A Wi-Fi module parameter adjustment method and system

Abstract

The present invention discloses a parameter adjustment method and system for a Wi-Fi module. The method comprises: collecting and obtaining multimodal interference samples; constructing a spatiotemporal annotated dataset; constructing an LSTM-attention model to predict interference and output prediction results; constructing a dynamic interference tensor based on real-time IQ sampling data, prediction results, and a historical interference database, and outputting enhanced features after fusion through confidence-weighted CP decomposition and gated feature fusion; dynamically generating a QoS vector based on device type and service requirements, evaluating service quality based on the final quality value, and dynamically optimizing service quality; constructing and updating a virtual queue based on the final quality value, the prediction results, and the current radio frequency parameters, monitoring quality deviations, and dynamically adjusting radio frequency parameters. Through the closed-loop coordination of interference modeling, service perception, and an optimization framework, the present invention solves the problem of Wi-Fi module self-adaptation in highly dynamic interference environments, making it more suitable for complex scenarios such as industrial wireless control systems.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, relates to a Wi-Fi module optimization technology, and specifically relates to a Wi-Fi module parameter adjustment method and system. Background Art

[0002] A Wi-Fi module is a hardware module that integrates wireless communication capabilities. It typically includes components such as a radio frequency chip, baseband chip, radio frequency front-end, and antenna. The Wi-Fi module receives digital baseband signals from the host device, modulates them, and transmits them through the antenna, enabling the device to send and receive wireless signals. In scenarios such as the Industrial Internet of Things, smart homes, and medical devices, Wi-Fi modules play a critical role in data transmission. Wi-Fi modules are subject to various interference sources during operation. For example, in a factory environment, Wi-Fi modules may be affected by multiple interference sources simultaneously, such as Bluetooth devices, motors, and metal obstacles. This external interference can negatively impact signal reception and data transmission. To ensure communication quality, the Wi-Fi module requires adjustments to its radio frequency parameters.

[0003] Patent publication number CN118250723B discloses a Wi-Fi module radio frequency adjustment method, apparatus, device, and storage medium. The method includes: controlling a sample Wi-Fi module to use various radio frequency parameters to receive sample signals under interference from various service objects; selecting two sample signals to form a signal pair based on the service object; labeling high-quality sample signals in the signal pair based on the service object; training a signal detection network based on the labeled signal pairs and radio frequency parameters; using the trained network to perform quality assessment on signals received by the target Wi-Fi module; and dynamically adjusting the radio frequency parameters based on the quality assessment results to optimize the Wi-Fi module's anti-interference capability and improve communication efficiency.

[0004] However, the above technical solution has some problems: Wi-Fi modules in environments such as factories are simultaneously affected by multiple types of interference sources such as Bluetooth, microwaves, and motors. The above solution only uses single-business object interference sample training and cannot quantify the coupling effect; there are differences in actual business needs. For example, video surveillance requires low latency, while sensor acquisition can tolerate higher latency. The quality assessment of the above technical solution does not distinguish between business priorities and cannot adapt to multi-business terminals. Therefore, it has limited use in high-interference scenarios such as industrial Internet of Things and smart medical care. Summary of the Invention

[0005] Purpose of the invention: To overcome the deficiencies in the prior art, a method and system for adjusting parameters of a Wi-Fi module are provided.

[0006] Technical Solution: To achieve the above objectives, the present invention provides a method for adjusting parameters of a Wi-Fi module, comprising the following steps:

[0007] S1: Collect and obtain multimodal interference samples;

[0008] S2: Construct a spatiotemporal annotation dataset based on multimodal interference samples;

[0009] S3: Extract historical interference sequences and environmental time series data from the spatiotemporal annotation dataset, build an LSTM-attention model for interference prediction, and output the prediction results;

[0010] S4: Based on the real-time IQ sampling data, prediction results and historical interference database, a dynamic interference tensor is constructed. Through confidence-weighted CP decomposition and gated feature fusion, the fused enhanced features are output as anti-interference features.

[0011] S5: Dynamically generates QoS vectors based on device type and service requirements, evaluates service quality using the final quality value, and dynamically optimizes service quality based on interference prediction results.

[0012] S6: Based on the final quality value, prediction results, and current RF parameters, a virtual queue is constructed and updated to monitor quality deviations and dynamically adjust RF parameters.

[0013] Furthermore, the process of obtaining the multimodal interference sample in step S1 includes:

[0014] A1: Initial Data Collection

[0015] Environmental parameters are collected through the environmental sensor array, RF parameters are collected through the sample Wi-Fi module register, and the original signal is collected through the Wi-Fi module RF front-end to obtain the original data packet. The original data packet includes a timestamp, IQ sampling data, environmental vector, and RF parameters.

[0016] A2: Data Preprocessing

[0017] After normalizing and filtering the original data packet, a pre-processed data packet is obtained. The pre-processed data packet includes a timestamp, IQ pre-processed data, an environment vector, and radio frequency parameters.

[0018] A3: Detect interference intensity and identify interference types;

[0019] A4: Perform data annotation on the pre-processed data packets, mark the interference intensity and interference type, and generate multimodal interference samples.

[0020] Furthermore, the construction of the spatiotemporal annotation dataset in step S2 includes:

[0021] B1: Time series alignment, arrange all interference samples in ascending order by timestamp, and obtain a time-aligned sample sequence through equal-interval resampling and intra-slot aggregation;

[0022] B2: Spatial position mapping: Perform spatial position mapping on the location information in the device metadata, add three-dimensional coordinates to each interference sample, and obtain a spatiotemporal correlation sequence;

[0023] B3: Perform data dimensionality reduction and compression on the high-dimensional IQ sampling data to obtain compressed signal features;

[0024] B4: Perform integrity check, physical range verification, type-strength consistency and signal-to-noise ratio verification on the interference samples, and output a qualified sample set;

[0025] B5: Standardize and encode qualified samples and output standardized samples;

[0026] B6: After encapsulating the dataset of all standardized samples, output the spatiotemporal annotation dataset.

[0027] Furthermore, the LSTM-attention model in step S3 includes an LSTM layer, an attention layer, and a fully connected layer. The LSTM layer extracts temporal features, the attention layer focuses on key interference features through a multi-head attention mechanism, and the fully connected layer outputs prediction results, which include predicted interference intensity, predicted interference type probability distribution, and confidence level.

[0028] The LSTM layer is h t , c t =LSTM(x t ,h t-1 ,c t-1 ), where h t is the hidden state, c t is the cell state, x t is the input feature at time t;

[0029] The attention layer is:

[0030]

[0031] s=∑α t h t

[0032] Among them, α t is the attention weight at time t, W a is the attention weight matrix, u T is the attention vector, s is the weighted feature aggregation;

[0033] The output of the fully connected layer is:

[0034]

[0035] P type =softmax(W o s+b o )

[0036] Among them, W o is the output weight matrix, b o is the output bias; the output dimension includes the predicted interference strength and the predicted interference type probability distribution P type .

