hardware platforms of varying costs, from high-end software-defined radios to simple terminals. Small film crews can leverage lightweight local models and cloud computing power to achieve the same anti-interference capabilities as large film crews, lowering the technical threshold and deployment costs, and promoting the widespread application and popularization of the technology.

[0034] 5. Visualized Scene Adaptation and Decision Support: The system provides scene feature similarity heatmaps and parameter optimization suggestions, offering technicians intuitive and convenient decision-making support. Technicians can quickly grasp the correlation between the current environment and historical scenes, such as "the current frequency band interference mode is 82% similar to the 2024 X Concert, recommending the reuse of parameter set A." In critical shooting scenarios with high communication reliability requirements, such as drone tracking and synchronous sound recording, parameters can be configured quickly and accurately, improving on-site deployment efficiency, ensuring communication quality, and reducing shooting errors caused by improper parameter configuration. Attached Figure Description

[0035] Figure 1 This is a logical block diagram of the federated transfer learning-driven multi-scenario communication parameter optimization system of the present invention.

[0036] Figure 2 This is a flowchart of the multi-scenario communication parameter optimization method driven by federated transfer learning according to the present invention. Detailed Implementation

[0037] The following detailed description illustrates the specific implementation method: As attached Figure 1 As shown: A multi-scenario communication parameter optimization system driven by federated transfer learning, including a central processing unit and data acquisition module, feature extraction and matching module, federated learning model module, encrypted transmission module and cloud-local collaboration module respectively connected to the central processing unit; The data acquisition module is used to collect radio environment data in different scenarios in real time. The feature extraction and matching module uses a convolutional neural network to perform hierarchical feature extraction on the collected data, and calculates the feature similarity between scenes using a cosine similarity algorithm; The federated learning model module divides the communication parameter model into globally shared parameters and local scene parameters, and uses a hierarchical aggregation strategy to update the model; During the federated learning data interaction process, the encrypted transmission module uses a homomorphic encryption algorithm to encrypt the transmitted data and designs a dynamic update strategy for the encryption key. The cloud-local collaboration module enables local devices to collect interference samples in real time for parameter fine-tuning, while the cloud aggregates optimization results from various scenarios to continuously iterate the global model.

[0038] In the feature extraction and matching module, when the feature similarity between the source and target scenes exceeds a threshold of 0.7, a transfer learning mechanism is automatically triggered. This explicit similarity threshold provides a quantitative standard for initiating transfer learning, avoiding blind transfer and improving its accuracy and effectiveness. This allows the system to quickly transfer effective anti-interference parameters between suitable scenes, reducing repetitive training in new scenes and significantly improving parameter adaptation efficiency. For example, in film and television production, if the similarity of interference modes between different studios exceeds 0.7, parameter transfer can improve anti-interference efficiency by 60%, significantly reducing shooting interruptions caused by frequency band conflicts.

[0039] The federated learning model module assigns higher weights to scenarios with complex interference when weighting and aggregating global parameters. This weighting strategy fully considers the complexity differences between different scenarios, giving more attention to scenarios with complex interference, and ensuring that the core algorithm has good universality even in complex environments. In complex interference scenarios such as film and television bases with dense multi-device environments, assigning higher weights can better aggregate model updates for these scenarios, making the global model more adaptable to complex interference environments and improving the performance of the communication system in such scenarios.

[0040] like Figure 2 As shown, the federated transfer learning-driven multi-scenario communication parameter optimization method of the present invention, employing the above-mentioned multi-scenario communication parameter optimization system, includes the following steps: S1: Real-time acquisition of radio environment data in different scenarios; S2: A convolutional neural network is used to extract hierarchical features from the collected data, focusing on capturing key dimensions including signal strength fluctuations, spectrum occupancy patterns, and interference duration. S3: Calculate the feature similarity between the source scene and the target scene using the cosine similarity algorithm. When the similarity exceeds the 0.7 threshold, the transfer learning mechanism is automatically triggered to transfer the effective anti-interference parameters in the source scene. S4: The communication parameter model is divided into global shared parameters and local scene parameters. Global parameters are aggregated in the cloud through federated learning to update the models of each scene. When weighted fusion, more weight is given to scenes with complex interference. Local parameters are allowed to be adaptively adjusted by each scene node according to local interference characteristics and the encrypted parameter gradient is uploaded. S5: In federated learning data exchange, homomorphic encryption algorithm is used to encrypt the transmitted parameter gradients, and the encryption key is automatically changed every 10 model iterations. S6: Local devices collect interference samples in real time, fine-tune parameters based on the transferred model, and summarize the optimization results of various scenarios in the cloud. The global model is continuously iterated through federated transfer learning.

