A Deep Learning-Based Adaptive Network Regulation Method and System for Mining Wireless Gateways

Abstract

This application provides a deep learning-based adaptive network adjustment method and system for mining wireless gateways, relating to the field of industrial gateway technology. The method includes: collecting network state parameters of the mining wireless gateway for feature analysis to construct a standardized network dataset; constructing a hybrid deep learning branch, training it using the standardized network dataset, and generating a network parameter prediction channel; analyzing the network state parameters, combining this with a MOSFET switch array for circuit control, amplifying the gain of the mining wireless gateway, and adaptively adjusting the network. This application solves the technical problem in existing technologies where wireless gateways are susceptible to interference in dynamic extreme environments, leading to adjustment lag and passive response, further affecting network performance. Based on deep learning and network state feature analysis, it intelligently optimizes the network performance of mining wireless gateways, improving the stability and efficiency of wireless communication in industrial gateways.

Description

Technical Field

[0001] This application relates to the field of industrial gateway technology, specifically to a deep learning-based adaptive network adjustment method and system for mining wireless gateways. Background Technology

[0002] In the actual deployment of wireless gateways in mines, the tunnel environment is complex and ever-changing, facing complex signal interference and environmental changes, which often leads to unstable wireless communication performance. Due to the special characteristics of the mining environment, such as electromagnetic interference deep underground, the influence of metal equipment, and changes in terrain, existing network adjustment methods often rely on traditional static or simple dynamic adjustment strategies. These methods use preset thresholds or classical control algorithms for lagging adjustments, which are difficult to cope with rapid changes in channel conditions. This often results in adjustment commands lagging behind environmental changes, causing communication link oscillations or even interruptions, severely restricting the stability and real-time performance of wireless communication in mines.

[0003] In summary, existing technologies suffer from the technical problem that wireless gateways are susceptible to interference in dynamic extreme environments, leading to delayed adjustments and passive responses, which further affects the network performance of mining wireless gateways. Summary of the Invention

[0004] The purpose of this application is to provide a deep learning-based adaptive network adjustment method and system for mining wireless gateways, in order to solve the technical problem in the prior art where wireless gateways are susceptible to interference in dynamic extreme environments, resulting in adjustment lag and passive response, which further affects the network performance of mining wireless gateways.

[0005] To achieve the above objectives, this application provides a method and system for adaptive network adjustment of mining wireless gateways based on deep learning.

[0006] Firstly, this application provides a deep learning-based adaptive network adjustment method for mining wireless gateways. This method is implemented through a deep learning-based adaptive network adjustment system for mining wireless gateways. The method includes: collecting network state parameters of the mining wireless gateway for feature analysis to obtain a multi-dimensional network feature dataset for preprocessing, and constructing a standardized network dataset; constructing a hybrid deep learning branch, training the hybrid deep learning branch using the standardized network dataset, and generating a network parameter prediction channel; analyzing the network state parameters through the network parameter prediction channel, outputting network adjustment parameters, and combining this with a MOS transistor switching array for circuit control to obtain a network control strategy for gain amplification of the mining wireless gateway, generating a gain signal for adaptive network adjustment of the mining wireless gateway.

[0007] Optionally, dual-frequency sensor nodes are alternately deployed on the roof and sidewalls of the mine roadway to construct a three-dimensional monitoring network; the data acquisition frequency is set to activate the three-dimensional monitoring network, and multiple sensor nodes are identified for sensor collision analysis; when multiple sensor nodes have data sensing collisions, multiple sensor nodes are connected to the mine wireless gateway for sensing according to the data transmission cycle using the time division multiple access method to obtain the network status parameters of multiple sensor nodes.

[0008] Optionally, the acquisition time data of the three-dimensional monitoring network is divided into equal-length time slots according to the data acquisition frequency; the equal-length time slots are matched with multiple sensor nodes to construct a time slot allocation table; multiple sensor nodes are allocated based on the time slot allocation table to obtain multiple transmission time slots; adjacent time slot overlap analysis is performed by traversing the multiple transmission time slots to obtain data transmission overlap parameters; data transmission conflicts are monitored according to the data transmission overlap parameters to obtain data collision results of multiple sensor nodes.

[0009] Optionally, multi-scale decomposition is performed based on the network state parameters to obtain multi-scale decomposition coefficients; noise identification is performed based on the multi-scale decomposition coefficients to obtain noise coefficient components; the network state parameters are filtered according to the noise coefficient components to obtain effective signal features; and signal propagation ray tracing is performed on the mining wireless gateway according to the effective signal features to obtain the multi-dimensional network feature dataset.

[0010] Optionally, numerical analysis is performed based on the multi-dimensional network feature dataset to obtain numerical distribution characteristic parameters; the min-max standardization method is used to map the numerical distribution characteristic parameters to determine the target mapping interval; Z-score standardization is performed based on the target mapping interval to obtain a standardized grid initial dataset; the standardization effect of the standardized grid initial dataset is verified, and the standardized grid initial dataset is updated in reverse to obtain the standardized network dataset.

[0011] Optionally, multiple convolutional layers and multiple pooling layers are alternately arranged to construct a convolutional neural network branch; a flattening layer is connected to the end of the convolutional neural network branch to generate a one-dimensional feature vector; a multi-layer gated recurrent unit network is set, with each layer of the multi-layer gated recurrent unit network containing multiple gated recurrent units to construct a recurrent neural network branch; a fully connected layer is connected to the end of the recurrent neural network branch to extract temporal features; the one-dimensional feature vector is concatenated with the temporal features to construct the hybrid deep learning branch.

[0012] Optionally, the standardized network dataset is divided into a training set, a validation set, and a test set; the training set, the validation set, and the test set are used to train the convolutional neural network branch and the recurrent neural network branch in segments to obtain multiple segmented training parameters; the convolutional neural network branch and the recurrent neural network branch are unfrozen and combined with the multiple segmented training parameters for end-to-end joint training to generate a joint training result; the standardized network dataset is then optimized and monitored based on the joint training result to construct the network parameter prediction channel.

[0013] Optionally, the network state parameters are input to the network parameter prediction channel: S1: Extract spatial sequence features of the network state parameters through the convolutional neural network branch; S2: Capture temporal sequence features of the network state parameters through the recurrent neural network branch; S3: Fuse the spatial sequence features and the temporal sequence features to generate a comprehensive network state evaluation result; Perform multidimensional network analysis based on the comprehensive network state evaluation result to generate network adjustment parameters; Analyze the network adjustment parameters using the MOS transistor switching array to generate a circuit control signal; Decompose the circuit control signal into multiple adjustment steps, and perform gain control on the circuit control signal according to the multiple adjustment steps to generate a target gain value; Perform progressive gain adjustment on the circuit control signal based on the target gain value to generate the gain signal.