[0037] Furthermore, the attention layer of the LSTM-attention model in step S3 focuses on key time points through a multi-head attention mechanism to calculate confidence, and the calculation formula is as follows:

[0038]

[0039] Among them, conf is the confidence level; P type,i is the current prediction type probability, is the historical average type probability;

[0040] The calculation formula is:

[0041]

[0042] Among them, t is the current time point, k is the historical time index, K is the sliding window size, i is the category label, δ is the indicator function, and y type Interference type.

[0043] Furthermore, the step S4 specifically includes:

[0044] C1: Dynamic tensor construction:

[0045] Perform short-time Fourier transform on the IQ sample data S:

[0046] F=STFT(S)

[0047] Convert the time domain signal to the time-frequency domain to obtain the frequency components of the signal at different time points;

[0048] By using the short-time Fourier transform signal F, channel state information C, and prediction result P red and historical interference database D hist Perform splicing construction, namely:

[0049] I = concat(F, C, P red , D hist )

[0050] The interference tensor I is a mathematical structure used to describe the characteristics of multi-source interference;

[0051] C2: Confidence-weighted CP decomposition:

[0052] Extract coupling features through CP decomposition:

[0053]

[0054] CP decomposition output coupling eigenvector v int =[a r ; b r ;c r ], where: a r is the interference source feature vector; b r is the frequency band attenuation vector; c r is the time dynamic vector; A vector for historical storage; is the updated vector; R is the rank factor, that is, the number of eigenvector combinations after decomposition;

[0055] C3: Three-branch heterogeneous feature extraction:

[0056] Time domain branch: Use LSTM+cavity convolution structure to extract time domain features f t ;

[0057] Frequency domain branch: Use wavelet packet transform + spectral attention mechanism to extract frequency domain features f f ;

[0058] Interference branch: Process v through the fully connected layer int , extract interference features f p ;

[0059] C4: Gated Fusion:

[0060] For the time domain feature f t , frequency domain characteristics f f and interference characteristics f p Fusion is performed to obtain enhanced features f fusion , expressed as follows:

[0061] f fusion =g☉f t +(1-g)☉f f +f p

[0062] g=σ(W g |f t ;f f ;f p )

[0063] Among them, g is the gate weight; W gIt is a learnable gating weight matrix used to adjust the importance of different branch features; σ is the Sigmoid function, which normalizes the weight value to between 0 and 1.

[0064] Furthermore, the process of dynamically generating the QoS vector in step S5 includes:

[0065] D1: Define the service requirement QoS vector q = [q lat ,q bw ,q rel ] T , used to describe the communication quality requirements of different services, where q lat ,q bw ,q rel Represent the weights of delay, bandwidth, and reliability respectively;

[0066] D2: Calculate basic quality: Basic quality is the basis of QoS vector q base , obtained by calculating the basic parameters of the signal:

[0067] q base =FC(f fusion )

[0068] Among them, FC is the mapping function of the fully connected layer;

[0069] D3: Dynamic QoS vector adjustment. The adjustment formula is:

[0070]

[0071] Get the adjusted business weight vector q adj , where SI is the interference severity index.

[0072] Furthermore, the calculation formula of the final quality value in step S5 is:

[0073] Q final =σ(W T q adj ))·q base

[0074] Among them, Q final is the final mass value; W T is the learnable quality weight matrix; σ is the Sigmoid function;

[0075] Dynamically optimize service quality based on interference prediction results, including:

[0076] Dynamically optimize service quality through online update of weight matrix. The online update formula of weight matrix is:

[0077]

[0078] Among them, α is the first learning rate, which controls the step size of weight matrix update; is the gradient of ListMLE sorting loss with respect to the weight matrix W, L rank Sorting loss for ListMLE.

[0079] Step S6 of this embodiment includes:

[0080] E1: Construct a virtual queue Z(t);

[0081] E2: Virtual queue update, the update formula is:

[0082] Z(t+1)=max[Z(t)+β(γQ th -Q final (t)),0]

[0083] Among them, Q th is the target quality threshold; Q final (t) is the final quality value at the current moment; β is the sensitivity coefficient; γ is the prediction compensation factor;

[0084] E3: Parameter adjustment execution.

[0085] Furthermore, the step E3 specifically includes:

[0086] F1: RF parameter adjustment:

[0087] RF is the radio frequency parameter vector, which is expressed by the formula:

[0088]

[0089] Calculate the radio frequency parameter adjustment value ΔRF, where ΔRF is used to adjust the radio frequency parameters of the Wi-Fi module; where η is the second learning rate; Indicates the final quality value Q final The gradient of the RF parameter RF; sat is the saturation function; conf is the confidence level; Z max is the maximum capacity threshold of the virtual queue;

[0090] F2: Power adjustment:

[0091] By formula:

[0092] P new =P old +ΔP×γ PA

[0093] Calculate the adjusted Wi-Fi module power P new , P old is the current transmit power of the Wi-Fi module, γ PA is the power efficiency factor;

[0094] ΔP is the power adjustment:

[0095]

[0096] in, is the mass gradient term, is the queue factor, conf is the confidence level, ΔP pre is the predicted compensation item;

[0097] F3: Channel switching:

[0098] The channel switching decision depends on the interference type:

[0099]

[0100] According to the channel switching probability formula:

[0101]

[0102] Calculate the channel switching probability p ch ; Among them, k2 is the sensitivity coefficient, e is the natural constant, C type is the channel type coefficient.

[0103] The present invention also provides a parameter adjustment system for a Wi-Fi module, comprising:

[0104] A data acquisition unit, used for acquiring original data packets;

[0105] The preprocessing unit is used to preprocess the original data packet, identify the interference intensity and interference type, perform data annotation, form interference samples, integrate the interference samples, and output a spatiotemporal annotation data set;

[0106] The interference prediction unit is used to predict the type and intensity of interference in advance based on the spatiotemporal annotated data set, quantify the physical characteristics and probability of future interference, and output the interference prediction results;

[0107] Interference tensor decomposition and feature fusion unit, used to integrate real-time signals and prediction results to generate anti-interference features;

[0108] The service perception quality assessment unit is used to dynamically generate QoS vectors based on device type and service requirements, dynamically adjust service weights based on prediction results, integrate anti-interference characteristics with the adjusted QoS vectors, and output the final quality value to evaluate the service quality and dynamically optimize service quality;

[0109] The optimization and adjustment unit is used to build and update a virtual queue based on the final quality value, prediction results, and current RF parameters, monitor quality deviations, and dynamically adjust RF parameters.

