Transient electromagnetic signal transmitting and receiving device

Abstract

The invention provides a transient electromagnetic signal transmitting and receiving device, and relates to the technical field of geophysical exploration, and the device comprises a signal transmitting coil which is disposed on the bottom surface of a frame and is used for transmitting a transient electromagnetic signal to a to-be-detected geologic body; the plurality of signal receiving coils are respectively arranged on the bottom surface and each side surface of the frame, and the signal receiving coils are used for receiving electromagnetic response signals fed back by the geologic body to be detected; and based on an improved GWO-LSTM algorithm, determining structure parameters of the signal transmitting coil and the signal receiving coil according to detection requirements. According to the method, optimal parameters are accurately matched through demand driving, the detection precision and the signal quality are remarkably improved, a physical mechanism and an AI algorithm are deeply fused, the model generalization ability is enhanced, dependence on expert experience is greatly reduced, the detection efficiency is improved, a complete intelligent closed loop is formed, a core technical support is provided for digital upgrading of equipment, and the method is worthy of popularization and application. The industrial problems of parameter solidification and performance limitation of a traditional method are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geophysical prospecting, in particular to a transient electromagnetic signal transmitting and receiving device. BACKGROUND

[0002] The transient electromagnetic method is an important method for resource exploration and engineering geological prospecting, and its effectiveness depends on the performance of the transmitting and receiving coils. The coil structure parameters (such as shape, size, and number of turns) of traditional devices are fixedly designed and determined by engineers' experience, which has significant limitations: Firstly, general fixed parameters cannot achieve optimal performance when facing different detection tasks, for example, there is an inherent contradiction between the requirements of shallow fine structure exploration and deep resource exploration for coils; Secondly, the coupling relationship between electromagnetic fields and complex geological structures is highly nonlinear, and relying on artificial experience or traditional numerical simulation for parameter optimization has high calculation cost and is difficult to find a global optimal solution; Finally, the existing technology lacks the ability to dynamically and adaptively adjust the coil parameters according to real-time detection needs, which limits the detection accuracy, efficiency, and adaptability to complex environments.

[0003] In addition, although some research attempts to use intelligent algorithms for optimization design, there are generally problems such as slow convergence speed, easy to fall into local optimum, and poor combination of algorithm search and physical model, which makes it difficult to meet the high-precision and high-efficiency detection needs in practical applications. Therefore, it is necessary to design a transient electromagnetic signal transmitting and receiving device. SUMMARY

[0004] In order to overcome the deficiencies of the prior art, the purpose of the present application is to provide a transient electromagnetic signal transmitting and receiving device.

[0005] To achieve the above-mentioned purpose, the present application provides the following solutions: The present application provides a transient electromagnetic signal transmitting and receiving device, comprising: a frame, the plurality of sides of which are oblique to the bottom surface thereof; a signal transmitting coil, arranged on the bottom surface of the frame, for transmitting a transient electromagnetic signal to a geological body to be detected; a plurality of signal receiving coils, respectively arranged on the bottom surface and each side of the frame, for receiving an electromagnetic response signal fed back by the geological body to be detected; The structure parameters of the signal transmitting coil and the signal receiving coils are determined according to the detection requirements based on an improved GWO-LSTM algorithm.

[0006] Preferably, the shape of the frame is a regular pyramid or a regular prism, the signal transmitting coil is concentrically arranged on the bottom surface of the frame, and the edges of the signal transmitting coil and the plurality of signal receiving coils are respectively arranged on the edge frames of the corresponding frame surfaces.

[0007] Preferably, the device further includes: a signal source and a plurality of storage units, wherein the signal source is connected to the signal transmitting coil and is used to transmit current to the signal transmitting coil to drive the signal transmitting coil to transmit the transient electromagnetic signal, and the storage units correspond one-to-one with the signal receiving coils, and the storage units are used to store the electromagnetic response signals received by the corresponding signal receiving coils.

[0008] Preferably, the structural parameters of the signal transmitting coil and the signal receiving coil are determined based on the detection requirements using the improved GWO-LSTM algorithm, specifically as follows: A structural parameter determination model is constructed based on the improved GWO-LSTM algorithm; Train the model to determine the structural parameters; The detection requirements are input into the structural parameters to determine the model, and the structural parameters are obtained. The signal transmitting coil and signal receiving coil of the transient electromagnetic signal transceiver are adjusted based on structural parameters.