[0041] In step S3, the transferred anti-interference parameters include frequency hopping sequences and power control strategies. Clearly defining the specific types of anti-interference parameters for transfer makes the transfer learning more targeted and practical. Frequency hopping sequences and power control strategies are common and effective anti-interference methods; transferring these parameters between similar scenarios can quickly improve the anti-interference capability of the target scenario.

[0042] In step S5, random noise is added to the gradient using differential privacy technology. The combination of differential privacy and encrypted transmission further enhances data privacy protection. Even if some information is leaked during data transmission, the added random noise makes it difficult for attackers to deduce the original data through model updates. This provides more comprehensive and reliable protection for data privacy in multi-institutional and multi-user collaborations, promoting collaborative cooperation among different entities in optimizing communication parameters.

[0043] The local device performs parameter fine-tuning based on real-time interference samples, including the signal-to-noise ratio and packet loss rate of the current frequency band. This real-time adjustment mechanism, using key interference indicators, allows for rapid response to real-time changes in the wireless environment, such as interference from temporarily intruding drones. This ensures the communication system maintains good communication performance even in the face of dynamically changing interference, improving the system's real-time adaptability and reliability.

[0044] In this process, the cloud platform aggregates optimization results from various scenarios, forming a closed loop of "scenario data accumulation - model optimization - parameter distribution." This closed-loop mechanism enables the system to self-improve and continuously evolve. For example, when a film crew discovers a new type of microwave interference during filming, their locally optimized anti-interference parameters are encrypted and uploaded. The cloud model learns this characteristic and pushes it to all similar scenarios, achieving collective evolution of the anti-interference algorithm. This continuously improves the entire system's anti-interference capabilities in various scenarios, enhancing the robustness and stability of the communication system.

[0045] The system provides scene feature similarity heatmaps and parameter optimization suggestions, helping technicians intuitively understand the correlation between the current environment and historical scenes. The visual presentation and parameter optimization suggestions greatly improve the system's usability and on-site deployment efficiency. In critical shooting scenarios with high communication reliability requirements, such as drone tracking and synchronous sound recording, technicians can quickly determine the correlation between the current environment and historical scenes based on the scene feature similarity heatmaps, and efficiently configure parameters according to the optimization suggestions, ensuring communication quality and reducing shooting errors caused by improper parameter configuration.

[0046] The system employs a local processing mechanism for real-time interference response and a cloud-based approach for large-scale model training, reducing the computational demands on on-site equipment. This flexible allocation strategy of computational resources is adaptable to hardware platforms with varying cost profiles. Smaller production crews can utilize lightweight local models and leverage cloud computing power to achieve the same level of interference resistance as larger crews. This lowers the technical barrier and deployment costs, enabling more scenarios to benefit from the system's optimization capabilities and promoting the widespread application and adoption of the technology.

[0047] Specifically, Example 1 This example illustrates a 2.4G wireless listening device scenario at a film and television shooting base.

[0048] I. System Deployment Data Acquisition Module: In each of the 10 independent sound stages at the film and television shooting base, 6 software-defined radio sensors (model: SDRplay RSPdx) are deployed. The sensors support 2.4G band signal acquisition, with a sampling frequency of 1kHz and a data transmission rate of 10Mbps. They collect real-time radio environment data such as the signal strength of wireless listening devices (-90-0dBm), the duty cycle of 2.4G band interference generated by lighting equipment (20%-40%), and the duration of single interference (2-5 seconds), ensuring coverage of the communication area of ​​each listening device in the sound stage.

[0049] Feature extraction and matching module: Employs a 3-layer convolutional neural network (CNN). The input layer receives a 128×128 spectrum matrix, the convolutional layers use the ReLU activation function, and the fully connected layers output a 64-dimensional feature vector. The cosine similarity calculation module is integrated into the edge computing node (NVIDIA Jetson Nano) to compare the feature vectors of the current scene with historical scenes (storing data from 100+ film crew shooting scenes) in real time. When the feature similarity between the source scene and the target scene exceeds a threshold of 0.7, a transfer learning mechanism is automatically triggered. For example, if the similarity between a newly built soundstage and a historical film crew A's soundstage reaches 0.83, the parameter transfer process is immediately initiated.