[0110] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0111] 1. The present invention solves the problem of Wi-Fi module adaptation in highly dynamic interference environments through the closed-loop coordination of interference modeling, service perception, and optimization framework. It is more suitable for complex scenarios such as industrial wireless control systems. Interference prediction (intensity / type / confidence) is used as control input to replace traditional passive detection, thereby improving response speed. The LSTM-attention model that integrates spatiotemporal features solves the temporal correlation problem of sudden interference. A four-dimensional tensor (real-time signal + prediction result + historical data) is constructed, and feature enhancement is achieved through confidence-weighted CP decomposition. A gating mechanism is introduced to fuse traditional features with prediction branches, thereby improving the discriminability of anti-interference features.

[0112] 2. Based on the prediction results, the present invention dynamically adjusts the QoS vector and updates the virtual queue, and adjusts the strategy and parameters in a targeted manner, which can respond to interference more quickly and accurately. In strong interference scenarios, the bit error rate is significantly reduced, the communication delay is more stable, and the energy efficiency of the equipment is improved. In different business scenarios, it can be dynamically adjusted according to business needs to ensure the communication quality of various services, such as the low latency requirements of medical equipment and the high reliability requirements of industrial sensors. The significant improvement in performance and business adaptability can effectively solve the communication problems in high-interference scenarios such as industrial Internet of Things and smart medical care. BRIEF DESCRIPTION OF THE DRAWINGS

[0113] Figure 1 is a flow chart of the method of the present invention;

[0114] Figure 2 is a flow chart of generating interference samples in the present invention;

[0115] Figure 3 is a flow chart of parameter and radio frequency adjustment in the present invention;

[0116] Figure 4 It is a structural diagram of the Wi-Fi module parameter adjustment system in the present invention. DETAILED DESCRIPTION

[0117] The present invention is further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0118] Example 1:

[0119] like Figure 1 As shown, this embodiment provides a method for adjusting parameters of a Wi-Fi module, including the following steps:

[0120] S1: Collect and obtain multimodal interference samples;

[0121] S2: Construct a spatiotemporal annotation dataset based on multimodal interference samples;

[0122] S3: Extract historical interference sequences and environmental time series data from the spatiotemporal annotation dataset, build an LSTM-attention model for interference prediction, and output the prediction results;

[0123] S4: Based on the real-time IQ sampling data, prediction results and historical interference database, a dynamic interference tensor is constructed. Through confidence-weighted CP decomposition and gated feature fusion, the fused enhanced features are output as anti-interference features.

[0124] S5: Dynamically generates QoS vectors based on device type and service requirements, evaluates service quality using the final quality value, and dynamically optimizes service quality based on interference prediction results.

[0125] S6: Based on the final quality value, prediction results, and current RF parameters, a virtual queue is constructed and updated to monitor quality deviations and dynamically adjust RF parameters to achieve adaptive parameter control.

[0126] In step S1 of this embodiment, multimodal interference sample collection is as follows: in an electromagnetic shielding environment, a programmable interference source generates an interference signal of a specified type and strength, and at the same time, the sample Wi-Fi module receives an OFDM test signal from a standard tester and synchronously collects environmental sensor data to obtain interference samples. The interference samples include: timestamp t, IQ sampling data S, environmental vector E (temperature, humidity, electromagnetic noise), interference intensity y strength , interference type y type (0: none, 1: Bluetooth, 2: microwave, 3: motor), RF parameters (power, channel, bandwidth);

[0127] In this embodiment, the electromagnetic shielding environment uses an anechoic laboratory, which effectively reduces external interference and ensures the accuracy of sample collection. A programmable interference array is deployed in the anechoic chamber to simultaneously generate three types of interference signals: Bluetooth frequency hopping (2.4GHz), microwave pulses (5.8GHz), and motor noise (0-1kHz harmonics). The programmable interference array can flexibly adjust interference parameters such as intensity and frequency to simulate realistic and complex interference scenarios.

[0128] The standard tester uses an 802.11ax signal generator. The OFDM (Orthogonal Frequency Division Multiplexing) signal is generated by the signal generator to simulate the wireless communication signal in actual application scenarios and is sent to the sample Wi-Fi module together with the interference signal.

[0129] In this embodiment, the sample Wi-Fi module is equipped with two antennas, each with 256 sampling points. The collected IQ sample data S represents the real-time signal. The signal reception stage increases the sampling rate of the OFDM signal received by the sample Wi-Fi module to 160MHz. This high sampling rate helps capture transient interference—short-lived, rapidly changing interference signals—providing richer data for subsequent interference analysis.

[0130] The data collected by environmental sensors include environmental parameters (temperature, humidity, electromagnetic background noise) and radio frequency parameters (transmit power, channel, bandwidth).

[0131] The purpose of multimodal interference sample collection is to obtain signal data containing multiple interference characteristics to provide a basis for interference modeling and analysis. After collection, signal data containing interference information is obtained.

[0132] Reference Figure 2 ,The process of obtaining multimodal interference samples includes:

[0133] A1: Initial Data Collection

[0134] Environmental parameters are collected through the environmental sensor array, RF parameters are collected through the sample Wi-Fi module register, and the original signal is collected through the Wi-Fi module RF front-end to obtain the original data packet. The original data packet includes timestamp, IQ sampling data, environmental vector (temperature, humidity, electromagnetic noise), and RF parameters (power, channel, bandwidth).

[0135] A2: Data Preprocessing

[0136] After normalizing and filtering the original data packet, a pre-processed data packet is obtained. The pre-processed data packet includes a timestamp, IQ pre-processed data, an environment vector, and radio frequency parameters.

[0137] The IQ sampling data is calibrated to eliminate hardware deviations, and then amplitude normalization and noise reduction filtering are performed. The amplitude and phase calibration is performed as follows:

[0138]

[0139] The normalized signal I' / Q' is obtained, with a mean of 0 and a standard deviation of 1. Here, I / Q are the original in-phase and quadrature signals, μI / μQ is the statistical mean of the original I / Q signal (reflecting the hardware offset), and σI / σQ is the standard deviation of the original I / Q signal (reflecting the signal amplitude fluctuation).

[0140] Through Butterworth filtering:

[0141]

[0142] The complex frequency domain transfer function of the filter H(s) is obtained, where ω c is the cutoff angular frequency; n is the filter order, which controls the steepness of the transition band; s is the complex frequency variable;

[0143] A3: Detect interference intensity and identify interference types;

[0144] The Wi-Fi module's built-in FFT (Fast Fourier Transform) unit scans full-band signals to detect interference.