[0014] Optionally, the network power assessment result and channel quality assessment result are obtained by parsing the comprehensive network status assessment result; power analysis is performed based on the network power assessment result to determine the power adjustment level; channel analysis is performed based on the channel quality assessment result to determine the channel switching parameters; the signal received strength is retrieved and the power adjustment level and the channel switching parameters are combined and adjusted to generate the network adjustment parameters.

[0015] Secondly, this application also provides a deep learning-based adaptive network adjustment system for mining wireless gateways, used to execute the deep learning-based adaptive network adjustment method for mining wireless gateways as described in the first aspect. The deep learning-based adaptive network adjustment system for mining wireless gateways includes: a parameter analysis module, used to collect network state parameters of the mining wireless gateway for feature analysis, obtain a multi-dimensional network feature dataset for preprocessing, and construct a standardized network dataset; a channel construction module, used to construct a hybrid deep learning branch, train the hybrid deep learning branch using the standardized network dataset, and generate a network parameter prediction channel; and a network adjustment module, used to analyze the network state parameters through the network parameter prediction channel, output network adjustment parameters, combine a MOS transistor switching array for circuit control, obtain a network control strategy to amplify the gain of the mining wireless gateway, and generate a gain signal to adaptively adjust the network of the mining wireless gateway.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0017] By collecting network state parameters from mining wireless gateways and performing feature analysis, a multi-dimensional network feature dataset is obtained and preprocessed to construct a standardized network dataset. A hybrid deep learning branch is then constructed and trained using the standardized network dataset to generate a network parameter prediction channel. This prediction channel is used to analyze the network state parameters, outputting network adjustment parameters. Combined with a MOSFET switching array for circuit control, a network control strategy is obtained to amplify the gain of the mining wireless gateway, generating a gain signal for adaptive network adjustment. In other words, by collecting network state parameters from mining wireless gateways and performing multi-dimensional feature analysis, a deep learning model is used to identify various interference factors in complex environments. Combined with MOSFET switching arrays for circuit control, intelligent gain amplification is achieved, and changes in network state are predicted in advance for proactive adjustment, thereby improving the stability and efficiency of wireless communication in mining industrial gateways.

[0018] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the adaptive network adjustment method for mining wireless gateways based on deep learning, as described in this application.

[0021] Figure 2 This is a schematic diagram of the adaptive network adjustment system for mining wireless gateways based on deep learning, as described in this application.

[0022] Figure labeling: Parameter analysis module 11, channel construction module 12, network adjustment module 13. Detailed Implementation

[0023] This application provides a deep learning-based adaptive network adjustment method and system for mining wireless gateways. It addresses the technical problem in existing technologies where wireless gateways are susceptible to interference in dynamic extreme environments, leading to adjustment lag and passive response, which further affects the network performance of mining wireless gateways. By collecting network state parameters of the mining wireless gateway and performing multi-dimensional feature analysis, a deep learning model is used to identify various interference factors in complex environments. Combined with MOSFET switching arrays for circuit control, intelligent gain amplification is implemented to predict network state changes in advance and proactively adjust the system, thereby improving the stability and efficiency of wireless communication in mining industrial gateways.

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0025] Example 1, please refer to the appendix. Figure 1 This application provides a deep learning-based adaptive network adjustment method for mining wireless gateways. The method is applied to a deep learning-based adaptive network adjustment system for mining wireless gateways, and specifically includes the following steps:

[0026] By collecting network status parameters from mining wireless gateways and performing feature analysis, a multi-dimensional network feature dataset is obtained, which is then preprocessed to construct a standardized network dataset.

[0027] Furthermore, this application also includes the following steps: alternately deploying dual-frequency sensor nodes on the roof and sidewalls of the mine roadway to construct a three-dimensional monitoring network; setting a data acquisition frequency to activate the three-dimensional monitoring network, identifying multiple sensor nodes for sensor collision analysis; when multiple sensor nodes experience data sensing collisions, accessing the mine wireless gateway with multiple sensor nodes according to the data transmission cycle using the time-division multiple access method to obtain the network status parameters of the multiple sensor nodes.

[0028] Furthermore, this application also includes the following steps: dividing the acquisition time data of the three-dimensional monitoring network into equal-length time slots according to the data acquisition frequency; matching the equal-length time slots with multiple sensor nodes to construct a time slot allocation table; allocating multiple sensor nodes based on the time slot allocation table to obtain multiple transmission time slots; traversing the multiple transmission time slots to perform adjacent time slot overlap analysis to obtain data transmission overlap parameters; monitoring data transmission conflicts according to the data transmission overlap parameters to obtain data collision results of multiple sensor nodes.

[0029] Specifically, in mine tunnels, signal transmission is affected by factors such as terrain, obstacles, and electromagnetic interference. Therefore, a scientific deployment of sensor nodes and data acquisition strategies are needed to ensure the accuracy and stability of monitoring data. Dual-frequency sensor nodes are alternately deployed on the mine roof and sidewalls, forming a three-dimensional monitoring network covering the entire mine tunnel. Mine tunnels are underground passages used for passage and ore transportation during mining operations. The roof refers to the rock strata above the tunnel, while the sidewalls are the rock walls on either side of the tunnel. In mines, the roof and sidewalls are key areas for signal propagation and sensor deployment. Roof nodes are primarily responsible for wide-area coverage, while sidewall nodes can effectively detect signal blind spots near the ground. This layout minimizes monitoring blind spots. Dual-frequency sensor nodes refer to sensor devices that use two frequency bands simultaneously, commonly 2.4GHz and 5GHz. Dual-frequency sensor nodes provide frequency diversity; when one band is interfered with, they can automatically switch to the other, enhancing communication robustness.

[0030] In mine tunnels, multiple sensor nodes need to synchronously collect environmental data and transmit it to a wireless gateway. Due to the limited bandwidth of wireless communication, if multiple sensor nodes transmit data simultaneously, data collisions may occur, affecting data integrity and transmission efficiency. To avoid this problem, time slot allocation is required to ensure that each sensor node transmits data at different times, thereby preventing data collisions.

[0031] The data acquisition frequency is set, which is the number of times the sensor collects data per second. During monitoring, a reasonable acquisition frequency ensures sufficient data acquisition while avoiding excessive useless information from excessively high frequencies. Simultaneously, all nodes of the three-dimensional monitoring network are activated. Based on the data sampling frequency, the acquisition time data of the three-dimensional monitoring network is divided into equal-length time slots, that is, time is divided into small segments of equal length. These equal-length time slots are matched with multiple sensor nodes to construct a time slot allocation table, clearly recording the correspondence between each sensor node and its dedicated transmission time slot. For example, sensor node 1 transmits data in time slot 1, sensor node 2 transmits data in time slot 2, and so on. In this way, different sensor nodes can transmit data in different time periods, avoiding signal conflicts.