[0050] The federated learning model module divides the communication parameter model into globally shared parameters (basic LS channel estimation algorithm) and local scene parameters (studio lighting interference compensation coefficient 1.2-1.5). When weighting and aggregating the global parameters, scenarios with complex interference (simultaneous use of 20 or more wireless eavesdroppers, multipath fading coefficient ≥1.8, 2.4GHz band interference duty cycle ≥30%) are assigned a weight of 1.5-1.8, while ordinary scenarios (not meeting any of the above conditions) are assigned a weight of 1.0. The cloud aggregation nodes use Alibaba Cloud ECS servers (8 cores, 32GB), and edge nodes are deployed in each studio, responsible for local parameter adjustment and gradient uploading.

[0051] Encrypted transmission module: The Paillier homomorphic encryption algorithm is used to encrypt the transmitted parameter gradients. The key length is 2048 bits, and the key is automatically updated every 10 model iterations. Gaussian random noise is added in combination with differential privacy technology. The noise parameters meet the requirements of mean 0 and variance 0.01-0.02. After the noise is added, the original information leakage rate of the gradient data is ≤0.1%, ensuring that the communication data of the eavesdropper is not leaked during transmission.

[0052] Cloud-Local Collaboration Module: Local devices (studio edge nodes) collect interference samples in real time for parameter fine-tuning. The cloud aggregates the optimization results from 10 studios to continuously iterate the global model, forming a closed loop of "scene data accumulation - model optimization - parameter distribution," with a loop cycle of ≤30 minutes. Specifically, the scene data accumulation frequency is ≥1 time / 5 seconds, the model iteration update time is ≤15 minutes, and the parameter distribution delay is ≤10 seconds. After 6 months of closed-loop operation, the accuracy rate for identifying new types of LED lighting interference increased from the initial 60% to ≥92% (an increase of ≥32 percentage points in accuracy).

[0053] II. Method Implementation Step S1 (Data Acquisition): The sensor collects radio environment data of the newly built photography studio (target scene), including the signal strength fluctuation of the wireless listening device by ±5dB, the interference intensity generated by the lighting equipment in the 2.45GHz band of the transmitter being +20dBm, and the signal-to-noise ratio dropping from 20dB to 12dB due to the superposition of multiple devices, and the packet loss rate increasing from 0.5% to 3%.

[0054] Step S2 (feature extraction): The 3-layer CNN performs hierarchical feature extraction on the collected data, focusing on capturing three key dimensions: “signal strength fluctuation frequency (0.5-2Hz), interference frequency band overlap (30%-50%), and multi-device interference superposition coefficient (1.2-1.8)”, and outputs the target scene feature vector V1.

[0055] Step S3 (Parameter Transfer): The similarity between V1 and the historical A film crew scene (source scene, feature vector V2) is calculated using the cosine similarity algorithm, resulting in a similarity of 0.83 (>0.7). The effective anti-interference parameters of film crew A are automatically transferred. Frequency hopping sequence: The frequency band covers the 2.4GHz-2.5GHz band (a total of 80 channels in the 2.4GHz-2.5GHz band; the output power range of our signal source is -10dBm to +20dBm, the output power range of the interference source is also -10dBm to +20dBm, and the power spectrum coordinate range of the detected software radio node is -90dBm to +0dBm), with a frequency hopping rate of 35-65 hops / second, ensuring that the dwell time of a single channel is shorter than the duration of the light interference (2-5 seconds), effectively avoiding interference; Power control strategy: Dynamic power adjustment range -10 to +20dBm, adapted to signal attenuation compensation at a communication distance of 50 meters in the studio, power control threshold accuracy ±0.1dBm, ensuring stable listening device signal.

