Efficient detection method of submarine cable based on neural network magnetic data inversion

Abstract

The application discloses a kind of based on neural network magnetic data inversion submarine cable efficient detection method, belong to the technical field of detection of submarine cable.The method includes: receiving magnetic anomaly time series signal, local time domain feature is extracted by one-dimensional convolution and obtains multi-channel one-dimensional convolution feature map by ReLU activation and normalization;Channel weight adaptive weighting is realized by SE channel attention module;Obtain double attention enhanced feature map by multi-head self-attention of lower triangular mask mechanism;After being compressed by global average pooling, respectively, output spatial position parameters by position parameter regression module, output magnetic anomaly detection probability by detection probability classification module.The application constructs end-to-end deep learning network, avoids the false solution and time-consuming problem caused by improper parameter selection of Euler deconvolution method, can accurately and efficiently identify and locate submarine cable magnetic anomaly, significantly improves detection accuracy and efficiency, and is suitable for fast detection of underwater magnetic targets such as submarine cable and unexploded ordnance.

Description

Technical Field

[0001] This invention belongs to the field of submarine cable detection technology, specifically relating to an efficient detection method for submarine cables based on neural network magnetic data inversion. Background Technology

[0002] In offshore wind power systems, submarine cables are a crucial component for transmitting electrical energy, and their quality directly affects the overall performance of power transmission and the lifespan of the wind power system. However, during use, submarine cables are subjected to long-term seawater erosion and wave scouring, resulting in decreased water resistance. Simultaneously, with insulation aging, leakage currents may occur in the cables, leading to localized temperature increases at fault points. Furthermore, submarine cable breakage accidents caused by external forces such as marine operations and ship anchor damage are frequent. Failure to promptly detect and address these problems seriously threatens the safe and stable operation of the cross-sea power grid, severely impacting the quality of offshore wind power systems and the economic benefits for enterprises. Currently, magnetic detection is the most widely used and effective detection technology, divided into active and passive methods (absolute magnetic field method and metal detection method). Commonly used equipment for marine magnetic detection includes resonant magnetometers, fluxgate magnetometers, and magnetic gradient meters. Active detection requires applying alternating current to the submarine cable to receive electromagnetic wave signals, which is susceptible to noise and signal attenuation. Mature products include the TSS-350 pipeline tracker, Smartrak and its upgraded version SmartSearch robot, and the Orion underwater pipeline survey system. In passive detection, the TSS-440 and TSS-660 pipeline trackers and the Type 5000 metal detector use the metal detection method, which does not require an external tracking signal, but is affected by the target size and magnetic materials. The G882 cesium optically pumped magnetometer is an absolute magnetic field detection method, with high sensitivity and fast sampling rate, suitable for different latitudes.

[0003] Among them, Euler deconvolution, a commonly used method for magnetic anomaly inversion, obtains cable locations by processing magnetic field data and sequentially performing field separation, denoising, and Euler deconvolution. However, the Euler method still has shortcomings, such as the susceptibility to spurious solutions. The selection of parameters such as the structure index, horizontal gradient filter coefficients, constraint distance, and convergence degree significantly affects the inversion accuracy and is time-consuming. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and disclose an efficient detection method for submarine cables based on neural network magnetic data inversion. This method solves the problems of spurious solutions, low inversion accuracy, and low efficiency of the Euler deconvolution magnetic anomaly inversion method. It is an end-to-end deep learning network that can build a fast and intelligent algorithm for identifying and locating magnetic anomalies in submarine cables.

[0005] To achieve the above objectives, the present invention employs the following technical solution: An efficient method for detecting submarine cables based on neural network magnetic data inversion includes the following steps: S1, Receive magnetic anomaly timing signal, extract local time domain features of the signal through one-dimensional convolution, process the local time domain features through ReLU activation function to obtain nonlinear features, and normalize the nonlinear features to obtain multi-channel one-dimensional convolution feature map. S2, the multi-channel one-dimensional convolutional feature map is first processed by global average pooling to obtain global information, the weight coefficients of each channel are extracted from the global information, and the multi-channel one-dimensional convolutional feature map and the weight coefficients of each channel are multiplied to obtain the optimized feature map. S3, after the optimized feature map is processed by the multi-head attention mechanism, multiple sets of attention weights are processed by the lower triangular mask mechanism to obtain a feature map with dual attention enhancement. S4, perform global average pooling on the feature map enhanced by dual attention to obtain a compressed low-dimensional global vector; S5, the spatial location parameters of the magnetic anomaly signal are obtained by the location parameter regression module through the low-dimensional global vector, and the probability of the presence of the submarine cable with the magnetic anomaly signal is obtained by the detection probability classification module.