[0009] Preferably, a structural parameter determination model is constructed based on the improved GWO-LSTM algorithm, specifically as follows: Improvements were made to the GWO algorithm and LSTM network structure; The improved GWO algorithm is used as the optimizer, and the improved LSTM network structure is used as the optimized object to obtain the hyperparameter combination. A structural parameter determination model is constructed by initializing the parameters of the LSTM network structure based on hyperparameter combination.

[0010] Preferably, the GWO algorithm is improved as follows: Based on the traditional GWO algorithm, Tent mapping is introduced as a random number generation mechanism; Based on the traditional GWO algorithm, a nonlinear convergence factor is introduced to replace the convergence factor; Based on the traditional GWO algorithm, an inertial weight is introduced to adjust the balance between global optimization and local optimization.

[0011] Preferably, the LSTM network structure is improved, specifically as follows: A memory function unit is added to each hidden unit of the recursive hidden layer neuron in the traditional LSTM network structure.

[0012] Preferably, the structural parameter determination model is trained, specifically as follows: Construct a training dataset, which consists of detection requirements and coil parameters; The parameter-determining model is fully trained based on the training dataset.

[0013] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention provides a transient electromagnetic signal transceiver, including a signal transmitting coil disposed on the bottom surface of a frame for transmitting transient electromagnetic signals to a geological body to be detected; and multiple signal receiving coils disposed on the bottom surface and each side surface of the frame, respectively, for receiving electromagnetic response signals fed back by the geological body to be detected; the structural parameters of the signal transmitting and receiving coils are determined according to the detection requirements based on an improved GWO-LSTM algorithm. This invention, by introducing an improved GWO-LSTM hybrid intelligent algorithm, achieves adaptive determination of the structural parameters of the transient electromagnetic coils, bringing multiple beneficial effects: It has achieved a leap from "fixed design" to "demand-driven". The device can output the optimal coil parameters in real time according to specific requirements such as detection depth, resolution, and environmental noise, which fundamentally improves the quality and relevance of detection data. Second, by improving the LSTM network structure (introducing memory functional units), the model has the ability to capture the long-term dependence of electromagnetic transient processes, and its parameter prediction results are more physically consistent and have stronger generalization ability. Third, the improved GWO algorithm (which integrates Tent chaotic mapping, nonlinear convergence factor and PSO mechanism) is used for hyperparameter optimization, which greatly enhances the global search capability and convergence speed, ensuring that the constructed LSTM prediction model itself is optimal or close to optimal, thereby guaranteeing the reliability of the final parameter decision. Fourth, a complete intelligent closed loop of "demand-simulation-optimization-deployment" has been formed, which greatly reduces the dependence on the experience of operators, improves detection efficiency and automation level, and provides core technical support for the intelligent upgrading of transient electromagnetic detection equipment. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram illustrating the process for determining the structural parameters of a transient electromagnetic signal transceiver device provided in an embodiment of the present invention. Figure 2 A schematic diagram illustrating the population generation effect of initializing the Tent chaotic map; Figure 3 This is a schematic diagram comparing convergence factors; Figure 4A schematic diagram of the improved LSTM network structure. Detailed Implementation

[0016] 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.

[0017] The purpose of this invention is to provide a transient electromagnetic signal transceiver that upgrades the traditional fixed coil to an intelligent adjustable mode. By demand-driven precise matching of optimal parameters, it significantly improves detection accuracy and signal quality. It deeply integrates physical mechanisms and AI algorithms to enhance model generalization capabilities. It greatly reduces reliance on expert experience and improves detection efficiency. It forms a complete intelligent closed loop, providing core technical support for the digital upgrade of equipment and effectively solving the industry problems of fixed parameters and limited performance in traditional methods.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] This invention provides a transient electromagnetic signal transceiver, comprising: a frame, wherein multiple sides of the frame are obliquely intersecting its bottom surface; A signal transmitting coil is disposed on the bottom surface of the frame and is used to transmit transient electromagnetic signals to the geological body to be detected; Multiple signal receiving coils are respectively disposed on the bottom surface and each side surface of the frame. The signal receiving coils are used to receive the electromagnetic response signals fed back by the geological body to be detected. Based on the improved GWO-LSTM algorithm, the structural parameters of the signal transmitting coil and the signal receiving coil are determined according to the detection requirements.