[0056] Step S4 (Federated Model Optimization): Global parameters: The model updates from 10 film studios are aggregated in the cloud. The special effects film studio (with complex interference scenes) using 25 wireless listening devices simultaneously is given a weight of 1.6 times, while the ordinary film studio is given a weight of 1.0 times. The basic LS channel estimation algorithm is optimized, reducing the channel estimation error from 10% to 3%. Local parameters: Based on the local light interference characteristics (2.45GHz band interference duty cycle 35%), the light interference compensation coefficient of the target scene edge node is adjusted from 1.2 to 1.4, and the parameter gradient is uploaded after encryption (aggregation calculation is completed in encrypted state).

[0057] Step S5 (Encrypted Transmission): The parameter gradient is encrypted using the Paillier algorithm, and the key is automatically updated during the 10th iteration; Gaussian random noise (mean 0, variance 0.015) is added to prevent attackers from back-digging the original eavesdropper signal data through the gradient, and the privacy protection level meets the industry data security standards.

[0058] Step S6 (Parameter Fine-tuning and Iteration): Local fine-tuning: Parameters are adjusted based on real-time interference samples of the current frequency band's signal-to-noise ratio (SNR) and packet loss rate. Specifically, the SNR monitoring accuracy is ±1dB, with an adjustment response time ≤10 seconds; the packet loss rate monitoring accuracy is ±0.1%, with an adjustment response time ≤5 seconds. After fine-tuning, the SNR increases from 12dB to ≥15dB (an improvement of ≥3dB), and the packet loss rate decreases from 3% to ≤0.5% (a decrease of ≥2.5 percentage points). Cloud-based iteration: The optimization results of 10 studios are aggregated, the global model is iterated, and the optimization parameters of the target scene (frequency hopping rate of 45 hops / second, power control threshold of -55dBm) are distributed to 3 other studios with similarity >0.6 to achieve the reuse of anti-interference parameters.

[0059] Advantages of this embodiment: Parameter adaptability: The parameter settings of frequency hopping sequence and power control strategy fully cover the anti-interference requirements of 2.4G wireless monitoring devices in the studio. After parameter migration, the anti-interference efficiency is improved by 60%, and there is no shooting interruption due to frequency band conflict. Privacy protection: A dual protection mechanism of homomorphic encryption and differential privacy ensures that the listening device data of 10 film crews is "usable but not visible", and the overall spectrum utilization rate of the filming base is improved by 38%; Deployment efficiency: The system provides a scene feature similarity heatmap (similarity display accuracy ±0.01), and displays the similarity of the target scene with the scene of film crew A in real time (0.83). It is recommended to reuse "anti-interference parameter set A". The on-site deployment time for technicians is ≤10 minutes. Computing power optimization: Local processing of real-time interference response with processing latency ≤10 seconds and hardware computing power requirements reduced by ≥50% (edge ​​node GPU utilization <40% and CPU utilization <50%); cloud processing of large-scale model training with single training time ≤15 minutes and hardware configuration of ≥8 cores and 32G servers to adapt to deployment needs with different costs.

[0060] Example 2 This example demonstrates a 2.4G wireless mixing console scenario for a large-scale concert.

[0061] I. System Deployment Data Acquisition Module: Sensors are deployed in the concert venue (including the main stage area, audience area, and backstage area). Ten fixed sensors (model: NI USRP B210) are deployed in the main stage area, one mobile sensor is deployed for every 800 people in the audience area, and four sensors are deployed in the backstage area. All sensors support 2.4G frequency band signal acquisition, sampling frequency of 2kHz, data transmission rate of 20Mbps, and real-time acquisition of data such as wireless mixing console signal strength, multipath fading coefficient (1.2-1.8), and signal attenuation (5-10dB) caused by crowd obstruction.

[0062] Feature extraction and matching module: adopts a 4-layer CNN, the input layer receives a 256×256 spectrum matrix and outputs a 128-dimensional feature vector; similarity calculation module is deployed on the venue edge cloud (Intel Xeon E5 server), which compares the features of the current scene with those of historical concert scenes (stores data of 50+ large-scale events) in real time. When the similarity exceeds the threshold of 0.7, transfer learning is triggered. For example, when the similarity between the main stage area and the main stage area of ​​the historical B concert reaches 0.78, parameter transfer is initiated.

