Deep learning enhancement method for geomagnetic signal data features

Abstract

The invention relates to the technical field of signal processing, in particular to a deep learning enhancement method for geomagnetic signal data features, which comprises the following steps: an acquisition step: acquiring a geomagnetic field signal; a preprocessing step: preprocessing the geomagnetic field signal to obtain a time sequence signal with variable data volume; a feature extraction step: performing data feature extraction on the time sequence signal to obtain a data feature extraction result of geomagnetic signal features; an enhancement step: performing deep learning processing on the data feature extraction result of the geomagnetic signal feature to obtain a data enhancement result of the geomagnetic field feature, and through deep learning enhancement, significantly improving the representation capability of the geomagnetic signal feature, so that the key feature in the signal is more obvious, and subsequent identification and classification are facilitated; a self-adaptive noise management mechanism enables the method to have strong resistance to environmental interference, and useful signals can be effectively extracted in a complex environment.

Description

Technical Field

[0001] The present invention relates to the field of signal processing technology, and in particular to a deep learning enhancement method for geomagnetic signal data features, which is applied to the fields of geomagnetic signal acquisition, processing, feature extraction and classification. Background Art

[0002] Geomagnetic signals are the manifestation of the Earth's magnetic field in a specific area, which contains rich geological, environmental and spatial location information. In recent years, geomagnetic signals have been widely used in navigation and positioning, mineral exploration, geological surveys, military reconnaissance and other fields. However, due to the common problems of weak signals, susceptibility to interference, and unclear features of geomagnetic signals, how to effectively extract and enhance the features of geomagnetic signals has become the focus and difficulty of research in this field.

[0003] Traditional geomagnetic signal processing methods mainly rely on filtering, denoising, Fourier transform and other technologies. Although they can improve signal quality to a certain extent, they often perform poorly when faced with geomagnetic signal processing in complex environments. In particular, when geomagnetic signals are affected by multiple factors such as electromagnetic interference, equipment errors, and environmental changes, traditional methods are difficult to effectively separate and extract useful signal features.

[0004] With the development of deep learning technology, its application in the field of signal processing has become a trend. In the existing technology, there have been some attempts to process geomagnetic signals based on methods such as convolutional neural networks (CNN) and recurrent neural networks (RNN), but these methods usually directly apply deep learning algorithms to the original signals and lack customized processing procedures for the characteristics of geomagnetic signals, resulting in limited model generalization ability, insufficient feature extraction, and sensitivity to noise.

[0005] In addition, most existing geomagnetic signal processing methods are single algorithms or fixed processes, lacking adaptive capabilities and making it difficult to cope with geomagnetic signals of different qualities and characteristics. At the same time, existing methods have limited effects on enhancing geomagnetic signal characteristics and cannot meet the needs of high-precision geomagnetic applications.

[0006] Therefore, there is an urgent need for a deep learning enhancement method designed for the characteristics of geomagnetic signals and with adaptive processing capabilities to improve the representation ability and classification accuracy of geomagnetic signal features. Summary of the invention

[0007] In response to the problems existing in the prior art, the present invention provides a deep learning enhancement method for geomagnetic signal data features. The method combines adaptive preprocessing, multi-mode segmented processing, progressive deep learning and feature verification mechanism, which can effectively improve the representation ability and classification accuracy of geomagnetic signal features.

[0008] The object of the present invention is to provide a method capable of effectively processing geomagnetic signals of different qualities, which method can automatically adjust the processing strategy according to the signal characteristics, including intelligent noise management and optimized segmentation processing, thereby improving the accuracy of subsequent feature extraction and classification.

[0009] Another object of the present invention is to provide a progressive deep learning architecture, which effectively improves the model training effect and generalization ability through three-subset division and progressive training strategy, and realizes high-quality enhancement of geomagnetic signal characteristics.