[0145] The interference intensity detection method is as follows:

[0146] 1) FFT spectrum analysis: perform fast Fourier transform on the IQ preprocessed data in the preprocessed data packet and calculate the frequency domain amplitude:

[0147]

[0148] Among them, signal_window is the time domain signal sampled by ADC; FFT() is fast Fourier transform; e -j2πkn / N is the complex exponential rotation factor, k is the index of the frequency domain component, n is the index of the time domain signal, and N is the length of the signal window;

[0149] 2) Calculate the full-band interference intensity I total :

[0150]

[0151] Where N1 is the length of the time domain signal; FFT(signal_window)[k] is the amplitude of the kth frequency point after fast Fourier transform;

[0152] By calculating the mean of the square of the signal frequency domain amplitude and performing a logarithmic transformation, the signal energy is converted into an interference intensity index that is easy to measure and compare, reflecting the overall interference level;

[0153] Based on the full-band interference intensity I total Classify the interference intensity level:

[0154] I total <-70, weak interference, can be ignored, severity level = 0;

[0155] -70≤I total ≤-60, moderate interference, severity level = 0.5;

[0156] I total >-60, strong interference, severity level = 1.

[0157] 3) Subcarrier interference positioning

[0158] For each subcarrier, the current subcarrier frequency domain amplitude FFT[k] is compared with the reference energy E[X clean [k] 2 ](factory calibration value), calculate the subcarrier interference intensity I sub , the formula is:

[0159]

[0160] Output the interference intensity distribution map of each subcarrier, which can intuitively reflect the interference situation on different subcarriers and facilitate the precise positioning of the interference frequency band;

[0161] 4) Interference type identification:

[0162] From the subcarrier interference intensity I sub Extract key features from the distribution, including dynamic range, mutation location, and energy entropy;

[0163] The dynamic range reflects the fluctuation amplitude of the subcarrier interference intensity. A large dynamic range indicates the presence of complex and variable interference sources, such as sudden electromagnetic pulse interference.

[0164] The mutation location is used to locate subcarriers where interference intensity changes significantly. This is done by calculating the difference in interference intensity between adjacent subcarriers. When the difference exceeds a set threshold of 6dB, the corresponding subcarrier position is identified as the mutation point. Detecting the mutation location helps quickly identify the specific frequency band where interference is occurring.

[0165] Energy entropy is used to measure the disorder of the subcarrier interference intensity distribution.

[0166] The above key features are comprehensively analyzed by a lightweight classifier to output the interference type;

[0167] interference type=0: no interference;

[0168] interference type=1: Bluetooth interference (frequency band interference);

[0169] Interference type = 2: Microwave interference (broadband noise);

[0170] Interference type = 3: Motor interference (harmonic interference).

[0171] The module's built-in FFT unit completes a scan every 2ms, and the output data structure includes the total interference intensity, subcarrier interference distribution, and interference type identification.

[0172] A4: Perform data annotation on the pre-processed data packets, mark the interference intensity and interference type, and generate multimodal interference samples.

[0173] The construction of the spatiotemporal annotation dataset in step S2 of this embodiment includes:

[0174] B1: Time series alignment, arrange all interference samples in ascending order by timestamp, and obtain a time-aligned sample sequence through equal-interval resampling and intra-slot aggregation;

[0175] B2: Spatial position mapping: Perform spatial position mapping on the location information in the device metadata, add three-dimensional coordinates to each interference sample, and obtain a spatiotemporal correlation sequence;

[0176] B3: Perform data dimensionality reduction and compression on the high-dimensional IQ sampling data to obtain compressed signal features;

[0177] B4: Perform integrity check, physical range verification, type-strength consistency and signal-to-noise ratio verification on the interference samples, and output a qualified sample set;

[0178] B5: Standardize and encode qualified samples and output standardized samples;

[0179] B6: After encapsulating all standardized samples into datasets, output the spatiotemporal annotation dataset D = {t, S, E, y type ,y strength}, where: t is the timestamp, S is the IQ sampling data, E is the environment vector, including temperature, humidity, electromagnetic background noise, y type Interference type, including Bluetooth, microwave, and motor. strength is the interference intensity, the dBm value corresponding to the interference source transmission power (set to -80 / -70 / -60 / -50dBm during collection).

[0180] Reference Figure 3 , Steps S3 to S6 are described in detail below:

[0181] In step S3 of this embodiment, the LSTM-attention model includes an LSTM layer, an attention layer, and a fully connected layer. The LSTM layer extracts temporal features, the attention layer focuses on key interference features through a multi-head attention mechanism, and the fully connected layer outputs prediction results, which include predicted interference intensity, predicted interference type probability distribution, and confidence level.

[0182] The interference prediction method in step S3 of this embodiment is as follows:

[0183] 1) Extract the interference intensity of the last five time points from the spatiotemporal annotation dataset D in reverse chronological order to form a historical interference sequence I hist =[I total(t-5),…,I total (t)]; extract the environmental vectors of the last five time points from the spatiotemporal annotation dataset D in reverse chronological order to form the environmental time series data E hist =[E(t-5),…,E(t)].

[0184] 2) Constructing LSTM-attention model and integrating historical interference sequence I hist and environmental time series data E hist , the LSTM layer extracts time series features, the multi-head attention mechanism captures the dependencies between sequences, and the fully connected layer outputs the prediction results;

[0185] The input layer of the model architecture is X = [I hist ;E hist ];

[0186] The LSTM layer is h t , c t =LSTM(x t ,h t-1 ,c t-1 ), where h t is the hidden state (temporal memory), c t is the cell state (long-term memory), x t is the input feature at time t;

[0187] The attention layer is:

[0188]

[0189] s=∑α t h t

[0190] Among them, α t is the attention weight at time t, W a is the attention weight matrix, u T is the attention vector, s is the weighted feature aggregation;

[0191] The output of the fully connected layer is:

[0192]

[0193] P type =softmax(W o s+b o )

[0194] Among them, W o is the output weight matrix, b o is the output bias; the output dimension includes the predicted interference strength and the predicted interference type probability distribution P type ;

[0195] Interference intensity prediction: Covers the common interference strength range of WiFi signals.

[0196] 3) The attention layer uses a multi-head attention mechanism to focus on key time points for confidence calculation. The calculation formula is as follows:

[0197]

[0198] Among them, conf is the confidence level; P type,i is the current prediction type probability, is the historical average type probability;

[0199] The calculation formula is:

[0200]

[0201] Among them, t is the current time point, k1 is the historical time index, K is the sliding window size, i is the category label, δ is the indicator function, and y type is the interference type;

[0202] The confidence levels are as follows:

[0203] conf=[0.9,1.0], the confidence level is extremely high, and the decision recommendation is: fully trust the prediction results;

[0204] conf=[0.7,0.9], the confidence level is high confidence, and the decision recommendation is: trust the prediction result;

[0205] conf=[0.5,0.7], the confidence level is medium confidence, and the decision recommendation is: partially adopt the prediction results;

[0206] conf=[0.3,0.5], the confidence level is low, and the decision recommendation is: adopt the prediction results with caution;

[0207] conf=[0,0.3], the confidence level is very low, and the decision recommendation is: ignore the prediction result.