[0020] The frame is shaped like a regular pyramid or a regular frustum. The signal transmitting coil is concentrically arranged with the bottom surface of the frame. The edges of the signal transmitting coil and multiple signal receiving coils are respectively arranged on the border of the corresponding frame surface.

[0021] The device further includes: a signal source and multiple storage units. The signal source is connected to the signal transmitting coil and is used to transmit current to the signal transmitting coil to drive the signal transmitting coil to transmit the transient electromagnetic signal. The storage units correspond one-to-one with the signal receiving coils and are used to store the electromagnetic response signals received by the corresponding signal receiving coils.

[0022] The device itself will not be described in detail here. The key point of this invention lies in determining the structural parameters of the signal transmitting coil and the signal receiving coil, which will be described in detail below: like Figure 1 As shown, the structural parameters of the signal transmitting coil and the signal receiving coil are determined based on the detection requirements using the improved GWO-LSTM algorithm, specifically: Step 1: Construct a structural parameter determination model based on the improved GWO-LSTM algorithm; Step 2: Train the model to determine the structural parameters; Step 3: Input the detection requirements into the structural parameters to determine the model and obtain the structural parameters; Step 4: Adjust the signal transmitting coil and signal receiving coil of the transient electromagnetic signal transceiver based on the structural parameters.

[0023] In step 1, a structural parameter determination model is constructed based on the improved GWO-LSTM algorithm, specifically as follows: Step 101: Improve the GWO algorithm and LSTM network structure, specifically as follows: First, the specific improvements to the GWO algorithm will be introduced: Since the GWO algorithm is existing technology, it will not be introduced here. Only its improvements will be discussed, which include three improvements: 1. Population initialization based on the Tent chaotic mapping method Tent mapping, through a simple mathematical iterative formula, can generate seemingly random but actually uniformly distributed sequences within a given interval. This characteristic allows the initial population generated based on Tent mapping to cover a wider solution space, effectively avoiding the trap of getting stuck in local optima in the early stages of the algorithm. Furthermore, the pseudo-randomness of Tent mapping also means that the generated sequences have a certain degree of unpredictability, which helps the algorithm maintain sufficient exploration capability during the search process, thereby further improving the global search efficiency and convergence accuracy. Therefore, introducing Tent mapping into intelligent optimization algorithms as a random number generation mechanism is an effective means to improve algorithm performance and enhance population diversity. The Tent formula used is as follows: (1) Where a = 0.5, xn(i, j) represents the information of the i-th object in the j-th dimension during the n-th iteration. The Tent chaotic mapping initialization population generation effect is as follows: Figure 2 As shown.

[0024] 2. Nonlinear convergence factor Although the Grey Wolf optimization algorithm performs well, it still suffers from a certain degree of insufficient local search capability. Specifically, global search capability is crucial for the algorithm to find the target, while local search capability determines the convergence speed. Therefore, the GWO algorithm needs to balance global and local search capabilities to achieve better optimization results. The size of the convergence factor 'a' determines the algorithm's convergence capability. When 'a' is greater than 1, the Grey Wolf will attempt to find a more advantageous position to avoid prey; when 'a' is less than 1, it will launch an attack in the hope of capturing prey. Increasing the time for the Grey Wolf to converge to the target based on the characteristics of 'a' enhances the algorithm's local search capability and improves convergence capability, as shown in the following formula: (2) In the formula, a1 and a2 represent the starting and ending points of the convergence factor a, respectively; h is a factor that adjusts the local search capability, setting a1 to 1, a2 to 0, and h to 0.25; t is the current iteration number; and tmax is the maximum iteration number. The changes in a can be seen from... Figure 3 Observations show that the improved 'a' decreases non-linearly and becomes 0 after half of the iterations. At this point, the gray wolves begin to attack their prey and perform a local search. In simple maps, the wolf pack can easily perform a global search, so a local search needs to be performed earlier. In complex maps, the value of 'h' can be adjusted appropriately to extend the timing of the local search or increase the number of iterations. By combining the overall performance of the algorithm with the local search properties, we can prevent them from becoming overly focused on a specific target, thereby achieving the goal of accurate search.