[0010] Another object of the present invention is to provide a verification mechanism based on feature distribution, which can automatically evaluate the quality of features and determine whether further enhancement is needed, thereby ensuring the reliability and stability of the processing results.

[0011] The present invention proposes a deep learning enhancement method for geomagnetic signal data features, including: The acquisition step includes: collecting geomagnetic field signals; The preprocessing step includes: preprocessing the geomagnetic field signal to obtain a time series signal with a variable data amount; The feature extraction step includes: extracting data features from the time series signal to obtain data feature extraction results of geomagnetic signal features; The enhancement step includes: performing deep learning processing on the data feature extraction result of the geomagnetic signal feature to obtain a data enhancement result of the geomagnetic field feature; The pre-processing step specifically includes: The geomagnetic field signal is sequentially subjected to sliding average calculation, normalization and detrending processing, wherein the sliding average calculation is used to reduce high-frequency noise in the geomagnetic field signal, the normalization is used to convert the geomagnetic field signal data into a standard normal distribution, and the detrending is used to eliminate the linear trend in the geomagnetic field signal; The processed geomagnetic field signal is subjected to standardized denoising data processing, and whether it is necessary to increase noise and interference is determined according to the noise and interference intensity; The processed geomagnetic field signal is processed into non-overlapping segments; The noise and interference intensity is obtained by calculating the variance of the geomagnetic field signal; when the variance is less than 0.01, it is determined that the noise and interference do not need to be increased; when the variance is greater than or equal to 0.01, it is determined that the noise and interference need to be increased.

[0012] Preferably, the non-overlapping segmentation process comprises: The variance value of the geomagnetic field signal after statistical processing; Calculate the number of valid segments of the geomagnetic field signal with a variance value less than 0.01; When the number of valid segments is greater than or equal to a threshold, performing non-overlapping segment processing on the geomagnetic field signal in an equal time segmentation manner; When the number of valid segments is less than a threshold, the geomagnetic field signal is processed in non-overlapping segments by adopting an equal variance segmentation method.

[0013] Preferably, the data volume of the time series signal obtained in the preprocessing step includes at least one of 160000, 10000 and 1600; wherein: When the data volume is 160,000, it contains a time series signal consisting of 2048 consecutive geomagnetic signal data points as a group; When the data volume is 10,000, it contains a time series signal consisting of 250 consecutive geomagnetic signal data points as a group; When the data volume is 1600, it contains a time series signal with 100 consecutive geomagnetic signal data points as a group.

[0014] Preferably, the data feature extraction in the feature extraction step adopts a random forest algorithm or a deep learning algorithm, and determines whether the timing signal needs to add noise and interference based on the processed feature vector or data distribution histogram.

[0015] Preferably, the enhancing step specifically comprises: Divide the data feature extraction results into three subsets, wherein the first subset accounts for 15%, the second subset accounts for 25%, and the remaining data subset is the third subset accounting for 60%; Inputting the second subset into a deep learning neural network for nonlinear classification to obtain a vector of geomagnetic signal features of the second subset; Inputting the first subset into the trained deep learning neural network for nonlinear classification; The third subset is input into a further trained deep learning neural network for nonlinear classification.

[0016] Preferably, the method further comprises: Calculating distribution histograms of the geomagnetic signal features of the three partial subsets after vector classification; Comparing the distribution histogram with the data feature extraction result in the feature extraction step; When the standard deviation between any ratio and the data feature extraction result is greater than or equal to 10%, noise and interference are added to all the data feature extraction results, and a deep learning algorithm is used for data enhancement; When the standard deviation between all ratios and the data feature extraction result is less than 10%, it is determined that no noise or interference needs to be added.

[0017] Preferably, the deep learning neural network performs nonlinear classification including: Classifying the vectors of the geomagnetic signal features into two-class, three-class or multi-class categories; Using deep learning DNN processing to obtain a trained deep learning classification model; The newly input vector is classified using the trained deep learning classification model.