[0063] The federated learning model module uses a multipath fading compensation algorithm for globally shared parameters and region-specific anti-interference thresholds for local scene parameters (main stage area: -10 to +10 dBm; audience area: -5 to +5 dBm). The main stage area (with densely packed multiple devices, using 30+ wireless devices simultaneously) is assigned a weight of 1.5-1.8, while the audience and backstage areas are assigned a weight of 1.0. The cloud platform uses Tencent Cloud CVM servers (16 cores, 64GB), with edge nodes deployed in the main stage and backstage areas.

[0064] Encrypted transmission module: Employs the BFV homomorphic encryption algorithm with a 1024-bit key length, updating the key every 10 iterations; Gaussian noise (mean 0, variance 0.01-0.02) is added, ensuring an original information leakage rate of ≤0.1%, thus protecting the communication privacy of the company's wireless mixing console.

[0065] Cloud-Local Collaboration Module: Local edge server (2 Intel Xeon E5, CPU utilization <50%) handles real-time fine-tuning, cloud iterates the global model, closed-loop cycle is 25 minutes, and after 6 months the accuracy of identifying interference from new wireless headphones is improved by ≥35 percentage points (from 60% to 95%).

[0066] II. Method Implementation Step S1 (Data Acquisition): Acquire data from the main stage area. The signal strength of the wireless mixing console is -10dBm. The 2.46GHz band is subject to sudden interference from the wireless headphones (strength -55dBm, lasting 1-3 seconds). Multipath fading results in a signal-to-noise ratio of 12dB and a packet loss rate of 4%.

[0067] Step S2 (feature extraction): A 4-layer CNN extracts the features of "overlap of company equipment signal and interference spectrum (40%-60%), multipath fading duration (5-10 milliseconds), and burst interference frequency (2.46GHz±5MHz)" and outputs vector V3.

[0068] Step S3 (Parameter Transfer): Calculate the similarity between V3 and the historical B concert scene vector V4, which is 0.78 (>0.7). Transfer the anti-interference parameters: Frequency hopping sequence: frequency band 2.4GHz-2.5GHz, frequency hopping rate 35-65 hops / second, avoiding sudden interference for 1-3 seconds; Power control strategy: Dynamic power adjustment range -10 to +20dBm, compensating for signal attenuation within 30 meters, power control threshold accuracy ±0.1dBm.

[0069] Step S4 (Model Optimization): The cloud aggregates the updates of the three regional models, with the main stage region having a weight of 1.7. The multipath fading compensation algorithm is optimized, reducing the error from 8% to 2%.

[0070] Step S5 (encrypted transmission): The gradient is encrypted using the BFV algorithm. The key is updated in the 10th iteration, and Gaussian noise with a variance of 0.02 is added to prevent data leakage.

[0071] Step S6 (Collaborative Optimization): Local fine-tuning: Parameters are adjusted based on real-time interference samples of the current frequency band's signal-to-noise ratio (SNR) and packet loss rate. Specifically, the SNR monitoring accuracy is ±1dB with an adjustment response time of 8 seconds; the packet loss rate monitoring accuracy is ±0.1% with an adjustment response time of 4 seconds. Using the migrated anti-interference parameters, the frequency hopping rate is fine-tuned from the initial 35 hops / second to 55 hops / second within 5 seconds, increasing the SNR in the main stage area from 12dB to 16dB (an improvement of ≥4dB) and reducing the packet loss rate from 4% to 0.6% (a decrease of ≥3.4 percentage points), completely eliminating the impact of sudden interference from wireless headphones (lasting 1-3 seconds) on the mixing console communication.

[0072] Cloud-based iteration: The optimization results of the main stage area, audience area, and backstage area are aggregated, and a hierarchical aggregation strategy is used to iterate the global model—the parameter gradient of the main stage area (complex interference scenario) is assigned a weight of 1.7 times, while the audience area and backstage area are assigned a weight of 1.0 times, and the multipath fading compensation algorithm is updated. The optimized anti-interference parameters of the main stage area (frequency hopping sequence: 2MHz interval, 55 hops / second; power control strategy: -10-20dBm dynamic range, ±0.1dBm accuracy) are distributed to the audience area (scene similarity 0.65) and the backstage area (scene similarity 0.62) to achieve 2.4G wireless mixing console communication coverage throughout the venue.