[0025] 3. Improved Individual Location Update Strategy The Grey Wolf Optimization (GWO) algorithm, inspired by the hunting behavior of grey wolves in nature, has demonstrated good performance in solving complex optimization problems. However, the traditional GWO algorithm mainly relies on simulating the leader wolf to guide the search process. This mechanism is relatively simplistic in controlling the balance between global and local searches, potentially limiting the algorithm's flexibility in exploring a broader solution space and finely searching for the optimal solution. To overcome this limitation, an inertial weight ω is introduced to adjust the balance between global and local optimization in the GWO model, achieving the best results. The inertial weight ω, as a dynamically adjustable parameter, adaptively adjusts the wolf pack's transition between global and local optimization based on the search needs at different stages. Specifically, in the initial exploration phase, a larger ω value encourages the wolf pack to traverse the solution space more extensively, enhancing global search capabilities. As the search progresses, gradually decreasing the ω value helps the wolf pack focus on potential optimal regions, improving local search accuracy. In this way, the optimal balance between global and local searches is achieved in the GWO algorithm, thereby improving the overall search capability and convergence speed, ultimately achieving the best results in solving complex optimization problems, as shown below: (3) (4) (5) This invention uses a proportional weighting method based on step-size Euclidean distance to measure the independence of a gray wolf pack, as shown below: (6) (7) (8) (9) In the formula, the values ​​of W1, W2 and W3 represent the learning rates of ω wolf on α wolf, β wolf and δ wolf, respectively.

[0026] The improved LSTM network structure will be introduced next, specifically as follows: Its structural diagram is as follows Figure 4 As shown, the improved LSTM network structure mainly consists of four layers: forget gate, input gate, update gate, and output gate. Figure 4 It can be seen that the improved LSTM network adds one memory function unit to each hidden unit recursively hidden layer neuron; Its input layer is as follows: it receives a fixed-length probe demand sequence, and converts a structured probe demand vector D into a sequence data Dseq=[d1, d2, ..., dT] through an embedding layer or repeated sampling. Each time step dt can contain demand information with different focuses (such as first inputting depth information, then inputting resolution constraints, etc.), thereby providing context for the model. Its output layer is a fully connected layer that outputs the predicted optimal coil parameter set Pott, for example, [Tsize, Tturns, Rbsize, ...].

[0027] The memory function unit is an auxiliary memory path embedded within the LSTM hidden unit. It is independent of the standard input gate, forget gate, and output gate workflow and is specifically designed to capture and store long-term dependencies related to electromagnetic transient processes. The transient electromagnetic secondary field is a complex signal containing multiple decaying components. The memory function unit acts like a "physical process recorder," specifically responsible for remembering features with long-term effects, such as "eddy current diffusion depth" and "response time constants of different geoelectric layers," enabling the network to understand complex causal relationships such as "deep exploration requires a larger emitted magnetic moment and a longer observation time."

[0028] Step 102: Using the improved GWO algorithm as the optimizer and the improved LSTM network structure as the optimized object, obtain the hyperparameter combination, specifically: Use the improved GWO to find the optimal architecture and training parameters for the LSTM model; 1. Define the GWO search space (the "wolf" dimension): The position Xi of each gray wolf is a vector representing a set of hyperparameters of the LSTM model: Xi = [lstm_layers, lstm_units, dense_layers, dense_units, learning_rate, dropout_rate, batch_size]; lstm_layers: {1, 2, 3} (integer, representing the number of LSTM layers); lstm_units: [32, 256] (integer, number of neurons per layer); learning_rate: [1e-5, 1e-2] (logarithmic scale); dropout_rate: [0.1, 0.5] (to prevent overfitting); batch_size: {32, 64, 128} (integer); 2. Improved GWO optimizer settings: Population size: 20-30 wolves; Maximum number of iterations: 50; The improvement strategies specifically include: (1) Initialization: Use the Tent chaotic map to initialize the wolf pack positions in the search space.

