Four-dimensional dynamic adaptive gating fusion method for mine communication

Abstract

The invention discloses a four-dimensional dynamic adaptive gating fusion method for mine communication, and belongs to the technical field of radio transmission. Comprising the following steps: step 1, carrying out four-dimensional state sensing to obtain a four-dimensional channel state vector s; 2, performing nonlinear normalized mapping on the four-dimensional channel state vector s; 3, inputting the normalized features in the step 2 into an ultra-lightweight neural network for weight decision making; and 4, applying stability constraint to the weight decided in the step 3, and realizing self-adaptive fusion of the output of the physical model and the data model. According to the method, the four-dimensional state of the channel can be sensed in real time, the signal compensation parameters are dynamically adjusted according to the four-dimensional state sensed in real time, the method plays a leading role in signal compensation, and the bit error rate in a mine communication scene is remarkably reduced. The problems that a traditional hybrid drive architecture fusion strategy cannot adapt to second-level violent changes of the underground channel state and is difficult to meet the real-time requirement of underground intrinsic safety equipment are solved.

Description

Technical Field

[0001] This invention belongs to the field of radio transmission technology, specifically relating to a four-dimensional dynamic adaptive gating fusion method for mine communication. Background Technology

[0002] With the accelerated advancement of intelligent mine construction, the demand for high reliability, low latency, and strong environmental adaptability in underground wireless communication systems is becoming increasingly urgent. In the complex and ever-changing tunnel environment, factors such as electromagnetic interference, dust obstruction, high-speed equipment movement, and multipath reflection cause the channel state to fluctuate drastically at the millisecond level. Traditional single-drive architectures (pure physical models or pure data models) are no longer sufficient to meet the stringent requirements for communication stability in safe production. Physical-driven methods rely on prior channel modeling, which, while possessing good interpretability and stability, lacks generalization ability in unstructured interference scenarios. Data-driven methods, leveraging the powerful feature extraction capabilities of deep learning, can fit complex nonlinear relationships, but are prone to mode collapse under extreme conditions due to a lack of physical constraints. Hybrid-drive architectures, by combining the advantages of both, can theoretically balance robustness and flexibility, and have become the core direction of mine communication evolution. Among them, the dynamic fusion mechanism of physical and data models is the key link that determines the performance limit of the hybrid architecture. Existing fusion strategies generally adopt static weighting or discrete scene switching mechanisms, which essentially simplify the highly dynamic channel response problem into a mapping problem of a finite state space. When faced with second-level jumps in channel parameters (such as a 40% increase in millimeter-wave penetration loss due to the start-up and shutdown of a coal mining machine, or a jump in multipath delay spread to 3.2 μs due to turning of a mine car), these methods often experience a two-order-of-magnitude increase in bit error rate due to lagging weight updates or switching oscillations, revealing serious response lag and stability defects. Some solutions attempt to introduce adversarial training or Bayesian optimization to achieve adaptation, but due to excessive computational complexity (single inference delay > 50 ms), they cannot be deployed on intrinsically safe edge devices. Other solutions pre-set weight lookup tables, which reduce the online computational burden, but due to the explosion of combinations in underground scenarios, they fall into the dual dilemma of incomplete coverage and approximation errors, and lack a smooth transition mechanism, which easily causes system performance jitter and threatens the security of communication links.

[0003] Existing technologies, such as the invention patent application CN120223546A, disclose a communication network design and configuration method suitable for mine ventilation scenarios. This method optimizes the communication network topology by establishing a quantitative relationship model between network construction costs and security performance, and determines the optimal network node configuration scheme through target optimization, thus significantly improving network economy while ensuring system reliability. However, although this method has made progress at the mine network topology optimization level, it still uses the aforementioned rigid strategy in the fusion layer of the hybrid-driven architecture, failing to solve the dynamic balance problem between physical mechanisms and data learning. Its core flaw is that it cannot achieve millisecond-level weight decision-making through lightweight intelligent agents, nor does it construct a physically meaningful state perception and normalization mechanism, and it lacks stability constraints on the weight change rate. This leads to severe oscillations in the fusion output when the system is dealing with typical operating conditions such as low signal-to-noise ratio (<5dB), strong multipath (τ>2μs), and high-speed movement (v>3m / s), with the bit error rate fluctuating as high as 47%, far exceeding the safety production allowable threshold.