[0006] A further improvement of the present invention is that: Preferably, in S1, the magnetic anomaly timing signal is processed by equal-length padding convolution before being processed by one-dimensional convolution.

[0007] Preferably, in S1, the horizontal axis of the multi-channel one-dimensional convolutional graph represents the position coordinates of the cable, and the vertical axis represents the value of the magnetic anomaly timing signal.

[0008] Preferably, in S2, during the global average pooling process, the temporal features of each channel in the multi-channel one-dimensional convolutional feature map are compressed into a single value, and the temporal features include peak features and gradient features; the single values ​​after compression of the temporal features of each channel are summarized to obtain global information.

[0009] Preferably, in S2, the weight coefficients of each channel are obtained through learning by a fully connected layer and a sigmoid activation function.

[0010] Preferably, in S3, the feature map obtained after processing multiple sets of attention weights through the lower triangular masking mechanism is normalized through residual connections and LayerNorm layers to obtain a feature map with dual attention enhancement.

[0011] Preferably, in S5, the position parameter regression module includes a horizontal position parameter regression module and a vertical position parameter regression module; Both the horizontal position parameter regression module and the vertical position parameter regression module include a fully connected layer and a regression output layer.

[0012] Preferably, in S5, the loss function of the location parameter regression module training process is the mean squared error loss function; the loss function of the detection probability classification module training process is the binary cross-entropy loss function.

[0013] A high-efficiency submarine cable detection device based on neural network magnetic data inversion includes: The signal processing module is used to receive magnetic anomaly time-series signals, extract local time-domain features of the signals through one-dimensional convolution, process the local time-domain features through the ReLU activation function to obtain nonlinear features, and normalize the nonlinear features to obtain a multi-channel one-dimensional convolution feature map. The SE channel attention module is used to first process the multi-channel one-dimensional convolutional feature map through global average pooling to obtain global information, extract the weight coefficients of each channel from the global information, and multiply the one-dimensional convolutional feature map and the weight coefficients of each channel to obtain an optimized feature map. The masked self-attention block module is used to process the optimized feature map through a multi-head attention mechanism, and then process multiple sets of attention weights through a lower triangular masking mechanism to obtain a feature map with dual attention enhancement. The feature compression layer is used to perform global average pooling on the feature map enhanced by dual attention to obtain a compressed low-dimensional global vector. The output module is used to obtain the spatial location parameters of the magnetic anomaly signal from the low-dimensional global vector through the location parameter regression module, and to obtain the probability of the presence of the submarine cable with the magnetic anomaly signal through the detection probability classification module.

[0014] Preferably, during the training process, the device inputs the training set into the device, performs forward propagation to calculate the output, calculates the error between the predicted value and the true value based on the loss function, calculates the gradient through the backpropagation algorithm, and updates the device parameters using the optimizer until the training process is completed.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) Compared with the Euler deconvolution method, which relies on the manual selection of parameters such as structural index and horizontal gradient filter coefficient and is prone to producing spurious solutions, this invention directly establishes the mapping relationship between magnetic anomaly time-series signals and cable positions by constructing an end-to-end deep learning network, avoiding the accuracy loss and time consumption problems caused by improper parameter selection, and significantly improving the accuracy and efficiency of submarine cable magnetic anomaly identification and positioning.

[0016] (2) The SE channel attention module is adopted to dynamically suppress irrelevant interferences such as seawater noise and magnetic sediments through the "compression-excitation-weighting" mechanism, and accurately focus on the cable-related features; the mask self-attention block realizes causal constraints through the lower triangular mask mechanism, avoids false feature learning, and enables the model to maintain stable positioning performance in complex marine environments.

[0017] (3) The dual-branch output design (position parameter regression module and detection probability classification module) realizes one operation and dual output, without the need to split the horizontal and vertical positioning process, further simplifying the operation steps, which can meet the rapid detection needs of large-scale offshore wind farms and help to discover cable fault hazards in a timely manner.