[0029] (2) Convergence factor a: adopt a nonlinear decreasing strategy, such as a=2 (1-cos(π)) t / (2 Tmax)), where t is the current iteration number and Tmax is the maximum iteration number.

[0030] (3) Position update: Integrating the PSO concept, an inertia weight ω (linearly decreasing from 0.9 to 0.4) and the individual historical best Pbest are introduced; 3. Fitness assessment: For each wolf Xi (i.e., for each set of hyperparameters): a. Model building: Based on the hyperparameters decoded by Xi, an improved LSTM model is instantiated. The improvements include adding memory functional units in the hidden layers to better capture the long-term dependence of electromagnetic transient processes. b. Rapid training and validation: The pre-defined training dataset is divided into a training set, a validation set, and a test set in a 70:15:15 ratio. Using the current hyperparameters Xi, the LSTM model is trained quickly on the training set (e.g., for only 10 epochs, with early stopping to prevent overfitting) to efficiently evaluate the potential of the set of hyperparameters, rather than achieving full convergence; the model's performance metrics are then computed on the validation set. c. Calculate fitness: Fitness(Xi) = 1 / (1 + MSEval + λ) MAEval) MSE val Predict parameters P for the model on the validation set. pred With the true optimal parameter P true Mean squared error; MAE val The mean absolute error is used as a supplementary indicator; λ is the weighting coefficient, which is 0.5 in this invention. A higher fitness value indicates that the set of hyperparameters X... i The better the performance of the defined LSTM model; 4. GWO Iterative Optimization Loop: The GWO algorithm iteratively updates the position of the wolf pack (i.e., the hyperparameter combination) based on the fitness value using an improved update formula. Record the position X of the alpha wolf with the highest fitness in each generation. α； The optimization terminates when the maximum number of iterations is reached or the fitness no longer improves significantly over multiple generations. Output: Globally optimal combination of hyperparameters X α ; Step 103: Initialize the parameters of the LSTM network structure based on hyperparameter combination, and construct a structural parameter determination model, specifically as follows: The optimal hyperparameter X was found using GWO. α We then reconstruct a completely new, structurally defined, improved LSTM model.

[0031] In step 2, the structural parameter determination model is trained, specifically as follows: Step 201: Construct a training dataset, which consists of detection requirements and coil parameters, specifically: 1. Define the multidimensional parameter space and sampling strategy (1) Definition and sampling of the detection demand space D: Latin hypercube sampling is used to ensure that sample points can uniformly cover the entire multidimensional space, avoid clustering, and obtain the maximum spatial coverage with the minimum number of samples. The multidimensional parameter space specifically includes: d target (Main detection depth): Range: 1m (shallow fine exploration) ~ 500m (deep resource exploration); Sampling strategy: Uniform sampling on a logarithmic scale, because the detection depth is approximately logarithmically related to the coil size, ensuring sufficient samples in shallow, medium, and deep regions. 10 (d) target )~Uniform(log 10 (1), log 10 (500)); resolution weight (Resolution preference): Range: 0~1; Physical meaning: 0 represents "pursuing deep exploration capabilities at all costs", and 1 represents "prioritizing the highest resolution for shallow areas"; Sampling strategy: uniform sampling, resolution weight ~Uniform(0,1); target type (Target geological body type): Types: Use One-hot encoding, such as [1,0,0] low-resistivity thin layer (e.g., water-bearing fault), [0,1,0] high-resistivity rock mass (e.g., granite), [0,0,1] locally good conductor (e.g., metallic ore body); Sampling strategy: Random sampling is performed according to a preset probability, for example, P=[0.4, 0.3, 0.3], to simulate the distribution of target bodies in actual exploration; noise level (Environmental noise level): Range: 0.1 (extremely quiet environment) ~ 0.9 (strong industrial interference zone); Sampling strategy: Uniform sampling; space constraint (Spatial constraints for device layout): Range: 0.5m (narrow alleyway) ~ 5.0m (open ground); Sampling strategy: Uniform sampling; (2) Definition and sampling of coil parameter space P: For each sampled D i We need to be under its constraints (especially space) constraint Randomly generate a large number of candidate coil parameters P j ; The specific parameters of the transmitting coil are as follows: T shape : Select randomly from {circle, square} with equal probability; T size For a circle with diameter and a square with side length, the sampling range is [0.2m, min(space)]. constraint [3.0], Sampling strategy: Uniform sampling; T turns : Integer, range [4, 20], sampling strategy: uniform sampling; T current Transmit current, range [1A, 30A], sampling strategy: uniform sampling; Bottom surface receiving coil parameter R b (Similar definition, but typically smaller in size than the transmitting coil); Side receiving coil parameter R s1 R s2 ...: (Similar definition, consider its tilt angle); Ultimately, 50,000 (D) units were generated via LHS. i P j ) combinations, where each D i Corresponding to approximately 100-200 sets of random P j .