[0032] Based on a time slot allocation table, multiple sensor nodes are allocated, with each node assigned its own transmission time slot, resulting in multiple transmission time slots. During time slot allocation, it's crucial to ensure that time slots of adjacent nodes do not overlap. Multiple transmission time slots are traversed to perform adjacent time slot overlap analysis, checking for any overlap. By calculating node clock drift, signal propagation delay, and transceiver switching time, the probability and extent of signal overlap at adjacent time slot boundaries are simulated and evaluated, thus obtaining data transmission overlap parameters. Data transmission overlap parameters are an indicator used to quantify the risk of inter-time slot interference; they can be understood as the length of time two consecutive data packets might overlap at the time slot boundary, typically measured in microseconds. For example, in a cross-section... In a simulated tunnel, 20 dual-frequency sensor nodes were deployed alternately, with a data acquisition frequency of 10Hz. The acquisition time data of the three-dimensional monitoring network was divided into 20 equal-length time slots, each with a length of 5ms. After the time slot allocation table was established, overlap analysis was performed to verify its reliability. It was assumed that the maximum drift of the node crystal oscillator was ±50ppm, the maximum signal propagation delay was 1μs, and the transceiver switching time was 10μs. The worst-case cumulative time error was calculated by analyzing all adjacent time slots (e.g., at the end of time slot 1 and the beginning of time slot 2). It was found that due to clock drift, after one minute of continuous transmission, the time slot boundary between the first and last nodes experienced a high error rate. The deviation was considered. Combined with other factors, the worst-case data transmission overlap parameter was calculated to be 28 μs. Based on this parameter, a high risk of data transmission conflicts was detected, and data collision results were obtained for the current allocation scheme under long-term operation.

[0033] Based on data transmission overlap parameters, potential data transmission conflict risks are monitored, and data collision results for all sensor nodes are obtained. If the overlap parameter exceeds a safety threshold, it indicates a potential problem with the current allocation scheme, requiring readjustment of time slot allocation or the introduction of a protection interval. When data collision results show multiple sensor nodes experiencing data collisions, a scheduling mechanism is activated. According to the data transmission cycle, time division multiple access (TDMA) is used to allocate a dedicated, fixed transmission time slot to each node. In other words, time slots are reallocated to sensor nodes based on the set data transmission cycle. For example, node 1 can be allocated to the first 0.5 seconds of cycle 1, and node 2 can be allocated to the last 0.5 seconds of cycle 1. In this way, the transmission time slots of all nodes are optimized, avoiding time overlap and reducing the possibility of data collisions. Time division multiple access (TDMA) is a communication technology that divides time into multiple time slots and allocates these time slots to different devices for communication. It divides the timeline into periodic frames, and each frame is further divided into several non-overlapping time slots. Each node can only transmit data within its assigned specific time slot, thus fundamentally avoiding data collisions.

[0034] Each node transmits data to the mine's wireless gateway in an orderly manner, as if scheduled to speak at fixed times. Based on the new time slot allocation, it ensures that each sensor node transmits data according to its assigned time slot. The system monitors for data transmission collisions and obtains network status parameters for each sensor node, including indicators reflecting network quality and stability such as signal strength, transmission delay, bandwidth, and packet loss rate. By alternately deploying dual-frequency sensor nodes and utilizing time-division multiple access (TDMA) for data transmission, data transmission collisions can be effectively avoided, improving the access efficiency of the mine's wireless gateway while ensuring accurate acquisition of network status parameters and real-time monitoring of wireless communication quality within the mine roadways.

[0035] Furthermore, this application also includes the following steps: performing multi-scale decomposition based on the network state parameters to obtain multi-scale decomposition coefficients; performing noise identification based on the multi-scale decomposition coefficients to obtain noise coefficient components; filtering the network state parameters according to the noise coefficient components to obtain effective signal features; and performing signal propagation ray tracing on the mining wireless gateway according to the effective signal features to obtain the multi-dimensional network feature dataset.

[0036] Specifically, multi-scale decomposition is performed on network state parameters, breaking down the original data into components at multiple scales. Multi-scale decomposition is a signal processing technique used to decompose an original signal into a set of components (scales) of different frequencies. Wavelet transform is a typical method that can analyze signals simultaneously in the time and frequency domains, thereby revealing short-term abrupt changes (high-frequency details) and long-term trends (low-frequency approximations) in the signal. For example, the original signal may contain some high-frequency noise or low-frequency stable signals. Through multi-scale decomposition, these signals can be separated, and multi-scale decomposition coefficients can be obtained. These coefficients reflect the characteristics of the signal at different scales.

[0037] Noise identification is performed based on multi-scale decomposition coefficients. Noise typically manifests as high-frequency or irregular fluctuations unrelated to the signal. The goal of noise identification is to separate these interfering components from the original signal. Noise identification algorithms identify the noise components in the signal based on the decomposed coefficients and calculate the noise figure components, reflecting the impact of noise on the signal. Noise usually originates from environmental electromagnetic interference or sensor errors. Through noise identification, useful signals and invalid noise can be separated from the original data. The noise figure components obtained through noise identification reflect the magnitude and impact of noise in the signal.

[0038] Filtering network state parameters based on noise figure components involves discarding or reducing noise figure components during signal reconstruction to obtain denoised, more accurate signal characteristics that better reflect real channel variations. These accurate signal characteristics can include parameters such as signal strength, delay, and bandwidth, accurately reflecting the performance of the wireless network. During filtering, noise figure components are subtracted, thereby improving signal quality.

[0039] Ray tracing is performed on the signal propagation of a mining wireless gateway based on the characteristics of the effective signal to simulate the propagation path of the wireless signal from the sensor to the gateway. Ray tracing simulation considers not only the straight-line path of the signal but also factors such as obstacles, reflections, and refractions in the environment. Through ray tracing, the propagation of the signal in the actual environment can be obtained, thereby evaluating the coverage and transmission quality of the wireless network. In wireless communication, signal propagation ray tracing is a simulation technique used to calculate the propagation path and process of a wireless signal from the transmitter to the receiver. For example, suppose 50 sensor nodes are deployed in a mine, the network sampling frequency is 20 times per second, and the signal strength data of the sensor nodes contains noise within one cycle. The network state parameters are decomposed into multiple scales, with five frequency bandwidths set as the decomposition scales. After decomposition, the decomposition coefficients are as follows: Scale 1 (low frequency) has an average signal strength of -60dBm and a delay of 50ms; Scale 2 (mid frequency) has an average signal strength of -62dBm and a delay of 55ms; Scale 3 (high frequency) has an average signal strength of -65dBm and a delay of 60ms. Noise identification revealed approximately 5% noise interference in the high-frequency component, with a noise figure of 0.2. High-frequency noise components were then removed through signal filtering, yielding the effective signal characteristics: a signal strength of -62 dBm and a delay of 52 ms. Signal propagation ray tracing simulation was performed, simulating the effect of multiple obstacles in the mine. Ray tracing results showed that the signal strength reaching the mine's wireless gateway after reflection and refraction was -58 dBm, with a delay of 54 ms.