[0018] (4) This invention can detect the location and potential faults of submarine cables in a timely and accurate manner (such as seawater erosion, insulation aging, and magnetic anomalies caused by external force damage), effectively avoiding accidents such as cable breakage and power grid shutdown caused by failure to detect faults in time; by accurately locating the cable position, it can also provide safety guidance for marine operations and ship navigation, reduce the risk of external force damage such as anchor damage, ensure the stable operation of offshore wind power systems, and thus maintain the power generation benefits of enterprises and the safety benefits of power grid operation.

[0019] (5) The method of the present invention uses the synergistic effect of technologies such as BatchNorm batch normalization, residual connection, and LayerNorm layer normalization to enable the model to maintain stable convergence in magnetic data training in different sea areas, different burial depths, and different cable types, thus avoiding gradient vanishing or parameter oscillation problems. Attached Figure Description

[0020] Figure 1 This is a flowchart of an efficient submarine cable detection method based on neural network magnetic data inversion used in this invention; Figure 2 This is a training flowchart of an efficient submarine cable detection device based on neural network magnetic data inversion according to the present invention. It shows ten consecutive steps in sequence: magnetic field data collection, dataset generation, dataset partitioning, neural network model construction, model training, model evaluation, performance judgment, actual data preprocessing, model application and location information output. Each step is connected in a logical order. Detailed Implementation Hereinafter, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," "third," or "fourth" may explicitly or implicitly include one or more of that feature.

[0021] The methods provided in this application can be applied to terminal devices such as mobile phones, tablets, wearable devices, in-vehicle devices, augmented reality (AR) / virtual reality (VR) devices, laptops, ultra-mobile personal computers (UMPCs), netbooks, and personal digital assistants (PDAs). This application does not impose any restrictions on the specific type of terminal device.

[0022] It should be noted that the terms "first," "second," etc., used in the specification and drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] See Figure 1 This invention discloses an efficient method for detecting submarine cables based on neural network magnetic data inversion. By collecting and standardizing magnetic field data from the sea area where the submarine cable is located, an end-to-end deep learning model is constructed, comprising a one-dimensional convolutional group, an SE channel attention module, and a masked self-attention block (MAB). After training and evaluation, the preprocessed actual magnetic field data is processed, simultaneously outputting the horizontal position detection probability and vertical distance of the submarine cable. The final positioning result is presented in an intuitive form, achieving efficient and accurate submarine cable detection. This method employs a multi-task deep learning network architecture oriented towards magnetic anomaly time-series signals. The overall approach achieves magnetic anomaly detection and target localization in an end-to-end manner. Each module is sequentially connected according to the data flow, forming a complete feature extraction, enhancement, and multi-task output chain. The method includes the following steps: S1, the input layer of the network receives the magnetic anomaly time-series signal. The signal is one-dimensional time-series data, corresponding to the magnetic field data that changes over time collected by the magnetic sensor, and is the raw data entry point for the entire network. The signal first enters a one-dimensional convolutional group. The local time-domain features of the signal are extracted through one-dimensional convolution. The specific local time-domain features are processed by the ReLU activation function to obtain nonlinear features. The nonlinear features are then normalized through batch normalization, finally generating a multi-channel one-dimensional convolutional feature map. The horizontal axis of this one-dimensional convolutional feature map represents the specific location of the cable, and the vertical axis represents the value of the magnetic anomaly time-series signal.

[0024] The entire one-dimensional convolution process described above is combined with the ReLU activation function to introduce nonlinearity, which enhances the feature representation ability of the network. The BatchNorm batch normalization operation accelerates network training and alleviates the gradient vanishing problem. Finally, it completes the extraction of local temporal features of magnetic anomaly time series signals and captures key local patterns such as abnormal peaks and gradient changes in the signals.

[0025] Specifically, the magnetic anomaly timing signal is magnetic anomaly data arranged in time sequence.

[0026] Furthermore, before performing convolution processing, the one-dimensional convolution group uses equal-length padding (padding=SAME) to ensure that the temporal length remains unchanged after convolution.