[0018] Preferably, the method further comprises: The data enhancement result of the geomagnetic field characteristics is input into a fully convolutional neural network, wherein the fully convolutional neural network comprises: A 1×1×1 convolution kernel is used to extract data channel features and enhance channel dimension features. A Sigmoid activation layer is added after the convolution layer. Perform rotation transformation in the two-dimensional plane and use maximum pooling for two-dimensional dimensionality reduction; Batch normalize the data and set up the feature mapping layer; Downsampling and high-density feature enhancement are performed in the target area, and classification is performed through a fully connected layer.

[0019] The deep learning enhancement method of geomagnetic signal data features provided by the present invention has the following beneficial effects: 1. Improve the ability to represent geomagnetic signal features: Through deep learning enhancement, the ability to represent geomagnetic signal features has been significantly improved, making the key features in the signal more obvious and facilitating subsequent identification and classification; 2. Enhanced anti-noise interference capability: The adaptive noise management mechanism makes the method highly resistant to environmental interference and can effectively extract useful signals in complex environments; 3. Improve classification accuracy: The progressive training method and algorithm fusion improve classification accuracy, making the classification results based on enhanced features more reliable; 4. Achieve efficient dimensionality reduction: The dedicated network architecture effectively reduces the dimension and computational complexity while retaining important signal information; 5. Possess adaptive processing capabilities: The method can automatically adjust the processing strategy according to the signal characteristics and adapt to signals of different quality and characteristics; 6. Enhanced generalization ability: The overall method has good generalization ability for geomagnetic signals of various characteristics and can be applied to different scenarios; 7. Improve signal-to-noise ratio: Preprocessing and enhancement techniques improve the signal-to-noise ratio of the signal, making the useful signal more prominent; 8. Optimize computing efficiency: The architecture design takes into account the computing efficiency of geomagnetic signal processing, reducing computing resource requirements while ensuring processing results. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of the process of the deep learning enhancement method of geomagnetic signal data features of the present invention; Figure 2 Detailed schematic diagram of the pretreatment step of the present invention; Figure 3 It is a schematic diagram of the process flow of non-overlapping segment processing of the present invention; Figure 4 It is a flow chart of the data feature extraction step of the present invention; Figure 5 A schematic diagram of three subset division and progressive learning of the present invention; Figure 6 Schematic diagram of the fully convolutional neural network architecture of the present invention; Figure 7 It is a schematic diagram of the fusion calculation of the RBF neural network of the present invention; Figure 8 The figure is a comparison diagram of the processing effect of the method of the present invention applied to typical geomagnetic signals. DETAILED DESCRIPTION

[0021] Please refer to the attached Figure 1-8 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, the present invention provides a deep learning enhancement method for geomagnetic signal data features, the method comprising the following steps: 1. Acquisition step S1: collecting geomagnetic field signals; 2. Preprocessing step S2: preprocessing the geomagnetic field signal to obtain a time series signal with a variable amount of data; 3. Feature extraction step S3: extracting data features from the time series signal to obtain data feature extraction results of geomagnetic signal features; 4. Enhancement step S4: Perform deep learning processing on the data feature extraction results of the geomagnetic signal features to obtain data enhancement results of the geomagnetic field features.

[0024] In the acquisition step S1, the present invention uses a professional geomagnetic sensor to collect geomagnetic field signals. Preferably, a high-precision three-axis magnetometer is used for acquisition to obtain geomagnetic field signal data in three dimensions of x, y, and z. The acquisition frequency is preferably 10-100 Hz to ensure that subtle changes in the geomagnetic field are captured. In a preferred embodiment, the acquisition frequency is set to 50 Hz, which can avoid generating too much redundant data while ensuring signal integrity.