[0040] By employing multi-scale decomposition technology, signal features at different levels are extracted from complex network state parameters, thereby identifying and removing noise from the signal. The noise identification and filtering steps effectively improve signal quality and remove interference factors. Signal propagation ray tracing technology can simulate signal propagation paths in real-world environments, helping to evaluate the coverage and performance of wireless networks.

[0041] Furthermore, this application also includes the following steps: performing numerical analysis based on the multi-dimensional network feature dataset to obtain numerical distribution characteristic parameters; using the min-max standardization method to map the numerical distribution characteristic parameters to determine the target mapping interval; performing Z-score standardization calculation based on the target mapping interval to obtain a standardized grid initial dataset; verifying the standardization effect of the standardized grid initial dataset, and updating the standardized grid initial dataset in reverse to obtain the standardized network dataset.

[0042] Specifically, numerical analysis is performed on a multi-dimensional network feature dataset to obtain the numerical distribution characteristic parameters for each feature dimension, including the global minimum, maximum, mean, and standard deviation. These numerical distribution characteristic parameters describe the data distribution characteristics, such as the minimum, maximum, mean, and standard deviation. The min-max standardization method is used to map the numerical distribution characteristic parameters to determine the target mapping interval. By calculating the minimum and maximum values ​​of each feature parameter, the standardized numerical range is set, typically [0,1]. A linear transformation formula is then used to transform the original data to the target mapping interval, handling data boundary conditions and preventing data overflow after standardization. For example, suppose the multi-dimensional network feature dataset contains three features: signal strength (-100dBm to -50dBm), multipath delay spread (10ns to 500ns), and channel load rate (0% to 100%). Numerical analysis was performed on a multi-dimensional network feature dataset. The signal strength had a minimum value of -98, a maximum value of -52, a mean of -75, and a standard deviation of 10; the delay spread had a minimum value of 15, a maximum value of 480, a mean of 150, and a standard deviation of 100; and the load rate had a minimum value of 0, a maximum value of 95, a mean of 30, and a standard deviation of 20. All three features were linearly mapped to the interval [0,1]. A signal strength value of -80dBm was converted to (-80 - (-98)) / (-52 - (-98)) = 18 / 46 = 0.39.

[0043] Z-score standardization is performed based on the mapped interval to make the distribution of each feature have a mean of 0 and a standard deviation of 1, thus obtaining a standardized grid initial dataset that is both dimensionless and conforms to a standard normal distribution. Z-score standardization is a data standardization method that makes the processed data have a mean of 0 and a standard deviation of 1. For example, to calculate the new mean and standard deviation of data in the interval [0,1], assuming the mean of the signal strength is 0.5 and the standard deviation is 0.2, then 0.39 is further transformed into (0.39-0.5) / 0.2=-0.55.

[0044] The initial standardized grid dataset is validated to check if it meets the expected standards, i.e., whether the data has been sufficiently standardized and without distortion due to computational errors during the standardization process. For example, if the numerical range of some features is too narrow during standardization, resulting in poor standardization, these data can be further adjusted or remapped. After validation, the initial standardized grid dataset is updated as necessary, including adjusting outlier data points or adjusting the data mapping according to new computational methods to ensure that the data accurately reflects the characteristics of the original data. For example, it checks whether there are features with a standard deviation close to 0 (meaning the feature is invalid) or whether there is distribution distortion due to extreme values. Based on the validation results, the initial standardized grid dataset is updated in reverse, performing a logarithmic transformation on the feature before the min-max step, and then re-executing the two-stage standardization process. The final standardized network dataset shows a good distribution with a mean of 0 and a standard deviation of 1 for all features. Through min-max standardization and Z-score standardization, the original network state feature data is transformed into a unified standard format, eliminating scale differences between features.

[0045] Construct a hybrid deep learning branch, train the hybrid deep learning branch using the standardized network dataset, and generate network parameter prediction channels.

[0046] Furthermore, this application also includes the following steps: alternating multiple convolutional layers and multiple pooling layers to construct a convolutional neural network branch; connecting a flattening layer to the end of the convolutional neural network branch to generate a one-dimensional feature vector; setting a multi-layer gated recurrent unit network, each layer of the multi-layer gated recurrent unit network containing multiple gated recurrent units to construct a recurrent neural network branch; connecting a fully connected layer to the end of the recurrent neural network branch to extract temporal features; concatenating the one-dimensional feature vector with the temporal features to construct the hybrid deep learning branch.

[0047] Specifically, multiple convolutional and pooling layers are alternately arranged to construct convolutional neural network branches, such as Conv1D→BN→ReLU→MaxPooling→Conv1D→BN→ReLU→MaxPooling. Each convolutional neural network branch includes multiple convolutional and pooling layers arranged alternately. Each convolutional layer uses a different sized convolutional kernel to extract spatial features at different scales. Larger kernels can extract a wider range of features, making them suitable for capturing spatial information over larger regions. Through multiple convolutional layers, multi-scale features from local to global perspectives can be captured.

[0048] Convolutional layers of varying sizes are responsible for extracting multi-scale spatial features. Batch normalization layers are placed after convolutions and before activation functions to accelerate model training and improve stability. Modified linear units are used as activation functions to introduce non-linearity. Subsequent max pooling preserves salient features and reduces dimensionality, thereby reducing computational cost and the risk of overfitting. The ends of the convolutional neural network branches are connected to flattening layers, transforming the final feature maps into one-dimensional feature vectors. Flattening layers flatten multi-dimensional input data into one-dimensional data for input into fully connected layers. In convolutional neural networks, flattening layers are typically located between convolutional layers and fully connected layers.

[0049] A multi-layer gated recurrent unit (GRU) network is constructed to build a branch of a recurrent neural network. The GRUs finely control the flow of information through their reset and update gates. Dropout layers are added to prevent overfitting, and the tanh activation function is used to process candidate states of the GRUs to generate new candidate memories. GRUs are a variant of recurrent neural networks that, by introducing reset and update gates, effectively capture long-term dependencies in time series while mitigating the vanishing / exploding gradient problem of traditional recurrent neural networks. The reset gate determines how new inputs are combined with previous memories, while the update gate determines how much of the past memory is retained. Through the control of reset and update gates, the multi-layer GRU network enables the model to remember useful information and forget irrelevant information.

[0050] A fully connected layer is connected to the end of the recurrent neural network branch to map the temporal data to the target dimension and extract the temporal features from the network. The one-dimensional feature vector output by the convolutional neural network branch is concatenated with the temporal features output by the recurrent neural network branch to form a comprehensive feature vector that integrates spatial and temporal information, thereby constructing a hybrid deep learning branch of convolutional neural network-long short-term memory network based on the attention mechanism.