[0208] 4) Output prediction result P red , quantify the physical characteristics and probability of occurrence of future interference;

[0209]

[0210] Among them, p b is the Bluetooth interference probability, p m is the probability of microwave interference, p c is the motor interference probability, ∑p=p b +pm +p c =1.

[0211] According to the interference prediction results, the decision and parameters are modified as follows:

[0212] Active defense: Predict interference events in advance and initiate protective measures before interference occurs.

[0213] Decision optimization: Provides forward-looking guidance for parameter adjustments to avoid performance fluctuations caused by passive responses.

[0214] Resource allocation: Emergency resources are activated based on the forecast intensity, and the deep protection mechanism is enabled when the forecast is high-confidence.

[0215] In step S4 of this embodiment, the interference tensor combines the real-time signal STFT time-frequency spectrum, the predicted interference intensity, the predicted interference type probability distribution and the historical interference data. The interference tensor decomposition and feature fusion method is as follows:

[0216] C1: Dynamic tensor construction:

[0217] Perform short-time Fourier transform on the IQ sample data S:

[0218] F=STFT(S)

[0219] Convert the time domain signal to the time-frequency domain to obtain the frequency components of the signal at different time points, which facilitates the analysis of the frequency characteristics of the interference;

[0220] By using the short-time Fourier transform signal F, channel state information C, and prediction result P red and historical interference database D hist Perform splicing construction, namely:

[0221] I = concat(F, C, P red , D hist )

[0222] The interference tensor I is a mathematical structure used to describe the characteristics of multi-source interference; the channel state information C describes the transmission characteristics of the signal from the transmitter to the receiver, such as attenuation and delay, and is obtained through the channel estimation technology of the Wi-Fi module; the historical interference database D hist The interference tensor is constructed by gradually accumulating data such as interference type, intensity, and frequency through long-term monitoring and recording of interference information in different scenarios. Combined with the time-frequency characteristics of the current signal, a more comprehensive interference tensor can be constructed.

[0223] C2: Confidence-weighted CP decomposition:

[0224] Extract coupling features through CP decomposition:

[0225]

[0226] CP decomposition output coupling eigenvector v int =[a r ; b r ;c r ], where: a r is the interference source feature vector, which comes from the spectrum sensing device and is used to characterize the type, intensity and other information of the interference source; b r is the frequency band attenuation vector, which is obtained according to the channel state information CSI and reflects the attenuation of different frequency bands during the transmission process; c r It is a time dynamic vector, provided by the historical interference database, reflecting the change pattern of interference over time; A vector for historical storage; is the updated vector; R is the rank factor, that is, the number of eigenvector combinations after decomposition. R is adaptively selected through AIC (Akaike Information Criterion). AIC is a statistic used for model selection that balances the goodness of fit and complexity of the model. The coupling features are extracted through CP decomposition.

[0227] C3: Three-branch heterogeneous feature extraction:

[0228] Time domain branch: Use LSTM+cavity convolution structure to extract time domain features f t LSTM can process long sequence data and effectively capture the long-term dependencies of signals in the time dimension, making it suitable for analyzing the changing trends of signals over time. Dilated convolution expands the receptive field of the convolution kernel without increasing the number of parameters, allowing for the acquisition of a wider range of time domain information.

[0229] Frequency domain branch: Use wavelet packet transform + spectral attention mechanism to extract frequency domain features f f Wavelet packet transform can perform more precise frequency domain decomposition of the signal, decomposing the signal into multiple frequency bands and obtaining the characteristics of different frequency bands; the spectral attention mechanism can automatically learn the importance of different frequency bands and highlight the frequency band information that is valuable for signal analysis;

[0230] Interference branch: Process v through the fully connected layer int , extract interference features f p ; The fully connected layer can further extract and integrate the coupling characteristics of interference;

[0231] C4: Gated Fusion:

[0232] For the time domain feature f t , frequency domain characteristics f f and interference characteristics f p Fusion is performed to obtain enhanced features f fusion , expressed as follows:

[0233] f fusion =g☉f t +(1-g)☉f f +f p

[0234] g=σ(W g |f t ;f f ;f p )

[0235] Among them, g is the gate weight; W g It is a learnable gating weight matrix used to adjust the importance of different branch features; σ is the Sigmoid function, which normalizes the weight value to between 0 and 1.

[0236] Gated fusion integrates feature information of different dimensions to more comprehensively describe interference and signal characteristics, thereby improving the accuracy and effectiveness of the anti-interference algorithm. The enhanced feature f obtained after fusion is fusion Subsequently, it will participate in operations such as business perception quality assessment to provide richer information for quality assessment and parameter adjustment.

[0237] In step S5 of this embodiment, the process of dynamically generating a QoS vector includes:

[0238] D1: Define the service requirement QoS vector q = [q lat ,q bw ,q rel ] T , used to describe the communication quality requirements of different services, where q lat ,q bw ,q rel Represent the weights of delay, bandwidth, and reliability respectively;

[0239] Business type refers to the business category carried by the device, such as video surveillance, sensor data acquisition, medical data transmission, etc.; q is dynamically generated according to the device type and business requirements, q lat ,q bw ,q rel The value is determined according to specific business needs.

[0240] For example, medical equipment has high latency requirements, lat The value is larger, q=[0.9,0.2,0.7] T , where the first element 0.9 represents the latency weight, reflecting the urgent need for low latency in medical equipment;

[0241] Industrial sensors have high requirements for reliability. rel The value is larger, q=[0.3,0.6,0.8] T ;

[0242] D2: Calculate basic quality: Basic quality is the basis of QoS vector q base , obtained by calculating the basic parameters of the signal (such as signal-to-noise ratio, bit error rate, etc.):

[0243] q base =FC(f fusion )

[0244] Among them, FC is the mapping function of the fully connected layer;

[0245] D3: Dynamic QoS vector adjustment. The adjustment formula is:

[0246]

[0247] Get the adjusted business weight vector q adj , where SI is the interference severity index; in this embodiment When, SI = 0.05; When, SI = 0.62; When, SI = 0.95;

[0248] The final quality value is calculated as:

[0249] Q final =σ(W T q adj )·q base

[0250] Among them, Q final is the final mass value; W T is a learnable quality weight matrix (the initial value is obtained by looking up the service type table), which is used to evaluate whether the current communication quality meets the service requirements; σ is the Sigmoid function;

[0251] Q final When the result is [0.95, 1.0], the quality level is excellent and the business requirements are fully met. The subsequent processing strategy is: maintain the current parameters and start the energy-saving mode: P new =P old -2dB, increase the modulation order;

[0252] Q final When the result is [0.85, 0.95], the quality level is good and the service is basically satisfied. The subsequent processing strategy is: fine-tuning power: ΔP = 0.5 × (0.95-Q final )×10, channel optimization;

[0253] Q finalWhen the result is [0, 0.85], the quality level is poor and the service quality does not meet the standard. The subsequent processing strategy is: emergency response, and the power is greatly increased: ΔP=min(10,(0.85-Q final ))×20, channel emergency switching;

[0254] Q final Subsequent participation in Lyapunov optimization adjustment operations serves as the basis for parameter adjustment.