[0027] S2, the one-dimensional convolutional feature map enters the SE channel attention module. This module achieves channel-level feature enhancement through a three-step mechanism of "compression-activation-weighting": First, global average pooling is used to compress the temporal features of each channel into a single value, completing the aggregation of global information and obtaining global information. The specific temporal features include peak features and gradient features, corresponding to the peak channel and gradient channel, respectively. Then, a fully connected layer and a sigmoid activation function are used to learn and output the weight coefficients of each channel, realizing adaptive modeling of the importance of different channels. Finally, the one-dimensional convolutional feature map is multiplied with the channel weights channel by channel to obtain an optimized feature map that highlights the temporal feature channels. The number of channels in the optimized feature map is the same as the number of input channels. This process strengthens the key feature channels for magnetic anomaly detection and localization tasks, suppresses redundant noise channels, and improves the effectiveness of features.

[0028] S3, Feature Map Input Mask Self-Attention Block (MAB), is the core module for modeling long-range temporal dependencies. To optimize the feature map, multiple sets of attention weights are first computed in parallel using a multi-head attention mechanism. Different attention heads (groups) map features to different subspaces, capturing long-range temporal correlations in different dimensions of the magnetic anomaly signal. After processing with multiple sets of attention weights, only the current time step is allowed to access historical temporal information, strictly adhering to the causal constraints of the magnetic anomaly signal and conforming to physical reality. Then, a lower triangular mask mechanism is used to process the multiple sets of attention weights, obtaining the feature map after processing. The feature map is then normalized using a LayerNorm layer via residual connections to alleviate the gradient vanishing problem in deep networks, stabilize the training process, and ensure model convergence efficiency. Finally, the modeling of the global temporal dependencies of the magnetic anomaly signal is completed, resulting in a feature map enhanced by dual attention.

[0029] S4, the feature map enhanced by dual attention, then enters the feature compression layer. This layer uses global average pooling to compress high-dimensional temporal features into low-dimensional global feature vectors. This significantly reduces the amount of subsequent computation while fully preserving key feature information, providing a unified feature basis for multi-task output. The key feature information includes peak features, gradient features, and temporal dependency features modeled by dual attention of the magnetic anomaly signal.

[0030] S5, after the feature compression layer, the network is divided into three parallel task branches to achieve multi-task joint learning: two branches are position parameter regression modules, namely the horizontal position parameter regression module and the vertical position parameter regression module, corresponding to the horizontal position and vertical distance regression tasks, respectively. Each branch maps global features through a fully connected layer, outputs the spatial position parameters of the target through the regression output layer, and finally completes the training and optimization of the regression task through the loss function; the third branch is the detection probability classification module, which is a binary classification task used to determine whether there is a target in the magnetic anomaly signal. The specific target is the submarine cable that generates the magnetic anomaly signal. The output is the detection probability, which represents the confidence that there is a target in the current signal sequence. A threshold is set. If the probability is greater than the threshold, it is determined that "a target exists" and the position regression branch is triggered; if the probability is less than the threshold, it is determined that "a target does not exist" and the subsequent localization process is skipped directly; the classification task is also trained through the corresponding loss function.

[0031] During this process, if the detection probability classification module fails to detect the target for an extended period of time, the entire detection process will terminate directly.

[0032] The network ultimately achieves end-to-end training through multi-task joint loss (a weighted sum of regression loss and classification loss), completing the simultaneous prediction of the existence of magnetic anomalies and the target location parameters. The entire architecture takes into account the causal constraints of time-series signals, adaptive feature enhancement, and multi-task collaborative optimization, and is suitable for the detection and localization of underground targets, unexploded ordnance, and underwater magnetic targets in the field of magnetic detection.

[0033] A second aspect of the present invention discloses a high-efficiency detection device for submarine cables based on neural network magnetic data inversion, the device comprising: See Figure 2 The initial training process for the aforementioned device includes the following steps: S1. Collect magnetic field data. Using professional magnetic field measuring instruments such as magnetometers, collect raw magnetic anomaly time-series signals containing information such as magnetic field strength at set time intervals in the sea area where submarine cables are laid, and record metadata such as measurement location to obtain basic magnetic field data for subsequent analysis and modeling.

[0034] S2 preprocesses the actual magnetic field data, performs interpolation on the actual collected magnetic field data to supplement missing or sparse data points, and uses data transformations such as Z-score normalization to improve the quality of the actual magnetic field data and provide a better data foundation for training the model.