[0004] Based on this, the present invention proposes an adaptive channel estimation method based on deep learning to solve the problems existing in the prior art and improve the efficiency of downhole wireless communication. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a four-dimensional dynamic adaptive gating fusion method for mine communication, which solves the problem that traditional hybrid-driven architecture fusion strategies cannot adapt to the drastic, second-level changes in underground channel conditions and meet the real-time requirements of intrinsically safe underground equipment.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A four-dimensional dynamic adaptive gating fusion method for mine communication includes:

[0008] Step 1: Perform four-dimensional state perception to obtain the four-dimensional channel state vector s;

[0009] Step 2: Perform nonlinear normalization mapping on the four-dimensional channel state vector s;

[0010] Step 3: Input the normalized features from Step 2 into an ultralightweight neural network for weight decision-making;

[0011] Step 4: Apply stability constraints to the weights determined in Step 3 to achieve accurate, stable, and adaptive fusion of the physical model and data model outputs, and ensure output weight constraints and stability.

[0012] In a preferred embodiment of the present invention, the four-dimensional channel state vector s in step 1 is obtained and constructed in real time based on the gated fusion device, and specifically includes signal-to-noise ratio (SNR), multipath delay spread (τ), device moving speed (v), and modulation type.

[0013] In a preferred embodiment of the present invention, step 2 compresses the four-dimensional channel state vector s of the original observation to the vicinity of the [0, 1] interval, which better reflects its physical nature and its sensitivity to the impact on system performance, by performing a nonlinear normalization mapping on the four-dimensional channel state vector s.

[0014] In a preferred embodiment of the present invention, step 2, which involves performing nonlinear normalization mapping on the four-dimensional channel state vector s, includes:

[0015] Step 2.1: Perform logarithmic compression mapping on the signal-to-noise ratio (SNR):

[0016]

[0017] Where 40 is the upper limit of the dynamic range of SNR;

[0018] Step 2.2: Perform exponential saturation mapping on the multipath delay spread τ:

[0019] τ norm =1-e -0.8τ ;

[0020] Step 2.3: Perform hyperbolic constraint mapping on the device's moving speed v:

[0021] v norm =tanh(0.3v);

[0022] Step 2.4: The modulation type feature vector directly participates in the concatenation without additional normalization processing;

[0023] The three normalized values ​​[SNR] norm ,τ norm ,v norm The input vector of the decision network is concatenated with the modulation type's encoded vector.

[0024] In a preferred embodiment of the present invention, step 3, the process of making weight decisions in the ultra-lightweight neural network, includes:

[0025] Step 3.1: Construct a lightweight decision network;

[0026] Establish an MLP network structure, input the vector obtained in step 2 into the MLP, and seamlessly embed the MLP into intrinsically safe communication equipment or edge computing nodes in the mine. The resulting lightweight MLP output physical branch weights are:

[0027] w p =0.6·Sigmoid(W2ELU(W1s+b1)+b2)+0.2;

[0028] Among them, w p The value range w represents the weight of the physical branch. p ∈[0.2,0.8]; Sigmoid(*) is the activation function; ELU(*) is the exponential linear unit activation function; W1 and W2 represent weight matrices; b1 and b2 represent bias terms; s represents the input vector;

[0029] Step 3.2: Dynamic weight decision-making based on the adaptive strategy of dynamic scenarios according to electromagnetic laws.

[0030] In a preferred embodiment of the present invention, the MLP is a three-layer network structure, including:

[0031] Input layer: 6 nodes, receiving a 6-dimensional normalized feature vector x;

[0032] Hidden layer: 7 nodes, using the ELU activation function;

[0033] Output layer: 1 node, using the Sigmoid activation function, its output is initialized. The constraint is in the interval (0,1).

[0034] In a preferred embodiment of the present invention, during the dynamic weight decision-making process, in a multipath-dominated scenario:

[0035] When the channel delay spread exceeds 2μs, the channel curvature leads to a multipath number N. path ≥8, Nakagami shape parameter m≤0.95, at which point the physical branch weight w p It automatically increases to 0.7-0.8, accurately reconstructs multipath delay and attenuation through a differentiable geometric model, and dominates signal compensation. The data branch uses a deep learning-based phase compensation algorithm to assist in correcting nonlinear phase distortion.

[0036] In a preferred embodiment of the present invention, during the dynamic weight decision-making process, in a noise-dominated scenario:

[0037] In low signal-to-noise ratio environments, data-driven branch weights w d The value is increased to 0.7-0.8. By setting a specific window in the complex domain, the signal features within the window are weighted and processed, thereby focusing on the complex domain window attention of local features. Combined with the characteristic of the GELU activation function to smoothly suppress noise when processing noisy signals, the signal after the complex domain window attention processing is subjected to nonlinear transformation, thereby dominating signal recovery; the physical branch provides channel prior constraints.