[0255] Dynamically optimize service quality based on interference prediction results, including:

[0256] Dynamically optimize service quality through online update of weight matrix. The online update formula of weight matrix is:

[0257]

[0258] Among them, α is the first learning rate, which controls the step size of weight matrix update; is the gradient of ListMLE sorting loss with respect to the weight matrix W, which is obtained by calculating the partial derivative of the loss function with respect to the weight matrix and is used to indicate the direction of weight matrix update; L rank After introducing the service priority constraint into the ListMLE sorting loss, the weight matrix can be adjusted according to the priority requirements of different services, thereby optimizing the quality assessment results.

[0259] Weight matrix updates adjust the elements in the matrix to accommodate different service requirements. Updates are typically triggered by detecting a change in service type or persistently substandard communication quality. Updates terminate when the weight matrix converges, meaning that after multiple updates, the quality assessment results no longer change significantly.

[0260] In step S6 of this embodiment, the parameter adjustment process includes:

[0261] E1: Construct a virtual queue Z(t);

[0262] The virtual queue Z(t) is used to characterize historical quality deviations, recording the deviation between the quality at previous moments and the target quality. The purpose of building a virtual queue is to monitor quality deviations and provide a basis for adjusting RF parameters, enabling the system to dynamically adapt to interference environments. Because target quality thresholds may vary for different devices, it is necessary to build a corresponding virtual queue based on the device type.

[0263] E2: Virtual queue update, the update formula of Lyapunov optimization is:

[0264] Z(t+1)=max[Z(t)+β(γQ th -Q final (t)),0]

[0265] Among them, Q th Q is the target quality threshold, which is set by the device type. Different devices set different target quality thresholds according to their own business needs. final (t) is the final quality value at the current moment; β is the sensitivity coefficient, which defaults to 0.05 and increases adaptively with the interference intensity. final <0.8Q th When , β is multiplied to accelerate the response, so that the system can adjust to the quality degradation more quickly; γ is the prediction compensation factor, γ = 1 + 0.3 (1-SI), SI is the interference severity index;

[0266] This update formula is a specific form of the Lyapunov optimization adjustment operation, which is used to achieve dynamic optimization of parameters by maintaining queue stability. The core of Lyapunov optimization is to minimize long-term costs while ensuring system stability by constructing a Lyapunov function, represented here by a virtual queue Z(t).

[0267] Since the interference intensity will affect β, the update of β is as follows:

[0268] β new =β×(1+0.5×severity level)

[0269] Effect: Accelerates response in the case of strong interference. The greater the interference intensity, the larger the β value, and the more sensitive the virtual queue Z(t) is to quality deviations.

[0270] E3: Parameter adjustment execution, including the following steps:

[0271] F1: RF parameter adjustment:

[0272] RF is a radio frequency parameter vector, which includes parameters such as transmit power, operating channel, and bandwidth. Its initial value is set according to the default configuration of the device and service requirements.

[0273] By formula:

[0274]

[0275] Calculate the RF parameter adjustment value ΔRF, which is used to adjust the Wi-Fi module's RF parameters, such as power, channel, and bandwidth, to optimize communication quality. η is the second learning rate, which has an initial value of 0.01 and decays with the number of training cycles to avoid over-adjustment. Indicates the final quality value Q final The gradient of the RF parameter RF is obtained by calculating the partial derivative of the quality value with respect to each RF parameter, indicating the direction of RF parameter adjustment; sat is the saturation function; conf is the confidence level; Z maxThe maximum capacity threshold of the virtual queue is used to prevent the queue value from growing infinitely and causing system instability. max The specific value of Z needs to be set in combination with the scene. In the industrial strong interference scene, Z max The value is 20, in general scenarios Z max The value is 15.

[0276] When Z(t) is small, the adjustment amount is small to ensure system stability; when Z(t) is large, the adjustment force is increased to quickly restore communication quality. The trigger condition for RF parameter adjustment is that the virtual queue Z(t) exceeds a certain threshold or the communication quality continues to be lower than the target quality; the termination condition is that the communication quality reaches the target quality threshold and the virtual queue is stable;

[0277] F2: Power adjustment:

[0278] By formula:

[0279] P new =P old +ΔP×γ PA

[0280] Calculate the adjusted Wi-Fi module power P new ;P old The current transmit power of the Wi-Fi module is obtained in real time by the module's power monitoring module; PA is the power efficiency factor, which is calculated by the PA nonlinear model and takes into account the nonlinear characteristics of the power amplifier, making the power adjustment more in line with the actual situation;

[0281] ΔP is the power adjustment amount, which is calculated based on the difference between the current interference intensity and the historical average interference intensity. The calculation formula is:

[0282]

[0283] in, is the mass gradient term, is the queue factor, conf is the confidence level, ΔP pre is the predicted compensation item;

[0284] Look-ahead compensation based on forecast strength is as follows:

[0285]

[0286] The trigger conditions for power adjustment are the same as those for RF parameter adjustment. The termination condition is that after power adjustment, the communication quality meets the target requirements and the power is stable.

[0287] F3: Channel switching:

[0288] The channel switching decision depends on the interference type:

[0289]

[0290] According to the channel switching probability formula:

[0291]

[0292] Calculate the channel switching probability p ch , used to decide whether to switch channels; where k2 is the sensitivity coefficient, and the channel switching probability is calculated based on the value of the virtual queue Z(t). When Z(t) is larger, the channel switching probability increases, and the system is more inclined to switch to other channels to avoid interference; e is a natural constant with a value of 2.718; C type is the channel type coefficient, which is used to adjust the decision probability according to the inherent type of the channel.

[0293] C type Interference prediction provides the source of historical interference data or predefined rules based on channel types. For example:

[0294] If a certain channel type (such as Wi-Fi channel) has a high frequency of interference in history, the interference prediction model will output a high interference probability, which will lead to the C type is set to a lower value (such as 0.7); on the contrary, for “clean” channel types (such as dedicated LTE bands), the interference prediction shows low risk, C type is set to a high value (such as 1.0 or higher).

[0295] Interference prediction is not only used to calculate Z(t), but also used to train or update C type For example:

[0296] In a machine learning driven model, interference prediction data can be used to classify channel types (e.g., through cluster analysis) and then assign a C type In this way, the decision takes into account both the real-time interference (via Z(t)) and the inherent interference risk of the channel type (via C type ), which improves the robustness of decision making.

[0297] One of the classifications of channel types is as follows:

[0298] High reliability channels (such as licensed frequency bands, dedicated channels): C type If it is greater than 1 or close to 1, it means that the channel has strong anti-interference ability and the probability is amplified;

[0299] Low reliability channels (such as unlicensed bands and shared channels): C type Less than 1 or close to 0, it means that the channel is susceptible to interference and the probability is suppressed;

[0300] Channel class code: C type Sometimes it is a numerical factor based on a map of discrete categories, where different categories correspond to different interference risks.