[0025] In the preprocessing step S2, the geomagnetic field signal collected in step S1 is preprocessed, and the geomagnetic field signal is preprocessed to obtain a time series signal with a variable data amount. Figure 2 As shown, the preprocessing step specifically includes: Step S201: performing sliding average calculation, normalization and detrending processing on the geomagnetic field signal collected in step S1 in sequence.

[0026] The sliding average calculation is used to reduce the high-frequency noise in the collected geomagnetic field signal. Preferably, the sliding average calculation adopts the following formula: , in, is the original geomagnetic field signal value, is the signal value after sliding average, is the sliding window size, Indicates the index of the signal point currently being processed, represents the index of each sampling point in the sliding window. In one embodiment of the present invention, The value is 5. This window size can effectively smooth short-term noise without excessive loss of signal details. The normalization adopts a standardized formula to convert the geomagnetic field signal data collected in step S1 into a standard normal distribution. The standardized formula is as follows: , in, is the signal value after sliding average, is the signal mean, is the signal standard deviation, is the signal value after standardization. Standardization makes signals of different amplitudes comparable, which is convenient for subsequent processing. The detrending method uses moving average filtering to remove the linear trend in the geomagnetic field signal collected in step S1. The detrending method uses the following formula: , in, is the normalized signal value, is the long window moving average, is the signal value after detrending. In a preferred embodiment, the long window size is 10% of the signal length. This setting can effectively capture the long-term trend of the signal without affecting the local characteristics.

[0027] Step S202: Standardize and de-noise the geomagnetic field signal after the sliding average calculation, normalization and detrending in step S201. Determine whether the geomagnetic field signal after the sliding average calculation, normalization and detrending in step S201 needs to add noise and interference according to the noise and interference intensity. When the result of the determination is that the noise and interference need to be added, the noise and interference are added to the geomagnetic field signal after the sliding average calculation, normalization and detrending in step S201; when the result of the determination is that the noise and interference do not need to be added, go to step S3.

[0028] The noise and interference intensity is obtained by calculating the variance of the geomagnetic field signal. When judging whether the geomagnetic field signal after the sliding average calculation, normalization and detrending in step S2 needs to add noise and interference, if the variance is less than 0.01, it is judged that the geomagnetic field signal after the sliding average calculation, normalization and detrending in step S2 does not need to add noise and interference; when the variance is greater than or equal to 0.01, it is judged that noise and interference need to be added.

[0029] The present invention uses a variance threshold of 0.01 to judge the signal quality. The selection of this threshold is based on a large amount of experimental data analysis. When the variance is less than 0.01, it indicates that the signal fluctuation is small and the features are not obvious enough. At this time, there is no need to add noise and interference; when the variance is greater than or equal to 0.01, it indicates that the signal fluctuation is large and the features may be masked by noise. It is necessary to add appropriate noise and interference for data enhancement to improve the robustness of the model.

[0030] Step S203: performing non-overlapping segment processing on the geomagnetic field signal obtained in step S202.

[0031] Combination Figure 3 As shown, the non-overlapping segmentation process includes: Step S2031: Counting the variance value of the geomagnetic field signal obtained by processing in step S202; Step S2032: Calculate the number of valid segments of the geomagnetic field signal with a variance value less than 0.01. When the valid segment number is greater than or equal to the threshold, the geomagnetic field signal is processed in non-overlapping segments using equal time segmentation; when the valid segment number is less than the threshold, the geomagnetic field signal is processed in non-overlapping segments using equal variance segmentation.

[0032] Preferably, the threshold is set to 30% of the total number of segments. For example, if the signal is divided into 100 segments, the threshold is 30. The selection of this threshold is based on statistical analysis to ensure that the segmentation process can adapt to different types of geomagnetic signals. When the number of valid segments accounts for a high proportion, the signal is relatively stable as a whole, and it is suitable to adopt equal time segmentation; when the number of valid segments accounts for a low proportion, the signal fluctuates greatly, and it is suitable to adopt equal variance value segmentation to better capture signal characteristics.