[0038] In a preferred embodiment of the present invention, during the dynamic weight decision-making process, in a high-speed movement scenario:

[0039] When the equipment moves at a speed exceeding 3 m / s, the maximum Doppler frequency shift f d >24Hz triggers a dual compensation mechanism: the physical branch compensates for Doppler frequency offset, and the data branch suppresses time-varying interference. The weight of the physical branch is increased to 0.6-0.7 to strengthen Doppler frequency offset compensation, while the data branch shrinks the attention window to suppress time-varying interference.

[0040] When the signal uses higher-order modulation, the increased constellation density requires higher feature resolution, the data branch weight is increased to 0.5-0.7 and the window is expanded to 64, and the physical branch optimizes the number of multipath components to provide fine priors;

[0041] If high-order modulation and high-speed movement coexist, the gating network can further coordinate and adjust the weights to achieve joint optimization of Doppler compensation and noise suppression.

[0042] In a preferred embodiment of the present invention, step 4, which applies stability constraints to the weights determined in step 3, includes:

[0043] Step 4.1: Raw output of the MLP Apply two constraints:

[0044] Step 4.2: Range Constraint: Linearly scale and clamp the weight values ​​within the interval [0.2, 0.8];

[0045] Step 4.3: Rate of Change Constraint: On a continuous time series, constrain the weights w. p The rate of change of the increment is required to be:

[0046] ∣Δω p / Δt∣·0.1 / ms;

[0047] Where, Δw p Represents weight w p The increment is the change in weight between the previous and next time steps; Δt represents the time increment.

[0048] Step 4.4: Output the constrained physical branch fusion weight w p The weight of the data branch is w d =1-w p These two weights are sent to the fusion execution unit to complete the processing of physical feature F. phy With data features F data The weighted summation and fusion are used to obtain the fused output Y. fusion :

[0049] Y fusion =wp ·F phy +w d ·F data .

[0050] Compared with existing technologies, this invention provides a four-dimensional dynamic adaptive gating fusion method for mine communication, which has the following advantages:

[0051] The method of this invention can perceive the four-dimensional state of the channel in real time and complete the weight optimization calculation within 10 milliseconds with the help of a lightweight decision network. With its accurate multipath modeling capability, the network plays a leading role in signal compensation, significantly reducing the bit error rate in the scenario, fundamentally solving the lag problem of static fusion strategies, and ensuring the timeliness and accuracy of signal detection.

[0052] This invention employs a miniature multilayer perceptron, significantly reducing the total number of parameters and the latency of a single inference iteration, while minimizing computational overhead. This lightweight design offers significant advantages, allowing the intelligent decision-making core to be easily embedded in any edge computing node with almost no increase in the system's overall latency and power consumption. It effectively resolves the contradiction between complex algorithms and limited resources, perfectly meeting the resource constraints of mining equipment.

[0053] This invention introduces a multiple constraint mechanism, on the one hand, to increase the output weight w p Strictly limiting the weights to a reasonable range of [0.2, 0.8] prevents system crashes due to the complete failure of a branch; on the other hand, limiting the rate of change of weights to no more than 0.1 / ms ensures smooth weight switching and avoids signal jitter or decision oscillation caused by weight jumps.

[0054] This invention transforms the original channel state parameters into more physically meaningful and gradient-sensitive features through nonlinear normalization mapping, which are then input into a lightweight MLP for decision-making. This mapping method enhances the system's ability to perceive critical operating conditions, ensures that the dynamic weight adjustment strategy strictly follows the laws of electromagnetic wave propagation, and achieves a dynamic optimal balance between the constraints of physical mechanisms and the flexibility of data learning. It solves the problems of traditional hybrid drive architecture fusion strategies being unable to adapt to the drastic, second-level changes in downhole channel states and failing to meet the real-time requirements of intrinsically safe downhole equipment. Attached Figure Description

[0055] Figure 1 This is a flowchart of the four-dimensional dynamic adaptive gating fusion method for mine communication according to the present invention;

[0056] Figure 2 Here are the BER diagrams for each method in the three scenarios of this invention;

[0057] Figure 3This is a diagram illustrating the evolution of physical branch weights and bit error rate when a mine car enters a turning area from a steady-state roadway according to the present invention.

[0058] Figure 4 This is a graph showing the impact of the dynamic gating mechanism of this invention on the bit error rate.