[0301] The trigger condition for channel switching is that the current channel is severely interfered with and the communication quality continues to decline; the switching content is to select a channel with less interference; the termination condition is that the communication quality is improved and stabilized after switching to the new channel. ch Used to evaluate whether channel switching is required. ch When p is larger, the system is more inclined to switch channels; when p ch When it is smaller, keep the current channel. ch Calculation results, if p ch If the set threshold is exceeded, the channel switching operation is triggered; otherwise, the current channel remains unchanged.

[0302] The decision rule is as follows: ch >0.7, perform channel switching; p ch ≤0.7, maintain the current channel.

[0303] The effects of interference characteristics on adjustments are as follows:

[0304] For interference intensity:

[0305] When there is strong interference, the adjustment strategy is power compensation + emergency channel switching;

[0306] When the interference is moderate, the adjustment strategy is power fine-tuning + service degradation;

[0307] When there is weak interference, the adjustment strategy is power optimization + service enhancement.

[0308] For interference types:

[0309] When Bluetooth interference occurs, the channel strategy is to switch to a 5GHz non-overlapping channel, and the power strategy is to reduce power to avoid conflict.

[0310] In the event of microwave interference, the channel strategy is to switch to the DFS channel, and the power strategy is to increase power to resist noise;

[0311] When there is motor interference, the channel strategy is to maintain the channel, and the power strategy is to significantly increase the power.

[0312] The updated radio parameters are written to registers in the Wi-Fi module. The Wi-Fi module makes adjustments when it detects that communication quality falls below the target quality threshold or when the virtual queue Z(t) exceeds a certain threshold. These adjustments include radio parameter adjustments, with the adjustment amount calculated and applied using an optimization algorithm. After the adjustments, communication quality is reassessed. Adjustments are considered complete when communication quality reaches the target quality threshold and remains stable, while the virtual queue Z(t) remains within a reasonable range.

[0313] Example 2:

[0314] like Figure 4 As shown, based on the method of Example 1, this embodiment provides a parameter adjustment system for a Wi-Fi module, including:

[0315] A data acquisition unit, used for acquiring original data packets;

[0316] In this embodiment, in an electromagnetically shielded environment, a programmable interference source generates an interference signal of a specified type and strength. The tester sends an OFDM test signal. The sample Wi-Fi module simultaneously collects the interference signal, the actual test signal, environmental parameters, and device RF parameters to obtain the original data packet. Environmental parameters include temperature, humidity, and electromagnetic noise, and RF parameters include rate, channel, and bandwidth.

[0317] The preprocessing unit is used to preprocess the original data packet, identify the interference intensity and interference type, perform data annotation, form interference samples, integrate the interference samples, and output a spatiotemporal annotation data set;

[0318] The interference prediction unit is used to predict the type and intensity of interference in advance based on the spatiotemporal annotated data set, quantify the physical characteristics and probability of future interference, and output the interference prediction results;

[0319] Interference tensor decomposition and feature fusion unit, used to integrate real-time signals and prediction results to generate anti-interference features;

[0320] The service perception quality assessment unit is used to dynamically generate QoS vectors based on device type and service requirements, dynamically adjust service weights based on prediction results, integrate anti-interference characteristics with the adjusted QoS vectors, and output the final quality value to evaluate the service quality and dynamically optimize service quality;

[0321] The optimization and adjustment unit is used to build and update a virtual queue based on the final quality value, prediction results, and current RF parameters, monitor quality deviations, and dynamically adjust RF parameters to achieve adaptive parameter control.

[0322] In this embodiment, the interference prediction unit extracts historical interference sequences and environmental time series data from the spatiotemporal annotation dataset, constructs an LSTM-attention model to perform interference prediction, and outputs prediction results including predicted interference intensity, probability distribution of interference type, and prediction confidence.

[0323] In this embodiment, the interference tensor decomposition and feature fusion unit constructs a dynamic interference tensor based on the real-time test signal, prediction results and historical interference database, and outputs the fused enhanced features through weighted CP decomposition and gated feature fusion, which is the anti-interference feature.

Claims

1. A method for adjusting parameters of a Wi-Fi module, characterized in that: The steps include: S1: Collect and obtain multimodal interference samples; S2: Construct a spatiotemporal annotation dataset based on multimodal interference samples; S3: Extract historical interference sequences and environmental time series data from the spatiotemporal annotation dataset, build an LSTM-attention model for interference prediction, and output the prediction results; S4: Based on the real-time IQ sampling data, prediction results and historical interference database, a dynamic interference tensor is constructed. Through confidence-weighted CP decomposition and gated feature fusion, the fused enhanced features are output as anti-interference features. S5: Based on the anti-interference characteristics, QoS vectors are dynamically generated according to device type and service requirements. The service quality is evaluated using the final quality value, and service quality is dynamically optimized based on the interference prediction results. S6: Based on the final quality value, prediction results, and current RF parameters, a virtual queue is constructed and updated to monitor quality deviations and dynamically adjust RF parameters.

2. The parameter adjustment method of a Wi-Fi module according to claim 1, characterized in that: The process of obtaining the multimodal interference sample in step S1 includes: A1: Initial Data Collection Environmental parameters are collected through the environmental sensor array, RF parameters are collected through the sample Wi-Fi module register, and the original signal is collected through the Wi-Fi module RF front-end to obtain the original data packet. The original data packet includes a timestamp, IQ sampling data, environmental vector, and RF parameters. A2: Data Preprocessing After normalizing and filtering the original data packet, a pre-processed data packet is obtained. The pre-processed data packet includes a timestamp, IQ pre-processed data, an environment vector, and radio frequency parameters. A3: Detect interference intensity and identify interference types; A4: Perform data annotation on the pre-processed data packets, mark the interference intensity and interference type, and generate multimodal interference samples.

3. The parameter adjustment method of a Wi-Fi module according to claim 2, characterized in that: The construction of the spatiotemporal annotation dataset in step S2 includes: B1: Time series alignment, arrange all interference samples in ascending order by timestamp, and obtain a time-aligned sample sequence through equal-interval resampling and intra-slot aggregation; B2: Spatial position mapping: Perform spatial position mapping on the location information in the device metadata, add three-dimensional coordinates to each interference sample, and obtain a spatiotemporal correlation sequence; B3: Perform data dimensionality reduction and compression on the high-dimensional IQ sampling data to obtain compressed signal features; B4: Perform integrity check, physical range verification, type-strength consistency and signal-to-noise ratio verification on the interference samples, and output a qualified sample set; B5: Standardize and encode qualified samples and output standardized samples; B6: After encapsulating the dataset of all standardized samples, output the spatiotemporal annotation dataset.