[0032] 2. High-fidelity electromagnetic forward modeling simulation This invention employs high-fidelity electromagnetic forward modeling simulation and uses automated scripts to drive data simulation. The automated process includes: (1) Script generates geometric model: based on (D i P j The coil parameters in the simulation software are used to automatically create a 3D model, including: Transmitting coil and receiving coil (accurately model their shape, size, and position); A layered geodetic model (containing at least an air layer and 2-3 layers with different electrical parameters). According to target type Embed a typical geological body (such as a low-resistivity thin plate) in the appropriate location. (2) Set up the physical field: Set up the transient magnetic field physical field and define the waveform of the transmitted current (such as a bipolar square wave, and adjust the turn-off time according to the detection depth). (3) Mesh generation: Adaptive mesh generation is adopted, and the mesh is refined near the coil and in the target geological body area to ensure calculation accuracy; (4) Solving calculations: A large number of simulation tasks are run in parallel on the HPC (High Performance Computing) cluster, with calculation time ranging from a few minutes to several hours; (5) Result extraction: After the simulation is completed, the automated script extracts the curve V(t) of the induced electromotive force (EMF) on each receiving coil decaying over time from the result file.

[0033] 3. Calculate performance indicators and determine optimal parameters (1) Performance index vector Q ij Calculation: Analyze the attenuation curve V(t) obtained from the simulation and calculate: SNR early Early signal-to-noise ratio (SNR): The average signal amplitude within the first time window after shutdown (e.g., 0.1ms~0.5ms) is divided by the simulated ambient noise during that time period (based on noise level). evel generate); SNR late Signal-to-noise ratio (SNR) for late-stage signals, taken as the SNR within the last time window (e.g., 10ms~50ms); Peak Response Peak response, relative to target type For the indicated target body, calculate the maximum difference between its response amplitude and the background field response amplitude; Blind Zone_Thickness : Detect the thickness of the blind zone and estimate the thickness of the shallow layer where signals cannot be effectively distinguished by analyzing the attenuation gradient and noise level of the early signal; (2) Determine the optimal parameter P truei : Define a dynamic objective function F: for each D i Its optimal parameter is P that maximizes the following function. j : F(Q) ij =w1 SNR early +w2 SNR late -w3 Blind Zone_Thickness +w4 Peak Response ; Dynamic weights [w1, w2, w3, w4]: These weights are based on D i resolution weight Dynamic adjustment; When resolution weight When approaching 1 (high resolution): significantly increase w1, increase w3, and decrease w2. This means prioritizing early signal protection and reducing blind spots.

[0034] When resolution_weight approaches 0 (deep detection): w2 is significantly increased, while w1 and w3 are decreased. This means that the signal-to-noise ratio of late-stage signals is prioritized.

[0035] For each D i From its corresponding 100-200 P j In the middle, choose to make F(Q) ij The largest one forms the final data pair (D). i P truei ).

[0036] Ultimately, a training dataset containing approximately 50,000 high-quality samples was obtained.