[0059] Among them, Figure 2 In the figure, Figure (a) shows the BER map of each method in the turning area, Figure (b) shows the BER map of each method in the equipment interference area, and Figure (c) shows the BER map of each method in the high-speed movement area. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Please see Figure 1 This invention provides a four-dimensional dynamic adaptive gating fusion method for mine communication. The essence of this method lies in a lightweight intelligent decision-making system at its core, executing the fusion process: first, the channel state is subjected to four-dimensional perception and physically meaningful nonlinear normalization; then, the normalized features are input into an ultra-lightweight neural network for weight decision-making; finally, stability constraints are applied to the decided weights, thereby achieving accurate, stable, and adaptive fusion of the physical model and the data model outputs. The execution body of the method is a gating fusion device integrated into the signal processing unit, which can continuously receive input from environmental sensors and a signal preprocessing module. The specific fusion process of the four-dimensional dynamic adaptive gating fusion method for mine communication includes the following steps:

[0062] Step 1: Use a gated fusion device to perform four-dimensional state perception and obtain the four-dimensional channel state vector s;

[0063] A four-dimensional channel state vector s is acquired and constructed in real time using a gated fusion device;

[0064] The four dimensions of this vector are:

[0065] (1) Signal-to-noise ratio (SNR): The signal-to-noise ratio value in the dynamic range of 0-40dB is obtained in real time through the signal quality monitoring module, with a focus on monitoring low signal-to-noise ratio scenarios caused by sudden changes in gas concentration;

[0066] (2) Multipath delay spread (τ): The multipath intensity distribution within 0-5μs is measured using a time-domain channel estimator, with particular attention to the delay jump caused by the turn of the roadway;

[0067] (3) Equipment moving speed (v): The speed change from 0 to 10 m / s is tracked by the Doppler frequency shift estimation module and correlated with the Doppler frequency shift effect;

[0068] (4) Modulation type: BPSK, QPSK and 16QAM are mapped to three-dimensional feature vectors, and high-order modulation triggers fine processing mode.

[0069] In one specific implementation, when 16QAM modulation is detected (feature vector [0, 0, 1]), the system automatically expands the attention window to 64 points and increases the number of physical branch paths to meet the fine feature analysis requirements of high-order modulation.

[0070] The system achieves intelligent decision-making by analyzing the four-dimensional channel state in real time:

[0071]

[0072] Step 2: Perform nonlinear normalization mapping on the four-dimensional channel state vector s;

[0073] To avoid excessive differences in the dimensions and numerical ranges of each dimension affecting the convergence and performance of the decision network, the acquired four-dimensional channel state vector s needs to be subjected to a physically meaningful nonlinear normalization mapping, compressing it to the vicinity of the [0, 1] interval. The purpose of this step is not simply numerical scaling, but to map the original observations to a feature space that better reflects their physical nature and sensitivity to the system performance, providing inputs with clear physical meaning and excellent gradient characteristics for subsequent intelligent decision-making.

[0074] Step 2.1: The specific mapping function is designed as follows:

[0075] Perform logarithmic compression mapping on the signal-to-noise ratio (SNR):

[0076]

[0077] Here, 40 is the upper limit of the dynamic range of SNR.

[0078] The effect of signal-to-noise ratio (SNR) on the system's bit error rate is not linear. In the low SNR region (0-10 dB), a 1 dB increase significantly improves performance; however, the improvement gradually saturates in the high SNR region. This logarithmic function perfectly matches this physical law, greatly enhancing the gradient sensitivity in the critical low SNR operating range. This allows the decision network to more accurately detect signal degradation caused by sudden changes in gas concentration, thus enabling it to react earlier.

[0079] Step 2.2: Perform exponential saturation mapping on the multipath delay spread τ:

[0080] τ norm =1-e -0.8τ ;

[0081] The impact of multipath delay spread on the signal also exhibits a saturation effect. When τ is small (0-2μs), its increase will lead to a sharp increase in inter-symbol crosstalk (ISI), which is the main factor affecting the bit error rate; when τ exceeds a certain value (e.g., >3μs), its marginal negative impact on performance will weaken. This exponential function focuses the mapping on the critical delay interval of 0-2μs, which has the most significant impact on the bit error rate, allowing the decision network to prioritize multipath degradation caused by factors such as cornering.

[0082] Step 2.3: Perform hyperbolic constraint mapping on the device's moving speed v:

[0083] v norm =tanh(0.3v);

[0084] The Doppler effect caused by velocity intensifies with increasing velocity, but a smooth upper limit constraint is required. The hyperbolic tangent function (Tanh) smoothly constrains the velocity in the interval [0, 1) and provides a smooth inflection point near v = 3-4 m / s. This inflection point corresponds to the threshold at which the Doppler frequency shift begins to significantly affect the performance of coherent demodulation, thereby triggering the decision network to strengthen the Doppler compensation strategy.