4. The parameter adjustment method of a Wi-Fi module according to claim 1, characterized in that: In step S3, the LSTM-attention model includes an LSTM layer, an attention layer, and a fully connected layer. The LSTM layer extracts temporal features, the attention layer focuses on key interference features through a multi-head attention mechanism, and the fully connected layer outputs prediction results, which include predicted interference intensity, predicted interference type probability distribution, and confidence level. The LSTM layer is h t , c t =LSTM(x t ,h t-1 ,c t-1 ), where h t is the hidden state, c t is the cell state, x t is the input feature at time t; The attention layer is: Among them, α t is the attention weight at time t, W a is the attention weight matrix, u T is the attention vector, s is the weighted feature aggregation; The output of the fully connected layer is: P type =softmax(W o s+b o ) Among them, W o is the output weight matrix, b o is the output bias; the output dimension includes the predicted interference strength and the predicted interference type probability distribution P type ; The attention layer of the LSTM-attention model uses a multi-head attention mechanism to focus on key time points for confidence calculation. The calculation formula is as follows: Among them, conf is the confidence level; P type,i is the current prediction type probability, is the historical average type probability; The calculation formula is: Among them, t is the current time point, k is the historical time index, K is the sliding window size, i is the category label, δ is the indicator function, and y type Interference type.

5. The parameter adjustment method of a Wi-Fi module according to claim 4, characterized in that: The step S4 specifically includes: C1: Dynamic tensor construction: Perform short-time Fourier transform on the IQ sample data S: F=STFT(S) Convert the time domain signal to the time-frequency domain to obtain the frequency components of the signal at different time points; By using the short-time Fourier transform signal F, channel state information C, and prediction result P red and historical interference database D hist Perform splicing construction, namely: I=concat(F,C,P red ,D hist ) The interference tensor I is a mathematical structure used to describe the characteristics of multi-source interference; C2: Confidence-weighted CP decomposition: Extract coupling features through CP decomposition: CP decomposition output coupling eigenvector v int =[a r ; b r ;c r ], where: a r is the interference source feature vector; b r is the frequency band attenuation vector; c r is the time dynamic vector; A vector for historical storage; is the updated vector; R is the rank factor, that is, the number of eigenvector combinations after decomposition; C3: Three-branch heterogeneous feature extraction: Time domain branch: Use LSTM+cavity convolution structure to extract time domain features f t ; Frequency domain branch: Use wavelet packet transform + spectral attention mechanism to extract frequency domain features f f ; Interference branch: Process v through the fully connected layer int , extract interference features f p ; C4: Gated Fusion: For the time domain feature f t , frequency domain characteristics f f and interference characteristics f p Fusion is performed to obtain enhanced features f fusion , expressed as follows: f fusion =g⊙f t +(1-g)⊙f f +f p g=σ(W g |f t ;f f ;f p ) Among them, g is the gate weight; W g It is a learnable gating weight matrix used to adjust the importance of different branch features; σ is the Sigmoid function, which normalizes the weight value to between 0 and 1.

6. The parameter adjustment method of a Wi-Fi module according to claim 5, characterized in that: The process of dynamically generating the QoS vector in step S5 includes: D1: Define the service requirement QoS vector q = [q lat ,q bw ,q rel ] T , used to describe the communication quality requirements of different services, where q lat ,q bw ,q rel Represent the weights of delay, bandwidth, and reliability respectively; D2: Calculate basic quality: Basic quality is the basis of QoS vector q base , obtained by calculating the basic parameters of the signal: q base =FC(f fusion ) Among them, FC is the mapping function of the fully connected layer; D3: Dynamic QoS vector adjustment. The adjustment formula is: Get the adjusted business weight vector q adj , where SI is the interference severity index; The calculation formula of the final quality value in step S5 is: Q final =σ(W T q adj )·q base Among them, Q final is the final mass value; W T is the learnable quality weight matrix; σ is the Sigmoid function; Dynamically optimize service quality based on interference prediction results, including: Dynamically optimize service quality through online update of weight matrix. The online update formula of weight matrix is: Among them, α is the first learning rate, which controls the step size of weight matrix update; is the gradient of ListMLE sorting loss with respect to the weight matrix W, L rank Sorting loss for ListMLE.

7. The parameter adjustment method of a Wi-Fi module according to claim 6, characterized in that: The step S6 comprises: E1: Construct a virtual queue Z(t); E2: Virtual queue update, the update formula is: Z(t+1)=max[Z(t)+β(γQ th -Q final (t)),0] Among them, Q th is the target quality threshold; Q final (t) is the final quality value at the current moment; β is the sensitivity coefficient; γ is the prediction compensation factor; E3: Parameter adjustment execution.

8. The method for adjusting parameters of a Wi-Fi module according to claim 7, wherein: The step E3 specifically includes: F1: RF parameter adjustment: RF is the radio frequency parameter vector, which is expressed by the formula: Calculate the radio frequency parameter adjustment value ΔRF, where ΔRF is used to adjust the radio frequency parameters of the Wi-Fi module; where η is the second learning rate; Indicates the final quality value Q final The gradient of the RF parameter RF; sat is the saturation function; conf is the confidence level; Z max is the maximum capacity threshold of the virtual queue; F2: Power adjustment: By formula: P new =P old +ΔP×γ PA Calculate the adjusted Wi-Fi module power P new , P old is the current transmit power of the Wi-Fi module, γ PA is the power efficiency factor; ΔP is the power adjustment: in, is the mass gradient term, is the queue factor, conf is the confidence level, ΔP pre is the predicted compensation item; F3: Channel switching: The channel switching decision depends on the interference type: According to the channel switching probability formula: Calculate the channel switching probability p ch ; Among them, k2 is the sensitivity coefficient, e is the natural constant, C type is the channel type coefficient.

9. A parameter adjustment system for a Wi-Fi module, characterized in that: include: A data acquisition unit, used for acquiring original data packets; The preprocessing unit is used to preprocess the original data packet, identify the interference intensity and interference type, perform data annotation, form interference samples, integrate the interference samples, and output a spatiotemporal annotation data set; The interference prediction unit is used to predict the type and intensity of interference in advance based on the spatiotemporal annotated data set, quantify the physical characteristics and probability of future interference, and output the interference prediction results; Interference tensor decomposition and feature fusion unit, used to integrate real-time signals and prediction results to generate anti-interference features; The service perception quality assessment unit is used to dynamically generate QoS vectors based on device type and service requirements, dynamically adjust service weights based on prediction results, integrate anti-interference characteristics with the adjusted QoS vectors, and output the final quality value to evaluate the service quality and dynamically optimize service quality; The optimization and adjustment unit is used to build and update a virtual queue based on the final quality value, prediction results, and current RF parameters, monitor quality deviations, and dynamically adjust RF parameters.

Images