[0073] Advantages of this embodiment: Interference resistance meets standards: migration parameters are adapted to the needs of concert scenarios, the 2.4GHz-2.5GHz frequency band, and the rate of 55 hops / second make the single-channel dwell time about 18.18 milliseconds, which can completely avoid 1-3 second level sudden interference; the power dynamic range of -10-20dBm compensates for the 5-10dB signal attenuation caused by crowd obstruction within 30 meters, and the threshold accuracy of ±0.1dBm ensures the stability of the mixing console signal (the number of device communication interruptions has been reduced from 6 times per day to 0 times, and the communication stability reaches 99.94%).

[0074] Privacy protection compliance: BFV homomorphic encryption (1024-bit key) + 10 iterations of key updates + Gaussian noise with a variance of 0.02 ensures that the leakage rate of raw information of mixing console data in the three areas is ≤0.1%. Privacy data such as the company's equipment UUID and communication frequency band policies are not leaked. The venue's 2.4G frequency band spectrum utilization rate is increased by 40%, achieving the collaborative goal of "data usable but invisible".

[0075] The closed-loop evolution is effective: the closed-loop cycle of "scene data accumulation (1 time / 5 seconds) - model optimization (≤15 minutes) - parameter delivery (≤8 seconds)" is 25 minutes. After running for 6 months, the recognition accuracy of interference from new wireless microphones (2.47GHz band) increased from 60% to 95% (an increase of 35 percentage points, exceeding the requirement of "≥32 percentage points" in claim 8), and the anti-interference algorithm continues to evolve.

[0076] High computing power adaptability: Local edge server (2 Intel Xeon E5) CPU utilization <50%, GPU utilization <40%, hardware cost about 4800 yuan (≤5000 yuan lightweight standard); cloud 16 core 64G server single model training time 12 minutes (≤15 minutes), small performance venues can obtain the same anti-interference capability as 10,000-person concerts by renting cloud computing power, reducing the technical threshold by 60%.

[0077] Highly efficient visualization deployment: The system generates a scene feature similarity heatmap (display accuracy ±0.01), and marks the similarity between the main stage area and the historical B concert scene as 0.78 (range 0-1) in real time. With the suggestion of "reusing anti-interference parameter set B", technicians can complete the parameter configuration of the entire venue within 10 minutes, avoiding sound mixing interruptions caused by parameter configuration errors (such as loss of the same-day sound recording signal).

[0078] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A federated transfer learning-driven multi-scenario communication parameter optimization system, characterized in that, It includes a data acquisition module, a feature extraction and matching module, a federated learning model module, an encrypted transmission module, and a cloud-local collaboration module; The data acquisition module is used to collect radio environment data in different scenarios in real time. The feature extraction and matching module uses a convolutional neural network to perform hierarchical feature extraction on the collected data, focusing on capturing key dimensions such as signal strength fluctuations, spectrum occupancy patterns, and interference duration. It also calculates the feature similarity between scenes using a cosine similarity algorithm. When the similarity exceeds a threshold of 0.7, a transfer learning mechanism is automatically triggered to transfer the source scene and verify the effective anti-interference parameters. The federated learning model module divides the communication parameter model into global shared parameters and local scene parameters, and adopts a hierarchical aggregation strategy for model updates. The global parameters are aggregated in the cloud through federated learning to update the models of each scene. When weighted fusion, higher weights are given to scenes with complex interference. For local parameters, each scene node is allowed to adaptively adjust according to local interference characteristics and upload the encrypted parameter gradient. During the federated learning data interaction process, the encrypted transmission module uses a homomorphic encryption algorithm to encrypt the transmitted data and designs a dynamic update strategy for the encryption key. The cloud-local collaboration module enables local devices to collect interference samples in real time for parameter fine-tuning, while the cloud aggregates optimization results from various scenarios to continuously iterate the global model.

2. The federated transfer learning-driven multi-scenario communication parameter optimization system according to claim 1, characterized in that, In the feature extraction and matching module, when the feature similarity between the source scene and the target scene exceeds a threshold of 0.7, the transfer learning mechanism is automatically triggered.

3. The federated transfer learning-driven multi-scenario communication parameter optimization system according to claim 1, characterized in that, When the federated learning model module performs weighted aggregation of global parameters, it assigns 1.5-1.8 times the weight to complex interference scenarios and 1.0 times the weight to ordinary scenarios. Among them, complex interference scenarios are defined as scenarios in which 20 or more wireless devices are used simultaneously, or where the multipath fading coefficient is ≥1.8, or where the 2.4G band interference duty cycle is ≥30%.