[0085] After mapping the device's moving speed, the modulation type feature vector is processed:

[0086] Step 2.4: The modulation type feature vector directly participates in the concatenation without additional normalization processing;

[0087] The three normalized values ​​[SNR] norm ,τ norm ,v norm The input vector of the decision network is concatenated with the modulation type's encoded vector.

[0088] The innovation of this step lies in the fact that it is not a simple linear scaling, but a nonlinear function designed by introducing prior physical knowledge, which enhances the gradient sensitivity of key operating conditions (such as low SNR and high latency), laying a solid foundation for subsequent lightweight decision-making.

[0089] Step 3: Input the normalized features from Step 2 into an ultralightweight neural network for weight decision-making;

[0090] Step 3.1: Construct a lightweight decision network;

[0091] To achieve millisecond-level response and embedded deployment, this embodiment designs a miniature multilayer perceptron (MLP) with extremely compressed total parameters but complete functionality as the gated fusion core. The vector obtained in step 2 is input into an extremely lightweight multilayer perceptron (MLP). This MLP network structure is extremely lightweight, employing a three-layer MLP network structure:

[0092] Input layer: 6 nodes, receiving the above 6-dimensional normalized feature vector x.

[0093] Hidden layers: 7 nodes, using the ELU (Exponential LinearUnit) activation function. ELU was chosen instead of ReLU because its smooth output characteristics in the negative range can alleviate the gradient vanishing problem and improve the training stability and expressive power of small network models.

[0094] Output layer: 1 node, using the Sigmoid activation function, its output is initialized. The constraint is in the (0,1) interval, which provides a basis for subsequent stability constraints.

[0095] Parameter quantity and control: The number of parameters in this network is strictly controlled at an extremely low level.

[0096] From input layer to hidden layer: 6 × 7 = 42 weights + 7 biases

[0097] From hidden layer to output layer: 7 × 1 = 7 weights + 1 bias.

[0098] The total number of parameters is 42 + 7 + 7 + 1 = 57 parameters (approximately 0.06KB), which meets the lightweight design goal.

[0099] Computational efficiency advantage: This lightweight network structure brings computational efficiency advantages.

[0100] The latency for a single forward inference is extremely low, at only 12 microseconds; the computational overhead is minimal, requiring only about 0.002 GFLOPs of computation per inference.

[0101] This lightweight design ensures that the intelligent gating core can be seamlessly embedded in intrinsically safe communication equipment or edge computing nodes in mines, achieving real-time intelligent integration strategies with almost no increase in total system latency and power consumption.

[0102] This lightweight network outputs an initial weight value. The output physics branch weights of a lightweight MLP are:

[0103] w p =0.6·Sigmoid(W2ELU(W1s+b1)+b2)+0.2;

[0104] Among them, w p Represents the physical branch weight, used to measure the proportion of the physical branch in the fusion strategy, and its value range is constrained by w. p ∈[0.2,0.8], incremental change rate ≤0.1 / ms. The Sigmoid(*) function is a commonly used activation function, used here to map the input to the (0,1) interval, performing a nonlinear transformation on the intermediate results of the lightweight network output to control the range and variation characteristics of the weights. ELU(*) is the exponential linear unit activation function, which, compared to some traditional activation functions, can alleviate the gradient vanishing problem to some extent, providing the network with more effective nonlinear expressive power. W1 and W2 represent weight matrices, used in the lightweight MLP operation to perform linear transformations on the input vector s and the intermediate layer outputs. b1 and b2 represent bias terms, used to offset the linear transformation, providing the network with more expressive power.

[0105] Step 3.2: Dynamic weight decision-making based on adaptive strategies for dynamic scenarios using electromagnetic laws;

[0106] Dynamic weighted decision-making strictly follows the laws of electromagnetic wave propagation. It utilizes the synergistic effect of four state variables—signal-to-noise ratio, multipath delay spread, device speed, and modulation type—to achieve an adaptive balance between physical and data-driven approaches through neural network learning. Its core mechanism can be analyzed through three typical scenarios: multipath dominance, noise dominance, and high-speed movement.

[0107] (1) Multipath-dominated scenarios (τ>2μs);

[0108] When the channel delay spread exceeds 2μs, the channel curvature leads to a multipath number N. path ≥8, Nakagami shape parameter m≤0.95. Since the smaller the value of m in the Nakagami distribution, the more severe the signal fading, the signal exhibits significant heavy-tailed fading characteristics. At this point, the physical branch weight w... p It automatically increases to 0.7-0.8, accurately reconstructs multipath delay and attenuation through a differentiable geometric model, dominates signal compensation, and the data branch assists in correcting nonlinear phase distortion.