[0051] For example, to predict signal strength in the next second, the input to a convolutional neural network-long short-term memory (LSM) network is standardized network features from the past 3 seconds (30 time points), such as signal strength, signal-to-noise ratio (SNR), and bit error rate (BER). The input data has a shape of (30, 3). It first passes through a Conv1D layer with 16 convolutional kernels of size 3 to extract local correlation features, followed by Batch Normalization (BN) and ReLU, and then downsampled using MaxPooling (pool_size=2). This is followed by another Conv1D layer with 32 convolutional kernels of size 5, then BN, ReLU, and MaxPooling. Finally, a flattening layer yields a one-dimensional feature vector with approximately 100 elements. The same input sequence is fed into a two-layer multi-level gated recurrent unit (GRU) network, with 50 units per layer and Dropout=0.2 between layers. The GRU learns the temporal dependencies of the sequence through its gating mechanism. The output of the final GRU layer is fed into a fully connected layer to extract a 50-dimensional temporal feature. The 100-dimensional vector from the convolutional neural network branch and the 50-dimensional vector from the recurrent neural network branch are concatenated into a 150-dimensional composite feature vector. An attention mechanism is introduced onto this fused feature to assign appropriate weights to the outputs of the recurrent neural network branches at different time steps, thereby allowing the convolutional neural network-long short-term memory network to focus more on the most critical historical moments for prediction.

[0052] By combining convolutional neural networks (CNNs) and recurrent neural networks (RNNs), this hybrid deep learning model can simultaneously extract spatial and temporal features. CNNs effectively capture the spatial structure of the input data, while RNNs handle the temporal dependencies within the data. Through an attention mechanism, the model can focus on key information, further enhancing its ability to extract spatiotemporal features and thus improving the network's performance on complex spatiotemporal data.

[0053] Furthermore, this application also includes the following steps: dividing the standardized network dataset into a training set, a validation set, and a test set; using the training set, the validation set, and the test set to perform segmented training on the convolutional neural network branch and the recurrent neural network branch respectively, to obtain multiple segmented training parameters; unfreezing the convolutional neural network branch and the recurrent neural network branch and performing end-to-end joint training with the multiple segmented training parameters to generate joint training results; and performing optimization monitoring training on the standardized network dataset based on the joint training results to construct the network parameter prediction channel.

[0054] Specifically, the standardized network dataset is divided into a training set, a validation set, and a test set in a certain ratio (e.g., 70:15:15). The training set is used to directly train the model parameters; the validation set is used to monitor the model performance during training, adjust hyperparameters, and select the best model to prevent overfitting; the test set is used to finally evaluate the model's generalization performance, reflecting its true performance on unknown data.

[0055] The convolutional neural network (CNN) and recurrent neural network (RNN) branches are trained in segments using training, validation, and test sets, respectively. Segmented training involves dividing the entire training process into multiple stages, with each stage training a different part separately and obtaining multiple segmented training parameters. When training the CNN branch, the training process primarily focuses on spatial feature extraction. The data in the training set is processed through convolutional and pooling layers, followed by batch normalization and activation functions to obtain spatial feature representations. When training the RNN branch, the training process focuses on temporal feature extraction. The RNN branch uses gating mechanisms to capture temporal patterns in the data and learn long-short-term dependencies. In other words, the CNN and RNN branches are trained independently or semi-independently using training sets. For example, the RNN branch can be fixed first, and the CNN branch can be trained using only the training set; then the CNN branch can be fixed, and the RNN branch can be trained using only the training set. This allows both branches to initially master the ability to extract spatial and temporal features and obtain their respective segmented training parameters.

[0056] Unfreezing the two branches allows previously frozen network layers to re-participate in training, enabling them to adjust parameters and improve model performance during joint training. End-to-end joint training combines the convolutional neural network (CNN) and recurrent neural network (RNN) branches with multiple segmented training parameters, optimizing their parameters. Through end-to-end training, the two networks influence each other and learn together, further enhancing the model's overall capabilities. During joint training, spatial features extracted by the CNN and temporal features captured by the RNN are fused in the model, forming a more powerful joint network. During training, the model gradually adjusts its parameters based on feedback from the training and validation sets, ultimately yielding optimized joint training results.

[0057] Based on the joint training results, in subsequent training cycles, the standardized network dataset is optimized and monitored to further improve the network's performance in dynamic environments, enhance its real-time prediction and adjustment capabilities, construct a network parameter prediction channel, receive input network state parameters, and generate predicted values ​​about the network status.

[0058] By employing a phased, post-joint training strategy, the problems of training difficulties, overfitting, and unstable convergence in complex hybrid models are addressed. Segmented training ensures that the two functionally distinct sub-networks—the convolutional neural network branch and the recurrent neural network branch—lay a solid foundation for feature extraction, avoiding gradient flow instability or branch laziness that may occur in the early stages of end-to-end training. Joint training then proceeds from this starting point, allowing for fine-tuning of network parameter prediction channels from a high starting point, accelerating convergence, and making it easier to find a superior global optimum or strong local optimum.

[0059] The network state parameters are analyzed through the network parameter prediction channel, and network adjustment parameters are output. Combined with the MOS transistor switching array for circuit control, a network control strategy is obtained to amplify the gain of the mining wireless gateway and generate a gain signal for adaptive network adjustment of the mining wireless gateway.

[0060] Furthermore, this application also includes the following steps: inputting the network state parameters into the network parameter prediction channel: S1: extracting spatial sequence features of the network state parameters through the convolutional neural network branch; S2: capturing temporal sequence features of the network state parameters through the recurrent neural network branch; S3: fusing the spatial sequence features and the temporal sequence features to generate a comprehensive network state evaluation result; performing multidimensional network analysis based on the comprehensive network state evaluation result to generate network adjustment parameters; analyzing the network adjustment parameters using the MOS transistor switching array to generate a circuit control signal; decomposing the circuit control signal into multiple adjustment steps, performing gain control on the circuit control signal according to the multiple adjustment steps to generate a target gain value; performing progressive gain adjustment on the circuit control signal based on the target gain value to generate the gain signal.

[0061] Furthermore, this application also includes the following steps: analyzing the comprehensive network status assessment results to obtain network power assessment results and channel quality assessment results; performing power analysis based on the network power assessment results to determine the power adjustment level; performing channel analysis based on the channel quality assessment results to determine the channel switching parameters; and combining and adjusting the power adjustment level and the channel switching parameters by retrieving the signal reception strength to generate the network adjustment parameters.

[0062] Specifically, the network state parameters of the collected mining wireless gateway are input into the network parameter prediction channel and then processed in parallel. Spatial sequence features of the network state parameters are extracted through a convolutional neural network (CNN) branch. The CNN kernels primarily slide along the time step dimension, focusing on analyzing the interrelationships and local combination patterns among the three parameters—signal strength, signal-to-noise ratio (SNR), and bit error rate (BER)—at the same time point. For example, it identifies typical transient interference features such as a relatively acceptable SNR but a sudden increase in BER. Temporal sequence features of the network state parameters are captured through a recurrent neural network (RNN) branch. A multi-layer gated recurrent unit network processes the input at each time step sequentially, using its internal gating mechanism to remember and forget information, thereby capturing temporal evolution patterns such as a continuous, slow decline in signal strength over the past five time steps or regular fluctuations in signal quality with a 2-second period.