[0037] Step 202: Perform full training of the parameter determination model based on the training dataset, specifically as follows: Finding the optimal hyperparameter combination X by improving GWO α Then, we move on to the final model training phase; 1. Model Reconstruction and Initialization To prevent ambiguity in the technical solution, the model reconstruction in step 1 will be described in detail here: (1) Model reconstruction: Based on X α The decoded parameters (e.g., LSTM layer number = 2, unit per layer = 128, Dropout = 0.2, ...) are used to re-instantiate a new, structurally determined improved LSTM model in a deep learning framework (e.g., TensorFlow / Keras or PyTorch). The structure of this model is consistent with the model prototype used for hyperparameter search, but the structure is fixed and optimal at this time. Weight initialization: Use optimized initialization methods such as HeNormal or GlorotUniform to initialize all weights of the model, without inheriting the weights of any model in the GWO fast evaluation phase; 2. Training process (1) Data preparation: Partitioning: The 50,000 samples in the training dataset are divided into a training set, a validation set, and a test set in a 70%:15%:15% ratio to ensure that the data distribution is consistent after partitioning; Standardization: For input D and output P true Z-score standardization is performed to accelerate convergence and improve model stability; (2) Training configuration: Optimizer: The AdamW optimizer is used, which combines Adam's adaptive learning rate and weight decay (regularization), and usually achieves better generalization performance. Initial learning rate: using X α The optimal learning rate (e.g., 0.001) was found. Learning rate scheduling: The ReduceLROnPlateau strategy is adopted. When the validation set loss no longer decreases for 5 consecutive epochs, the learning rate is halved. A minimum learning rate (e.g., 1e-7) is set, and scheduling stops when the learning rate is lower than this value. Loss function: Use smoothed L1 loss, which is less sensitive to outliers than MSE and makes the training process more stable; Early stopping: Monitor the validation set loss. If the validation set loss does not decrease for 15 consecutive epochs, terminate training and backtrack to the model weights of the epoch with the lowest validation set loss. (3) Training execution and monitoring: Batch size: Use X α The optimal batch size found (e.g., 64); Number of training rounds: Set a large value (e.g., 200), and the early stopping mechanism will automatically control the actual number of rounds; Monitoring: During training, the training set loss and validation set loss curves are plotted in real time. A healthy training process should be characterized by both curves decreasing smoothly and eventually stabilizing. The validation set loss should not be significantly higher than the training set loss; otherwise, it indicates overfitting. 3. Final model evaluation and solidification (1) Performance evaluation: After training, load the best model saved from early stopping and perform a final evaluation on the test set: Calculate MSE and MAE; Conduct case studies, select several test samples, and compare the P values ​​predicted by the model. pred With the real P true And observe its physical plausibility (e.g., whether the predicted coil size increases with increasing probe depth); (2) Model solidification: Export the final model that meets the performance requirements (including its computational graph structure and all weight parameters) as a .h5 or .pt file. This file is the intelligent core that can be integrated into the actual device. (3) Deployment and Inference: Load this model file into the embedded system or industrial control computer of the device. When the user inputs a new detection requirement D, the system only needs to perform the forward propagation of the model once to output the optimal coil parameters P within milliseconds. opt This drives the hardware to perform adaptive reconfiguration.

[0038] This invention provides an embodiment: To verify the effectiveness of this invention, an advanced geological prediction scenario for a coal mine roadway is used as an example. The detection task is to predict water-bearing structures within 80 meters in front of the tunnel face. It is known that the roadway space is narrow and the deployment space is limited to 1.5 meters. 1. Input detection requirements The specific requirements for this probe are input through the device's human-machine interface, generating a digital vector D: d target (Target depth): 80 meters; resolution weight (Resolution preference): 0.8 (emphasizing high resolution for accurate identification of water-bearing faults); target type (Target type): [1, 0, 0] (Low-resistivity thin film); noise level (Noise level): 0.6 (moderate industrial interference); space constraint (Spatial constraints): 1.5 meters; 2. Model Inference and Parameter Determination The vector D is input into a fully trained structure parameter determination model (based on an improved GWO-LSTM) deployed in the device's embedded system. The model completes forward propagation in milliseconds and outputs the optimal coil parameter set P. opt : Transmitting coil: square in shape, with a side length of 1.2 meters, 10 turns, and a transmitting current of 18A; Bottom receiving coil: square, with a side length of 0.8 meters, using an edge-reinforced turn distribution pattern; Side receiving coil: rectangular, measuring 0.8m × 0.6m; 3. Coil Adaptive Reconstruction and Detection The control system of the device is based on P opt The adjustable driving parameter coil system is reconfigured, the mechanical telescopic mechanism adjusts the coil to the specified size, the electronic switch matrix switches to the specified number of turns distribution, and after the reconfiguration is completed, the device automatically performs transient electromagnetic detection; 4. Result Verification Compared with the conventionally preset general coil parameters (square coil with a side length of 1m and 12 uniform turns), under the same environment, the parameters determined by this invention improved the early signal-to-noise ratio of the secondary field signal by about 35% and the late signal-to-noise ratio by about 20%, and reduced the effective detection blind zone by about 15 meters. Subsequent drilling verification showed that the positioning accuracy of a 2-meter-wide water-bearing fault at 65 meters was better than that of the traditional parameter configuration, proving that the structural parameters determined by the method described in this invention can significantly improve the detection performance and the accuracy of geological prediction.