[0109] (2) Noise-dominated scenarios (SNR < 5dB);

[0110] In low signal-to-noise ratio environments, data-driven branch weights w d The value is increased to 0.7-0.8. The attention of local features is focused through the complex domain window, and the noise suppression characteristics of GELU activation dominate the signal recovery. The physical branch provides channel prior constraints to avoid misjudgment of the data model due to noise interference.

[0111] (3) High-speed movement scenario (v>3m / s);

[0112] When the equipment moves at a speed exceeding 3 m / s, the maximum Doppler frequency shift f d The 24Hz trigger employs a dual compensation mechanism: the physical branch compensates for Doppler frequency offset, while the data branch suppresses time-varying interference. The physical branch weight is increased to 0.6-0.7 to enhance Doppler frequency offset compensation, while the data branch shrinks its attention window to suppress time-varying interference. When the signal uses higher-order modulation (e.g., 16QAM), the increased constellation density requires higher feature resolution. The data branch weight is increased to 0.5–0.7, and the attention window is expanded to 64, while the physical branch optimizes the number of multipath components to provide fine priors. If higher-order modulation and high-speed motion coexist, the gating network further coordinates the weight adjustments to achieve joint optimization of Doppler compensation and noise suppression.

[0113] Step 4: Apply stability constraints to the weights determined in Step 3 to achieve accurate, stable, and adaptive fusion of the physical model and data model outputs, and ensure output weight constraints and stability.

[0114] Step 4.1: Raw output of the miniature multilayer perceptron (MLP) Two constraints are applied to ensure system stability:

[0115] Step 4.2: Range Constraint: The weight values ​​are linearly scaled and clamped within the interval [0.2, 0.8]. This constraint ensures that, under any circumstances, the physical branch and the data branch will not be completely abandoned, preventing model crashes due to the temporary failure of a single branch.

[0116] Step 4.3: Rate of Change Constraint: On a continuous time series, constrain the weights w. p The rate of change of the increment is required

[0117] ∣Δω p / Δt∣·0.1 / ms;

[0118] This formula represents the weight w. p The absolute value of the rate of change of the increment does not exceed 0.1 per millisecond. Wherein, Δw p Represents weight w p The increment is the change in weight between two consecutive time steps; Δt represents the time increment, which is the time interval between two consecutive time steps. This constraint is implemented through a first-order low-pass filter or smooth interpolation algorithm to ensure smooth weight switching without jumps and avoid introducing signal jitter. This constraint is guaranteed by the differentiability of the weight generation function: the derivative of the Siqmoid output layer of the gated MLP, combined with the Lipschitz constant of the state normalization function (e.g., ...). The system's weight change rate under extreme state transitions (ΔSNR>20dB) is suppressed to below 0.08 / ms.

[0119] Step 4.4: Output the constrained physical branch fusion weight w p The weight of the data branch is then determined as w. d =1-w p These two weights are sent to the fusion execution unit to complete the processing of physical feature F. phy With data features F data The weighted summation and fusion are used to obtain the fused output Y. fusion :

[0120] Y fusion =w p ·F phy +w d ·F data .

[0121] Example 1:

[0122] Please see Figures 2-4 To verify the feasibility of the four-dimensional dynamic adaptive gating fusion method for mine communication described above, this embodiment conducts a simulation experiment on the method.

[0123] 1. Comprehensive adaptability testing was conducted for complex mining scenarios, covering three typical scenarios: turning areas (strong multipath), equipment interference areas (low SNR), and high-speed movement (Doppler effect). The performance of various signal processing methods was evaluated in depth. In these three scenarios, the bit error rate (BER) of each method was tested according to a unified testing procedure. The test results are as follows: Figure 2 As shown.

[0124] Experimental results show that multipath effects cause severe signal distortion in the turning region. Traditional physical-driven methods (such as GSM) are limited by prior models, resulting in high bit error rates and difficulty in effectively handling complex multipath situations. Purely data-driven GAN detectors, lacking physical constraints, exhibit 'mode collapse' (i.e., the detector cannot stably output reasonable results) under multipath interference, with drastic fluctuations. The dual-drive method proposed in this invention accurately models the multipath channel (w) through physical branches. p =0.8), combined with data branch adaptive learning signal features, it exhibits excellent multipath robustness.