[0033] Step S2033: performing non-overlapping segmentation processing on the geomagnetic field signal obtained in step S202 to obtain a time-series segmented signal of the geomagnetic signal.

[0034] The data volume of the time series signal obtained in the preprocessing step includes at least one of 160000, 10000 and 1600. Specifically including: 1. When the data volume is 160000, it contains a time series signal consisting of 2048 consecutive geomagnetic signal data points as a group; 2. When the data volume is 10,000, it contains a time series signal consisting of 250 consecutive geomagnetic signal data points as a group; 3.1600 data volume, a time series signal containing 100 consecutive geomagnetic signal data points as a group.

[0035] The present invention designs three different time series signals with different data volumes, which are suitable for different application scenarios and computing resource conditions. The data volume of 160,000 is suitable for scenarios that require high-precision analysis and can capture subtle changes in signals; the data volume of 10,000 is suitable for general application scenarios and strikes a balance between accuracy and computing efficiency; the data volume of 1,600 is suitable for scenarios with limited resources or requiring fast processing, and although the accuracy is slightly lower, the processing speed is faster.

[0036] In the feature extraction step S3, Figure 4 As shown, data feature extraction is performed on the time series signal of step S2, and data feature extraction is performed on the time series signal using a data feature extraction algorithm to obtain a data feature extraction result of the geomagnetic signal feature.

[0037] The data feature extraction in the feature extraction step adopts a random forest algorithm or a deep learning algorithm, and determines whether the timing signal needs to add noise and interference based on the processed feature vector or data distribution histogram.

[0038] In a preferred embodiment of the present invention, the implementation of the random forest algorithm includes the following steps: 1. Randomly select n samples from the original training set through the bootstrap method to build a decision tree; 2. Randomly select m features at each node and select the best feature for splitting; 3. Each tree grows to its maximum without pruning; 4. Repeat the above steps to generate multiple decision trees to form a random forest; 5. For new time series signals, pass them to each decision tree for prediction, and obtain the final prediction result by voting or averaging.

[0039] Preferably, the number of decision trees is set to 100, and the feature selection number m is set to the square root of the total number of features. These parameter settings can avoid overfitting while ensuring model performance.

[0040] When using deep learning algorithms for feature extraction, it is preferred to use a convolutional neural network (CNN) structure. The basic structure of a CNN includes: 1. Input layer: receives the preprocessed timing signal; 2. Convolution layer: multiple convolution kernels are used to extract signal features; 3. Pooling layer: reduces feature dimensions and improves model robustness; 4. Fully connected layer: maps features to feature space; 5. Output layer: output the extracted feature vector.

[0041] In the CNN structure, the convolution kernel size is preferably 3, 5, or 7, and the number of convolution layers is preferably 3-5. These settings can capture signal features of different scales.

[0042] After processing by random forest algorithm or deep learning algorithm, feature vector or data distribution histogram will be obtained. The present invention judges whether the time series signal needs to add noise and interference based on these results. The judgment criteria are as follows: 1. When the feature distribution is too concentrated (the entropy value is less than 0.5), it indicates that the features are not rich enough and noise and interference need to be added; 2. When the distinction between features is insufficient (the sum of the main feature weights is less than 0.7), it is necessary to add noise and interference; 3. When the dimension validity of the feature vector is insufficient (the effective dimension ratio is less than 0.6), it is necessary to add noise and interference.

[0043] The thresholds of these judgment criteria are determined based on a large number of experiments and data analysis, and can effectively evaluate the quality of feature extraction and guide subsequent processing.

[0044] In the enhancement step S4, combined with Figure 5 As shown, the enhancement step specifically includes: Step S401: Divide the data feature extraction results of the geomagnetic signal features obtained by the data feature extraction in step S3 into three subsets, wherein the first subset accounts for 15%, the second subset accounts for 25%, and the remaining data subset is used as the third subset and accounts for 60%; This three-subset division ratio (15% to 25% to 60%) is one of the innovative points of this invention. After a large number of experiments, it has been verified that this ratio allocation can improve the generalization ability of the model while ensuring the training effect. The first subset (15%) is used for fine tuning, the second subset (25%) is used for initial training, and the third subset (60%) is used for global optimization. The three subsets work together to achieve the optimization of model performance.