[0063] This method fuses the spatial sequence features representing instantaneous states from the output of convolutional neural network (CNN) branches with the temporal sequence features representing historical patterns from the output of recurrent neural network (RNN) branches. Through a concatenation operation, a comprehensive network state evaluation result is generated. For example, the spatial features obtained from the CNN and the temporal features obtained from the RNN are concatenated into a new feature vector, which is then further processed through a fully connected layer to finally output the comprehensive network state evaluation result. This comprehensive network state evaluation result encompasses various aspects of network performance, such as signal strength, latency, and bandwidth utilization.

[0064] The comprehensive network status assessment results are analyzed to obtain network power assessment results and channel quality assessment results. Network power assessment results are an evaluation of the signal transmission power level required to maintain or achieve the target communication quality in the current and future period. This is typically related to network signal strength, transmission distance, and communication quality, reflecting the power requirements of network equipment at a given moment. Channel quality assessment results are a comprehensive evaluation of the quality of the wireless channel currently and in the future. This includes not only considering the current state but also predicting its future availability, including predictions of interference levels, multipath effects, and bandwidth utilization. This helps determine whether the channel is suitable for current data transmission and is usually quantified using indicators such as signal-to-noise ratio (SNR) or bit error rate (BER).

[0065] Power analysis is performed based on network power assessment results. A specific power adjustment level is determined according to the severity of the power deficit. For example, the required 5dB power increase is mapped to a level 3 power increase command. The power adjustment level is determined by dividing the power adjustment amount into several discrete levels based on the power assessment results, such as fine-tuning, medium-level increase, and strong-level increase. This helps simplify the control logic and improve stability.

[0066] Channel analysis is performed based on channel quality assessment results. From a pre-set list of clean channels, an optimal channel switching parameter is determined. When a channel switch is deemed necessary, the switching parameter specifies the target channel, determining whether a switch is needed and which channel to switch to. Channel switching is typically triggered by factors such as a signal-to-noise ratio below a certain threshold or a high bit error rate. If the current channel quality is below a preset standard, the model will automatically recommend switching to a channel with higher quality.

[0067] The system retrieves the received signal strength in real time and combines the power adjustment level and channel switching parameters for comprehensive adjustment. This means using the received signal strength as input to help determine the current network condition. Based on the comprehensive analysis of power adjustment and channel switching, corresponding network adjustment parameters are generated, including transmission power adjustment values, channel switching commands, and PGA gain control values. The transmission power adjustment value indicates the network transmission power that needs adjustment, and whether to increase or decrease power output is determined based on network assessment results. The channel switching command instructs the network to switch to another channel, typically triggered when the current channel quality is poor. The PGA gain control value controls the signal gain, typically used to adjust the gain of the signal amplifier to optimize signal transmission quality.

[0068] The received network adjustment parameters are sent to the control logic unit of the MOSFET switch array for parsing. The MOSFET driver circuit generates corresponding gate control voltages to precisely control the conduction state of each switch in the MOSFET switch array, thereby changing the configuration of the RF front-end matching network, attenuation network, or amplification circuit. In other words, the MOSFET switch array is used to control power and gain regulation. Based on the adjustment parameters, the gate voltage of the MOSFETs is adjusted to control the switching on and off. Through gate voltage control, the switching state of the MOSFETs can be adjusted in real time. Different switching states correspond to different power output and signal gain levels. To ensure adjustment accuracy, the output power value is monitored in real time, and the monitoring results are fed back to the control logic, forming a closed-loop power control loop. In other words, dynamic adjustments are made based on the difference between the actual power value and the target power value, forming a closed-loop power control loop. The circuit control signal is a signal generated based on the network adjustment parameters, used to control the circuit's operating state, directly affecting power output, signal gain, channel selection, etc.

[0069] The circuit control signal is decomposed into multiple adjustment steps, and gain control is applied to the circuit control signal according to these multiple adjustment steps to generate a target gain value. These multiple adjustment steps are the step sizes for gain or power adjustment, typically adjusted gradually in preset increments to adjust the target parameter. Smaller step sizes result in higher adjustment precision. The target gain value determines the final signal amplification factor. After the target gain value is generated, progressive gain adjustment is performed, allowing the gain value to gradually transition towards the target gain value, avoiding over-adjustment and ensuring a smooth improvement in signal transmission quality. Through progressive gain adjustment, signal quality is stably optimized in a real-time network environment.

[0070] The circuit control signal is progressively adjusted based on the target gain value to generate a gain signal that reflects the final adjustment result of power and gain in the network. This signal is then transmitted to the network device as a feedback signal to optimize its signal transmission.

[0071] For example, the network adjustment parameters include power adjustment +7dB, channel switching: channel 36, and PGA gain +3dB. The analysis process of the MOSFET switch array is as follows: For power adjustment, the drive circuit generates a specific combination of gate control voltages to control different conduction states of the four MOSFET switches, switching the operating bias point of the transmitter power amplifier from Class A to Class AB, achieving a +5dB adjustment; simultaneously, the other two MOSFETs switch the π-type attenuation network on the RF path, providing an additional +2dB adjustment. Real-time monitoring of the power meter readings forms a closed-loop control, ensuring that the synthesized output power is accurately +7dB. For the +3dB adjustment of the PGA gain, to avoid instantaneous changes, it is decomposed into six 0.5dB adjustment steps, executed every 10ms. Gain control is performed according to this step sequence, generating a gradual change curve from the current value to the target gain value. Based on this target value, the gain is gradually adjusted, and the required gain signal is smoothly generated within 60ms, with a smooth transition in received signal strength without any glitches or oscillations.

[0072] Adaptive network adjustment of the mining wireless gateway based on gain signal ensures optimal communication link performance in complex mining environments, thereby maintaining stable and efficient communication quality. The system automatically adjusts network transmission power, channel selection, and gain control based on real-time network conditions. Precise control using a MOSFET switching array effectively regulates network signal strength and quality, ensuring stable network performance in dynamically changing and complex environments. Adaptive gain signal adjustment enables real-time network optimization based on environmental changes, improving network coverage, stability, and overall communication quality.

[0073] In summary, the deep learning-based adaptive network adjustment method for mining wireless gateways provided in this application has the following technical effects: By collecting network state parameters of the mining wireless gateway and performing feature analysis, a multi-dimensional network feature dataset is obtained and preprocessed to construct a standardized network dataset; a hybrid deep learning branch is constructed, and the standardized network dataset is used to train the hybrid deep learning branch to generate a network parameter prediction channel; the network state parameters are analyzed through the network parameter prediction channel to output network adjustment parameters; combined with a MOSFET switch array for circuit control, a network control strategy is obtained to amplify the gain of the mining wireless gateway, generating a gain signal for adaptive network adjustment. In other words, by collecting network state parameters of the mining wireless gateway and performing multi-dimensional feature analysis, a deep learning model is used to identify various interference factors in complex environments. Combined with a MOSFET switch array for circuit control, intelligent gain amplification is performed, and changes in network state are predicted in advance for proactive adjustment, thereby improving the stability and efficiency of wireless communication for mining industrial gateways.