[0125] Strong noise in the equipment interference zone poses a significant challenge to signal detection. Physically driven methods suffer from limited noise suppression capabilities, leading to a sharp increase in bit error rate (BER). Among data-driven methods, CNN-LSTM, with its data learning advantage, performs slightly better than physically driven methods, but still falls short of the low BER requirements of mine communication. The GAN detector again succumbs to noise interference, exhibiting the worst performance. In contrast, the dual-driven method of this invention enhances the data branch weights (w) through a window attention mechanism. d=0.72), effectively suppressing noise and reducing BER.

[0126] In high-speed mobile scenarios, the rapid movement of devices triggers a significant Doppler effect, exacerbating the time-varying characteristics of signals. Traditional methods struggle to track signal changes in real time, and GSM's high bit error rate (BER) makes it unsuitable for the frequency offset challenges brought about by high-speed movement. While data-driven methods based on deep learning offer some improvement, they still have limitations. This invention employs a dual-drive approach, combining physics and data, utilizing a physical model to compensate for the Doppler frequency shift. Simultaneously, it incorporates data branch learning of time-varying signal characteristics to reduce the BER, achieving efficient adaptation to high-speed mobile scenarios.

[0127] 2. To verify the stability of the dynamic gating mechanism during scene switching, this experiment simulates the dynamic process of a mine car entering a turning area from a steady-state roadway (0-20ms), and monitors the physical branch weights (w) in real time. p The evolution of bit error rate (BER) and incremental constraint strategies are analyzed to improve system stability.

[0128] Experimental conditions: time delay spread τ = 3.2 μs in the turning region, Nakagami-m fading parameter m = 0.95, modulation scheme 16QAM, initial state steady-state tunnel (w p =0.5, w d =0.5).

[0129] Compared with the traditional method (without incremental constraints): the traditional method's weights abruptly increased from 0.5 to 0.8 within 3ms (a change rate of 0.1 / ms, violating stability constraints), then oscillated to 0.6 due to overshoot before finally stabilizing at 0.8; this drastic weight jump resulted in a BER oscillation of 47%, far exceeding the 3% fluctuation range of this method. This result verifies the reliability of the dynamic gating mechanism in complex scenario switching, providing stability assurance for real-time communication of mining mobile equipment.

[0130] 3. Disable dynamic gating;

[0131] The dynamic gating mechanism in the model was turned off to explore its impact on network performance. The dynamic gating mechanism aims to adaptively adjust the fusion weights of different modalities, enhance the transmission of key information, and suppress noise and redundant information.

[0132] Experimental results are as follows Figure 4 As shown, disabling dynamic gating significantly increases the bit error rate. This indicates that dynamic gating is crucial for improving detection accuracy, effectively filtering and fusing multimodal information, and reducing the bit error rate.

[0133] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A four-dimensional dynamic adaptive gating fusion method for mine communication, characterized in that, include: Step 1: Perform four-dimensional state perception to obtain the four-dimensional channel state vector s; Step 2: Perform nonlinear normalization mapping on the four-dimensional channel state vector s; Step 3: Input the normalized features from Step 2 into an ultralightweight neural network for weight decision-making; Step 4: Apply stability constraints to the weights determined in Step 3 to achieve accurate, stable, and adaptive fusion of the physical model and data model outputs, and ensure output weight constraints and stability.

2. The four-dimensional dynamic adaptive gating fusion method for mine communication as described in claim 1, characterized in that, The four-dimensional channel state vector s in step 1 is obtained and constructed in real time based on the gated fusion device, and specifically includes the signal-to-noise ratio (SNR), multipath delay spread (τ), device moving speed (v), and modulation type.

3. The four-dimensional dynamic adaptive gating fusion method for mine communication as described in claim 1, characterized in that, Step 2 involves performing a nonlinear normalization mapping on the four-dimensional channel state vector s, compressing the original observation's four-dimensional channel state vector s to the vicinity of the [0, 1] interval, which better reflects its physical nature and its sensitivity to the system's performance.

4. The four-dimensional dynamic adaptive gating fusion method for mine communication as described in claim 3, characterized in that, Step 2, the process of performing nonlinear normalization mapping on the four-dimensional channel state vector s, includes: Step 2.1: Perform logarithmic compression mapping on the signal-to-noise ratio (SNR): Where 40 is the upper limit of the dynamic range of SNR; Step 2.2: Perform exponential saturation mapping on the multipath delay spread τ: t norm =1-e -0.8τ ; Step 2.3: Perform hyperbolic constraint mapping on the device's moving speed v: v norm =tanh(0.3v); Step 2.4: The modulation type feature vector directly participates in the concatenation without additional normalization processing; The three normalized values ​​[SNR] norm ,τ norm ,v norm The input vector of the decision network is concatenated with the modulation type's encoded vector.