[0039] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0040] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A transient electromagnetic signal transceiver, characterized in that, include: The frame has multiple sides that intersect its base at an angle. A signal transmitting coil is disposed on the bottom surface of the frame and is used to transmit transient electromagnetic signals to the geological body to be detected; Multiple signal receiving coils are respectively disposed on the bottom surface and each side surface of the frame. The signal receiving coils are used to receive the electromagnetic response signals fed back by the geological body to be detected. Based on the improved GWO-LSTM algorithm, the structural parameters of the signal transmitting coil and the signal receiving coil are determined according to the detection requirements.

2. The transient electromagnetic signal transceiver according to claim 1, characterized in that, The frame is shaped like a regular pyramid or a regular frustum. The signal transmitting coil is concentrically arranged with the bottom surface of the frame. The edges of the signal transmitting coil and multiple signal receiving coils are respectively arranged on the border of the corresponding frame surface.

3. The transient electromagnetic signal transceiver according to claim 2, characterized in that, The device further includes: a signal source and multiple storage units. The signal source is connected to the signal transmitting coil and is used to transmit current to the signal transmitting coil to drive the signal transmitting coil to transmit the transient electromagnetic signal. The storage units correspond one-to-one with the signal receiving coils and are used to store the electromagnetic response signals received by the corresponding signal receiving coils.

4. The transient electromagnetic signal transceiver according to claim 3, characterized in that, Based on the improved GWO-LSTM algorithm, the structural parameters of the signal transmitting coil and the signal receiving coil are determined according to the detection requirements, specifically: A structural parameter determination model is constructed based on the improved GWO-LSTM algorithm; Train the model to determine the structural parameters; The detection requirements are input into the structural parameters to determine the model, and the structural parameters are obtained. The signal transmitting coil and signal receiving coil of the transient electromagnetic signal transceiver are adjusted based on structural parameters.

5. The transient electromagnetic signal transceiver according to claim 4, characterized in that, A structural parameter determination model is constructed based on the improved GWO-LSTM algorithm, specifically as follows: Improvements were made to the GWO algorithm and LSTM network structure; The improved GWO algorithm is used as the optimizer, and the improved LSTM network structure is used as the optimized object to obtain the hyperparameter combination. A structural parameter determination model is constructed by initializing the parameters of the LSTM network structure based on hyperparameter combination.

6. The transient electromagnetic signal transceiver according to claim 5, characterized in that, The GWO algorithm is improved as follows: Based on the traditional GWO algorithm, Tent mapping is introduced as a random number generation mechanism; Based on the traditional GWO algorithm, a nonlinear convergence factor is introduced to replace the convergence factor; Based on the traditional GWO algorithm, an inertial weight is introduced to adjust the balance between global optimization and local optimization.

7. The transient electromagnetic signal transceiver according to claim 6, characterized in that, The LSTM network structure is improved as follows: A memory function unit is added to each hidden unit of the recursive hidden layer neuron in the traditional LSTM network structure.

8. The transient electromagnetic signal transceiver according to claim 7, characterized in that, The model for determining structural parameters is trained as follows: Construct a training dataset, which consists of detection requirements and coil parameters; The parameter-determining model is fully trained based on the training dataset.

Images