[0074] Example 2: Based on the same inventive concept as the deep learning-based adaptive network adjustment method for mining wireless gateways in Example 1, this application also provides a deep learning-based adaptive network adjustment system for mining wireless gateways. Please refer to the appendix. Figure 2 The deep learning-based adaptive network adjustment system for mining wireless gateways includes:

[0075] The parameter analysis module 11 is used to perform feature analysis by collecting network state parameters of the mining wireless gateway, obtain a multi-dimensional network feature dataset for preprocessing, and construct a standardized network dataset. The channel construction module 12 is used to construct a hybrid deep learning branch, train the hybrid deep learning branch using the standardized network dataset, and generate a network parameter prediction channel. The network adjustment module 13 is used to analyze the network state parameters through the network parameter prediction channel, output network adjustment parameters, combine them with a MOS transistor switching array for circuit control, obtain a network control strategy to amplify the gain of the mining wireless gateway, and generate a gain signal to perform adaptive network adjustment of the mining wireless gateway.

[0076] Furthermore, the parameter analysis module 11 in the deep learning-based adaptive network adjustment system for mine wireless gateways is also used for: alternately deploying dual-frequency sensor nodes on the roof and sidewalls of the mine roadway to construct a three-dimensional monitoring network; setting a data acquisition frequency to activate the three-dimensional monitoring network, identifying multiple sensor nodes for sensor collision analysis; when multiple sensor nodes experience data sensing collisions, accessing the mine wireless gateway with multiple sensor nodes according to the data transmission cycle using time-division multiple access method to obtain the network status parameters of the multiple sensor nodes.

[0077] Furthermore, the parameter analysis module 11 in the deep learning-based adaptive network adjustment system for mining wireless gateways is also used for: dividing the acquisition time data of the three-dimensional monitoring network into equal-length time slots according to the data acquisition frequency; matching the equal-length time slots with multiple sensor nodes to construct a time slot allocation table; allocating multiple sensor nodes based on the time slot allocation table to obtain multiple transmission time slots; traversing the multiple transmission time slots to perform adjacent time slot overlap analysis to obtain data transmission overlap parameters; and monitoring data transmission conflicts according to the data transmission overlap parameters to obtain data collision results of multiple sensor nodes.

[0078] Furthermore, the parameter analysis module 11 in the deep learning-based adaptive network adjustment system for mining wireless gateways is also used for: performing multi-scale decomposition based on the network state parameters to obtain multi-scale decomposition coefficients; performing noise identification based on the multi-scale decomposition coefficients to obtain noise coefficient components; filtering the network state parameters according to the noise coefficient components to obtain effective signal features; and performing signal propagation ray tracing on the mining wireless gateway according to the effective signal features to obtain the multi-dimensional network feature dataset.

[0079] Furthermore, the parameter analysis module 11 in the deep learning-based adaptive network adjustment system for mining wireless gateways is also used for: performing numerical analysis based on the multi-dimensional network feature dataset to obtain numerical distribution characteristic parameters; using the min-max standardization method to map the numerical distribution characteristic parameters to determine the target mapping interval; performing Z-score standardization calculation based on the target mapping interval to obtain a standardized grid initial dataset; verifying the standardization effect of the standardized grid initial dataset, and updating the standardized grid initial dataset in reverse to obtain the standardized network dataset.

[0080] Furthermore, the channel construction module 12 in the deep learning-based adaptive network adjustment system for mining wireless gateways is also used for: alternating multiple convolutional layers and multiple pooling layers to construct a convolutional neural network branch; connecting a flattening layer to the end of the convolutional neural network branch to generate a one-dimensional feature vector; setting up a multi-layer gated recurrent unit network, each layer of which contains multiple gated recurrent units to construct a recurrent neural network branch; connecting a fully connected layer to the end of the recurrent neural network branch to extract temporal features; and concatenating the one-dimensional feature vector with the temporal features to construct the hybrid deep learning branch.

[0081] Furthermore, the channel construction module 12 in the deep learning-based adaptive network adjustment system for mining wireless gateways is also used to: divide the standardized network dataset into a training set, a validation set, and a test set; use the training set, the validation set, and the test set to perform segmented training on the convolutional neural network branch and the recurrent neural network branch respectively to obtain multiple segmented training parameters; unfreeze the convolutional neural network branch and the recurrent neural network branch and perform end-to-end joint training with the multiple segmented training parameters to generate joint training results; and perform optimization monitoring training on the standardized network dataset based on the joint training results to construct the network parameter prediction channel.

[0082] Furthermore, the network adjustment module 13 in the deep learning-based adaptive network adjustment system for mining wireless gateways is also used for: inputting the network state parameters to the network parameter prediction channel: S1: extracting spatial sequence features of the network state parameters through the convolutional neural network branch; S2: capturing temporal sequence features of the network state parameters through the recurrent neural network branch; S3: fusing the spatial sequence features with the temporal sequence features to generate a comprehensive network state evaluation result; performing multidimensional network analysis based on the comprehensive network state evaluation result to generate network adjustment parameters; analyzing the network adjustment parameters using the MOS transistor switching array to generate a circuit control signal; decomposing the circuit control signal into multiple adjustment steps, performing gain control on the circuit control signal according to the multiple adjustment steps to generate a target gain value; and performing progressive gain adjustment on the circuit control signal based on the target gain value to generate the gain signal.

[0083] Furthermore, the network adjustment module 13 in the deep learning-based adaptive network adjustment system for mining wireless gateways is also used for: analyzing the comprehensive network status evaluation results to obtain network power evaluation results and channel quality evaluation results; performing power analysis based on the network power evaluation results to determine the power adjustment level; performing channel analysis based on the channel quality evaluation results to determine the channel switching parameters; and combining and adjusting the power adjustment level and the channel switching parameters by retrieving the signal reception strength to generate the network adjustment parameters.