[0035] S3 generates a dataset. After removing abnormal and erroneous data points from the collected raw magnetic field data, it is organized and stored according to standard data storage formats such as CSV and HDF5, transforming the raw data into a dataset format suitable for neural network processing.

[0036] S4. Divide the dataset into training, validation, and test sets using random partitioning or stratified sampling methods, dividing the generated dataset into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. This ensures that the data distribution characteristics of each subset are similar, preparing data subsets for different purposes such as training, performance validation, and effect testing of the neural network model.

[0037] S5, building as Figure 2 The neural network model shown selects a network architecture adapted to the time-series data based on the characteristics of the problem. The specific structure is as follows: Input layer: Receives magnetic anomaly timing signals as the raw input data for the model.

[0038] One-dimensional convolution group: used to extract local features of magnetic anomaly signals by first performing convolution with equal-length padding, combined with ReLU activation function and batch normalization.

[0039] The SE channel attention module first compresses the feature map through global average pooling, then completes the excitation process through a fully connected layer and a sigmoid activation function, outputting channel weights. Finally, through the weighted operation of the feature map and the channel weights, key features are enhanced and redundant information is suppressed to obtain an optimized feature map. The Masked Self-Attention Block (MAB) module integrates a multi-head attention mechanism and uses a lower triangular masking strategy to process and optimize the feature map. It only allows the current position to focus on historical time-series information, which matches the causal characteristics of magnetic anomaly signals.

[0040] Residual connections and LayerNorm modules: connect the input and output of the masked self-attention block, and work with layer normalization operations to alleviate the gradient vanishing problem and stabilize model training; Feature compression layer: The features processed above are compressed using global average pooling to extract core feature information; Output layer: It is divided into two branches. The first is the vertical distance regression module composed of fully connected layers, which outputs the vertical distance prediction results of submarine cables. The second is the horizontal position detection probability classification module (binary classification) composed of fully connected layers, which outputs the detection probability of the horizontal position of submarine cables.

[0041] During the training process of this device, the regression module uses the mean squared error loss function, and the classification module uses a loss function adapted to binary classification tasks, such as the binary cross-entropy loss function, to jointly optimize the model performance.

[0042] Based on the PyTorch deep learning framework, code was written to complete the construction of the above structure, forming a neural network capable of processing magnetic anomaly time-series signals and locating submarine cables.

[0043] S6, Training the Model: Input the training set data into the constructed neural network model, calculate the output through forward propagation, calculate the error between the predicted value and the true value based on the loss function, calculate the gradient through the backpropagation algorithm, update the model parameters using the optimizer (Adam), and repeat this process until the loss function converges or the set number of training rounds is reached, allowing the model to learn the correlation pattern between magnetic field data and the location of submarine cables, and adjust the parameters to optimize model performance.

[0044] S7 evaluates the model through numerical testing and ablation experiments. In numerical testing, validation set data is input into the trained model to calculate evaluation metrics such as mean squared error and mean absolute error to measure the model's prediction accuracy on unseen data. Ablation experiments involve changing or removing some components of the model, retraining and evaluating the model, analyzing the impact of these components on the model's performance, comprehensively evaluating the performance of the trained model, and understanding the contribution of each part of the model to the performance.

[0045] S8. Determine whether the requirements are met by comparing the indicators obtained from the model evaluation with the pre-set performance thresholds to determine whether the model has reached the expected performance standard, and then decide on the subsequent operations.

[0046] Specifically, for horizontal position regression and vertical distance regression tasks, mean squared error (MSE) and mean absolute error (MAE) are used as evaluation metrics to measure the deviation between the predicted value and the true position parameters; for target detection binary classification tasks, accuracy, precision, recall, F1 score and AUC value are used as evaluation metrics to evaluate the model's ability to identify submarine cable targets.

[0047] S9, Application Model (Processing Actual Data, Locating Submarine Cables): Once the model meets the requirements, the pre-processed actual magnetic field data is input into the trained model. The model processes the data according to the learned patterns and outputs information related to the location of the submarine cable. The trained model is then used to process the actual data to locate the submarine cable.

[0048] S10 outputs submarine cable location information, presenting the submarine cable location information output by the model in an intuitive form such as text files and map annotations, making it convenient for users to obtain and use submarine cable location information.