4. A federated transfer learning-driven method for optimizing communication parameters across multiple scenarios, characterized in that: The multi-scenario communication parameter optimization system according to claim 1 includes the following steps: S1: Real-time acquisition of radio environment data in different scenarios; S2: A convolutional neural network is used to extract hierarchical features from the collected data, focusing on capturing key dimensions including signal strength fluctuations, spectrum occupancy patterns, and interference duration. S3: Calculate the feature similarity between the source scene and the target scene using the cosine similarity algorithm. When the similarity exceeds the 0.7 threshold, the transfer learning mechanism is automatically triggered to transfer the effective anti-interference parameters in the source scene. S4: The communication parameter model is divided into global shared parameters and local scene parameters. Global parameters are aggregated in the cloud through federated learning to update the models of each scene. When weighted fusion, more weight is given to scenes with complex interference. Local parameters are allowed to be adaptively adjusted by each scene node according to local interference characteristics and the encrypted parameter gradient is uploaded. S5: In federated learning data exchange, homomorphic encryption algorithm is used to encrypt the transmitted parameter gradients, and the encryption key is automatically changed every 10 model iterations. S6: Local devices collect interference samples in real time, fine-tune parameters based on the transferred model, and summarize the optimization results of various scenarios in the cloud. The global model is continuously iterated through federated transfer learning.

5. The federated transfer learning-driven multi-scenario communication parameter optimization method according to claim 4, characterized in that, In step S3, the anti-interference parameters for migration include a frequency hopping sequence and a power control strategy; wherein, the frequency hopping sequence includes a frequency band of 2.4GHz-2.5GHz and a frequency hopping rate of 35-65 hops / second, and the power control strategy includes a dynamic power adjustment range of -10-20dBm and a power control threshold accuracy of ±0.1dBm.

6. The federated transfer learning-driven multi-scenario communication parameter optimization method according to claim 4, characterized in that, In step S5, Gaussian random noise is added to the gradient using differential privacy technology. The noise parameters satisfy the following: mean 0, variance 0.01-0.02, and the original information leakage rate of the gradient data after adding noise is ≤0.1%.

7. The federated transfer learning-driven multi-scenario communication parameter optimization method according to claim 4, characterized in that, When the local device performs parameter fine-tuning, it uses real-time interference samples based on the signal-to-noise ratio and packet loss rate of the current frequency band. The signal-to-noise ratio includes a monitoring accuracy of ±1dB and an adjustment response time of ≤10 seconds. The packet loss rate includes a monitoring accuracy of ±0.1% and an adjustment response time of ≤5 seconds. After parameter fine-tuning, the signal-to-noise ratio is improved by ≥3dB and the packet loss rate is reduced by ≥2.5 percentage points.

8. The federated transfer learning-driven multi-scenario communication parameter optimization method according to claim 4, characterized in that, When the cloud aggregates the optimization results of various scenarios, it forms a closed loop of "scenario data accumulation - model optimization - parameter distribution", with a closed loop cycle of ≤30 minutes: the scenario data accumulation frequency is ≥1 time / 5 seconds, the model iteration update time is ≤15 minutes, the parameter distribution delay is ≤10 seconds, and after 6 months of closed loop operation, the accuracy of identifying new interference is improved by ≥32 percentage points.

9. The federated transfer learning-driven multi-scenario communication parameter optimization method according to claim 4, characterized in that, The system provides scene feature similarity heatmaps and parameter optimization suggestions. The heatmap similarity display accuracy is ±0.

01. The heatmap can display the similarity between the current environment and historical scenes in real time, with a similarity range of 0-1. After the parameter set recommended based on similarity is reused, the on-site deployment time is ≤10 minutes.

10. The federated transfer learning-driven multi-scenario communication parameter optimization method according to claim 4, characterized in that, Local processing of real-time interference response requires a processing latency of ≤10 seconds and hardware computing power requirements of GPU utilization <40% or CPU utilization <50%. Cloud processing of large-scale model training requires a single training time of ≤15 minutes and hardware configuration of ≥8 cores and 32G servers, reducing the computing power requirements of on-site equipment by ≥50%.

Images