5. The four-dimensional dynamic adaptive gating fusion method for mine communication as described in claim 1, characterized in that, Step 3, the process of making weight decisions in the ultralightweight neural network, includes: Step 3.1: Construct a lightweight decision network; Establish an MLP network structure, input the vector obtained in step 2 into the MLP, and seamlessly embed the MLP into intrinsically safe communication equipment or edge computing nodes in the mine. The resulting lightweight MLP output physical branch weights are: w p =0.6·Sigmoid(W2ELU(W1s+b1)+b2)+0.2; Among them, w p The value range w represents the weight of the physical branch. p ∈[0.2,0.8]; Sigmoid(*) is the activation function; ELU(*) is the exponential linear unit activation function; W1 and W2 represent weight matrices; b1 and b2 represent bias terms; s represents the input vector; Step 3.2: Dynamic weight decision-making based on the adaptive strategy of dynamic scenarios according to electromagnetic laws.

6. The four-dimensional dynamic adaptive gating fusion method for mine communication as described in claim 5, characterized in that, MLP is a three-layer network structure, including: Input layer: 6 nodes, receiving a 6-dimensional normalized feature vector x; Hidden layer: 7 nodes, using the ELU activation function; Output layer: 1 node, using the Sigmoid activation function, its output is initialized. The constraint is in the interval (0,1).

7. The four-dimensional dynamic adaptive gating fusion method for mine communication as described in claim 5, characterized in that, In the process of making dynamic weight decisions, in multipath-dominated scenarios: When the channel delay spread exceeds 2μs, the channel curvature leads to a multipath number N. path ≥8, Nakagami shape parameter m≤0.95, at which point the physical branch weight w p It automatically increases to 0.7-0.8, accurately reconstructs multipath delay and attenuation through a differentiable geometric model, and dominates signal compensation. The data branch uses a deep learning-based phase compensation algorithm to assist in correcting nonlinear phase distortion.

8. The four-dimensional dynamic adaptive gating fusion method for mine communication as described in claim 5, characterized in that, In the process of making dynamic weight decisions, under noise-dominated scenarios: In low signal-to-noise ratio environments, data-driven branch weights w d The value is increased to 0.7-0.

8. By setting a specific window in the complex domain, the signal features within the window are weighted and processed, thereby focusing on the complex domain window attention of local features. Combined with the characteristic of the GELU activation function to smoothly suppress noise when processing noisy signals, the signal after the complex domain window attention processing is subjected to nonlinear transformation, thereby dominating signal recovery; the physical branch provides channel prior constraints.

9. The four-dimensional dynamic adaptive gating fusion method for mine communication as described in claim 5, characterized in that, In the process of making dynamic weight decisions, in high-speed mobile scenarios: When the equipment moves at a speed exceeding 3 m / s, the maximum Doppler frequency shift f d >24Hz triggers a dual compensation mechanism: the physical branch compensates for Doppler frequency offset, and the data branch suppresses time-varying interference. The weight of the physical branch is increased to 0.6-0.7 to strengthen Doppler frequency offset compensation, while the data branch shrinks the attention window to suppress time-varying interference. When the signal uses higher-order modulation, the increased constellation density requires higher feature resolution, the data branch weight is increased to 0.5-0.7 and the window is expanded to 64, and the physical branch optimizes the number of multipath components to provide fine priors; If high-order modulation and high-speed movement coexist, the gating network can further coordinate and adjust the weights to achieve joint optimization of Doppler compensation and noise suppression.

10. The four-dimensional dynamic adaptive gating fusion method for mine communication as described in claim 1, characterized in that, Step 4 involves applying stability constraints to the weights determined in Step 3, including: Step 4.1: Raw output of the MLP Apply two constraints: Step 4.2: Range Constraint: Linearly scale and clamp the weight values ​​within the interval [0.2, 0.8]; Step 4.3: Rate of Change Constraint: On a continuous time series, constrain the weights w. p The rate of change of the increment is required to be: ∣Δω p / Δt∣·0.1 / ms; Where, Δw p Represents weight w p The increment; Δt represents the time increment; Step 4.4: Output the constrained physical branch fusion weight w p The weight of the data branch is w d =1-w p These two weights are sent to the fusion execution unit to complete the processing of physical feature F. phy With data features F data The weighted summation and fusion are used to obtain the fused output Y. fusion : Y fusion =w p ·F phy +w d ·F data 。

Images