[0049] The proposed method for submarine cable detection based on neural network magnetic data inversion is tailored to the characteristics of submarine magnetic anomaly data, the core requirements for cable location, and the complex marine environment. The entire process is specifically adapted, with each module's function and technical objective forming a closed loop. Combined with a mature deep learning technology system, it possesses full feasibility. The method also has the following advantages: (1) The magnetic anomaly signal generated by the submarine cable exhibits continuous temporal characteristics with the measurement time and navigation trajectory. Using the magnetic anomaly temporal signal as the input layer, it directly matches the original data form without the need for additional complex data form transformation. The subsequent one-dimensional convolution group adopts the padding=SAME mode, which can extract local features (such as magnetic anomaly peaks and gradient abrupt change points) while keeping the output temporal length consistent with the input, avoiding the loss of positional information. The ReLU activation function can enhance the model's ability to express nonlinear features and adapt to the complex mapping relationship between magnetic anomaly signals and cable positions. The BatchNorm batch normalization operation can standardize the input distribution of each layer, accelerate training convergence and suppress overfitting, solve the gradient divergence problem that is prone to occur in the training of temporal data, and lay the foundation for subsequent accurate positioning.

[0050] (2) In marine exploration scenarios, there are various irrelevant signals such as seawater flow noise, seabed magnetic sediment interference, and ship navigation interference. The SE channel attention module uses a complete logic of "global average pooling (compressing feature dimensions and retaining global information) → fully connected layer + Sigmoid activation (dynamically generating channel weights) → feature weighting (strengthening cable-related feature channels and weakening interference channels)" to accurately filter out magnetic anomaly features strongly related to cable location and shield invalid interference information. This design fully conforms to the physical logic that "marine magnetic data must be denoised before analysis," ensuring that the features learned by the model have practical engineering significance rather than being misled by interference signals.

[0051] (3) The propagation of magnetic anomaly signals follows the causal law of "historical time sequence affecting the current signal" (e.g., the magnetic anomaly intensity of the previous measurement point will affect the signal characteristics of adjacent points). The Masked Self-Attention Block (MAB) adopts the "lower triangular mask" strategy, which only allows the current position to focus on historical time sequence information and prohibits reverse dependence on future data, which fully conforms to the physical propagation characteristics of the signal and avoids the model learning meaningless false associations. At the same time, the residual connections of the modules can directly transmit shallow features, effectively alleviating the gradient vanishing problem in deep network training; LayerNorm layer normalization can stabilize the input distribution of each layer, avoid parameter oscillation during training, ensure that the model remains stable in multiple rounds of training, and finally form a localization model with strong generalization ability.

[0052] (4) Submarine cable positioning requires the simultaneous acquisition of two core types of information: horizontal position (planar coordinates) and vertical distance (burial depth). The dual-branch output layer is specifically designed: the vertical distance regression module directly outputs continuous burial depth values ​​through a fully connected layer, adapting to the "distance quantization" requirement; the horizontal position detection probability module (binary classification) outputs position probabilities through a cross-entropy loss function, quickly determining whether a cable exists in the detection area corresponding to the current time series. The two work together to achieve full coverage of the "horizontal + vertical" positioning dimensions. This design eliminates the need for additional processing steps; a single calculation is sufficient to obtain complete location information, perfectly matching the positioning requirements of actual engineering projects.

[0053] (5) The neural network model of this invention is built on the PyTorch deep learning framework, which has a complete time series data processing interface, a rich optimizer library (such as Adam, stochastic gradient descent) and mature distributed training support, which can quickly realize the encoding, training and deployment of the model. At the same time, the magnetic field data acquisition (magnetometer), data format conversion (CSV, HDF5), interpolation preprocessing (linear interpolation, spline interpolation) and other technologies involved in the method are all mature technologies in the field of marine exploration. The equipment is readily available and the operation is standardized, forming a seamless connection with the neural network model, ensuring that the entire exploration process can be engineered and implemented.