[0084] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Figure 1The deep learning-based adaptive network adjustment method and specific examples for mining wireless gateways in Example 1 are also applicable to the deep learning-based adaptive network adjustment system for mining wireless gateways in this embodiment. Through the foregoing detailed description of the deep learning-based adaptive network adjustment method for mining wireless gateways, those skilled in the art can clearly understand the deep learning-based adaptive network adjustment system for mining wireless gateways in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0085] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0086] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A deep learning-based adaptive network adjustment method for mining wireless gateways, characterized in that, include: By collecting network status parameters of mining wireless gateways for feature analysis, a multi-dimensional network feature dataset is obtained, which is then preprocessed to construct a standardized network dataset. Construct a hybrid deep learning branch, train the hybrid deep learning branch using the standardized network dataset, and generate network parameter prediction channels; The network state parameters are analyzed through the network parameter prediction channel, and network adjustment parameters are output. The circuit is controlled by combining the MOS transistor switching array to obtain the network control strategy. The gain of the mining wireless gateway is amplified, and the gain signal is generated to perform adaptive network adjustment of the mining wireless gateway. The process of constructing hybrid deep learning branches includes the following methods: Multiple convolutional layers and multiple pooling layers are arranged alternately to construct branches of a convolutional neural network; A flattening layer is connected to the end of the branch of the convolutional neural network to generate a one-dimensional feature vector; A multi-layer gated recurrent unit network is configured, wherein each layer of the multi-layer gated recurrent unit network contains multiple gated recurrent units, thereby constructing a recurrent neural network branch; A fully connected layer is connected to the end of the recurrent neural network branch to extract temporal features; The one-dimensional feature vector is concatenated with the temporal features to construct the hybrid deep learning branch; The network state parameters are analyzed through the network parameter prediction channel to output network adjustment parameters. Combined with a MOS transistor switching array for circuit control, a network control strategy is obtained to amplify the gain of the mining wireless gateway and generate a gain signal. The method includes: The network state parameters are input into the network parameter prediction channel: S1: Extract spatial sequence features of network state parameters through the branches of the convolutional neural network; S2: Capture the time-series features of the network state parameters through the branches of the recurrent neural network; S3: The spatial sequence features and the time sequence features are fused to generate a comprehensive evaluation result of the network state; Based on the comprehensive evaluation results of the network state, multidimensional network analysis is performed to generate network adjustment parameters; The circuit control signal is generated by analyzing the MOS transistor switching array in conjunction with the network adjustment parameters. The circuit control signal is decomposed into multiple adjustment steps, and the gain of the circuit control signal is controlled according to the multiple adjustment steps to generate a target gain value. The gain signal is generated by progressively adjusting the gain of the circuit control signal based on the target gain value.

2. The adaptive network adjustment method for mining wireless gateways based on deep learning as described in claim 1, characterized in that, The process and methods for collecting network status parameters of a mining wireless gateway include: Dual-frequency sensor nodes are alternately deployed on the roof and sidewalls of mine roadways to construct a three-dimensional monitoring network; Set the data acquisition frequency to activate the three-dimensional monitoring network, and identify multiple sensor nodes for sensor collision analysis; When multiple sensor nodes experience data sensing collisions, the multiple sensor nodes are connected to the mining wireless gateway using time-division multiple access according to the data transmission cycle to obtain the network status parameters of the multiple sensor nodes.

3. The adaptive network adjustment method for mining wireless gateways based on deep learning as described in claim 2, characterized in that, The method includes setting a data acquisition frequency to activate the three-dimensional monitoring network and identifying multiple sensor nodes for sensor collision analysis. The data acquisition time data of the three-dimensional monitoring network is divided into equal-length time slots according to the data acquisition frequency. The equal-length time slots are matched with multiple sensing nodes to construct a time slot allocation table; Multiple sensor nodes are allocated based on the time slot allocation table to obtain multiple transmission time slots; By traversing the multiple transmission time slots, adjacent time slot overlap analysis is performed to obtain data transmission overlap parameters; Based on the data transmission overlap parameter, data transmission conflicts are monitored to obtain data collision results from multiple sensor nodes.

4. The adaptive network adjustment method for mining wireless gateways based on deep learning as described in claim 1, characterized in that, Feature analysis was performed by collecting network status parameters from mining wireless gateways to obtain a multi-dimensional network feature dataset. The methods included: Multi-scale decomposition is performed based on the network state parameters to obtain multi-scale decomposition coefficients. Noise is identified based on the multi-scale decomposition coefficients to obtain noise figure components; The network state parameters are filtered according to the noise figure components to obtain effective signal characteristics; Based on the effective signal characteristics, signal propagation ray tracing is performed on the mining wireless gateway to obtain the multi-dimensional network feature dataset.

5. The adaptive network adjustment method for mining wireless gateways based on deep learning as described in claim 1, characterized in that, Methods include: obtaining multi-dimensional network feature datasets, preprocessing them, and constructing standardized network datasets. Numerical analysis is performed on the multi-dimensional network feature dataset to obtain numerical distribution characteristic parameters. The min-max normalization method is used to map the numerical distribution characteristic parameters to determine the target mapping interval. Z-score standardization calculation is performed based on the target mapping interval to obtain the initial standardized grid dataset; Verify the standardization effect of the initial standardized grid dataset, and update the initial standardized grid dataset in reverse to obtain the standardized network dataset.

6. The adaptive network adjustment method for mining wireless gateways based on deep learning as described in claim 1, characterized in that, The method for training the hybrid deep learning branch using the standardized network dataset to generate network parameter prediction channels includes: The standardized network dataset is divided into a training set, a validation set, and a test set; The training set, the validation set, and the test set are used to train the convolutional neural network branch and the recurrent neural network branch in segments to obtain multiple segmented training parameters; Unfreeze the convolutional neural network branch, combine the recurrent neural network branch with the multiple segmented training parameters, and perform end-to-end joint training to generate joint training results; Based on the joint training results, the standardized network dataset is optimized and monitored for training to construct the network parameter prediction channel.

7. The adaptive network adjustment method for mining wireless gateways based on deep learning as described in claim 1, characterized in that, Based on the comprehensive network state evaluation results, multidimensional network analysis is performed to generate network adjustment parameters. The method includes: The network power assessment result and the channel quality assessment result are obtained by analyzing the comprehensive network status assessment result. Based on the network power assessment results, power analysis is performed to determine the power adjustment level. Based on the channel quality assessment results, channel analysis is performed to determine the parameters to be switched to the channel. The signal reception strength is retrieved, and the power level to be adjusted and the channel switching parameters to be combined and adjusted to generate the network adjustment parameters.

8. A deep learning-based adaptive network adjustment system for mining wireless gateways, characterized in that: The step of implementing the deep learning-based adaptive network adjustment method for mining wireless gateways according to any one of claims 1 to 7, wherein the deep learning-based adaptive network adjustment system for mining wireless gateways comprises: The parameter analysis module is used to perform feature analysis by collecting network status parameters of the mining wireless gateway, obtain a multi-dimensional network feature dataset for preprocessing, and construct a standardized network dataset. The channel construction module is used to construct a hybrid deep learning branch, train the hybrid deep learning branch using the standardized network dataset, and generate network parameter prediction channels. The network adjustment module is used to analyze the network status parameters through the network parameter prediction channel, output network adjustment parameters, combine with the MOS transistor switching array for circuit control, obtain a network control strategy to amplify the gain of the mining wireless gateway, and generate a gain signal to perform adaptive network adjustment of the mining wireless gateway.

Images