[0054] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0055] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0056] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A highly efficient method for detecting submarine cables based on neural network magnetic data inversion, characterized in that, Includes the following steps: S1, Receive magnetic anomaly timing signal, extract local time domain features of the signal through one-dimensional convolution, process the local time domain features through ReLU activation function to obtain nonlinear features, and normalize the nonlinear features to obtain multi-channel one-dimensional convolution feature map. S2, the multi-channel one-dimensional convolutional feature map is first processed by global average pooling to obtain global information, the weight coefficients of each channel are extracted from the global information, and the multi-channel one-dimensional convolutional feature map and the weight coefficients of each channel are multiplied to obtain the optimized feature map. S3, after the optimized feature map is processed by the multi-head attention mechanism, multiple sets of attention weights are processed by the lower triangular mask mechanism to obtain a feature map with dual attention enhancement. S4, perform global average pooling on the feature map enhanced by dual attention to obtain a compressed low-dimensional global vector; S5, the spatial location parameters of the magnetic anomaly signal are obtained by the location parameter regression module through the low-dimensional global vector, and the probability of the presence of the submarine cable with the magnetic anomaly signal is obtained by the detection probability classification module.

2. The efficient detection method for submarine cables based on neural network magnetic data inversion according to claim 1, characterized in that, In S1, the magnetic anomaly time-series signal is processed by equal-length padding convolution before being processed by one-dimensional convolution.

3. The efficient detection method for submarine cables based on neural network magnetic data inversion according to claim 1, characterized in that, In S1, the horizontal axis of the multi-channel one-dimensional convolutional graph represents the position coordinates of the cable, and the vertical axis represents the value of the magnetic anomaly timing signal.

4. The efficient detection method for submarine cables based on neural network magnetic data inversion according to claim 1, characterized in that, In S2, during the global average pooling process, the temporal features of each channel in the multi-channel one-dimensional convolutional feature map are compressed into a single value. The temporal features include peak features and gradient features. By summing up the compressed individual values ​​of the timing characteristics of each channel, global information can be obtained.

5. The efficient detection method for submarine cables based on neural network magnetic data inversion according to claim 1, characterized in that, In S2, the weight coefficients of each channel are obtained through learning the fully connected layer and the Sigmoid activation function.

6. The efficient detection method for submarine cables based on neural network magnetic data inversion according to claim 1, characterized in that, In S3, the feature map obtained after processing multiple sets of attention weights through the lower triangular mask mechanism is then processed by residual connections and LayerNorm layers for normalization to obtain a feature map with dual attention enhancement.

7. The efficient detection method for submarine cables based on neural network magnetic data inversion according to claim 1, characterized in that, In S5, the position parameter regression module includes a horizontal position parameter regression module and a vertical position parameter regression module; Both the horizontal position parameter regression module and the vertical position parameter regression module include a fully connected layer and a regression output layer.

8. The efficient detection method for submarine cables based on neural network magnetic data inversion according to claim 1, characterized in that, In S5, the loss function for the training process of the location parameter regression module is the mean squared error loss function; the loss function for the training process of the detection probability classification module is the binary cross-entropy loss function.

9. A high-efficiency detection device for submarine cables based on neural network magnetic data inversion, characterized in that, include: The signal processing module is used to receive magnetic anomaly time-series signals, extract local time-domain features of the signals through one-dimensional convolution, process the local time-domain features through the ReLU activation function to obtain nonlinear features, and normalize the nonlinear features to obtain a multi-channel one-dimensional convolution feature map. The SE channel attention module is used to first process the multi-channel one-dimensional convolutional feature map through global average pooling to obtain global information, extract the weight coefficients of each channel from the global information, and multiply the one-dimensional convolutional feature map and the weight coefficients of each channel to obtain an optimized feature map. The masked self-attention block module is used to process the optimized feature map through a multi-head attention mechanism, and then process multiple sets of attention weights through a lower triangular masking mechanism to obtain a feature map with dual attention enhancement. The feature compression layer is used to perform global average pooling on the feature map enhanced by dual attention to obtain a compressed low-dimensional global vector. The output module is used to obtain the spatial location parameters of the magnetic anomaly signal from the low-dimensional global vector through the location parameter regression module, and to obtain the probability of the presence of the submarine cable with the magnetic anomaly signal through the detection probability classification module.

10. The high-efficiency submarine cable detection device based on neural network magnetic data inversion according to claim 9, characterized in that, During the training process, the device inputs the training set, performs forward propagation to calculate the output, calculates the error between the predicted value and the true value based on the loss function, calculates the gradient through the backpropagation algorithm, and uses the optimizer to update the device parameters until the training process is completed.

Images