[0045] Step S402: inputting the second subset into a deep learning neural network for nonlinear classification to obtain a vector of geomagnetic signal features of the second subset; Step S403: inputting the first subset into the trained deep learning neural network for nonlinear classification to obtain a vector of geomagnetic signal features of the first subset; Step S404: inputting the third subset into a further trained deep learning neural network for nonlinear classification to obtain a vector of geomagnetic signal features of the third subset; The method further comprises: Step S405: calculating distribution histograms of the geomagnetic signal features of the three partial subsets after vector classification; Step S406: comparing the distribution histogram with the data feature extraction result in the feature extraction step; Step S407: When the standard deviation between any ratio and the data feature extraction result is greater than or equal to 10%, noise and interference are added to all the data feature extraction results, and a deep learning algorithm is used for data enhancement; when the standard deviation between all ratios and the data feature extraction result is less than 10%, it is determined that noise and interference do not need to be added.

[0046] The present invention uses a 10% standard deviation threshold for judgment, which is the optimal value determined based on a large amount of experimental data. When the standard deviation value is greater than or equal to 10%, it indicates that the feature distribution has changed significantly, and it is necessary to enhance the adaptability of the model by adding noise and interference; when the standard deviation value is less than 10%, it indicates that the feature distribution is relatively stable and no additional enhancement is required.

[0047] The deep learning neural network performs nonlinear classification including: Step S4021: classifying the vectors of the geomagnetic signal features into two-classification, three-classification or multiple-classification processing; Step S4022: using deep learning DNN processing to obtain a trained deep learning classification model; Step S4023: Classify the newly input vector using the trained deep learning classification model.

[0048] In a preferred embodiment of the present invention, the structure of the DNN model is designed as follows: 1. Input layer: the number of nodes is equal to the feature vector dimension; 2. Hidden layer: 3-5 hidden layers are used, and the number of nodes in each layer is 128, 64, 32, 16, and 8 respectively (gradually decreasing from the input layer to the output layer); 3. Output layer: the number of nodes is equal to the number of classification categories; 4. Activation function: The hidden layer uses the ReLU activation function, and the output layer uses the Softmax activation function; 5. Optimizer: Adam optimizer, the initial learning rate is set to 0.001, and the learning rate decay strategy is adopted; 6. Loss function: cross entropy loss function.

[0049] The training process of the DNN model is as follows: 1. Use the second subset (25%) for initial training with 100 iterations; 2. Use the first subset (15%) for fine-tuning with 50 iterations; 3. Use the third subset (60%) for global optimization with 200 iterations.

[0050] This progressive training strategy can give full play to the advantages of three-subset division, gradually optimize model performance, and improve feature representation capabilities.

[0051] The method further includes: inputting the data enhancement result of the geomagnetic field characteristics into a fully convolutional neural network, wherein the fully convolutional neural network includes: 1. Use 1×1×1 convolution kernel to extract data channel features, enhance channel dimension features, and add Sigmoid activation layer after convolution layer; 2. Perform rotation transformation in the two-dimensional plane and use maximum pooling for two-dimensional dimensionality reduction; 3. Perform batch normalization on the data and set the feature mapping layer; 4. Downsampling and high-density feature enhancement are performed in the target area, and classification is performed through the fully connected layer.

[0052] like Figure 6 As shown, the fully convolutional neural network structure designed in the present invention is optimized for the characteristics of geomagnetic signals and has the following advantages: 1.1×1×1 convolution kernel design: This design can effectively extract the relationship between data channels, enhance channel dimension features, and is suitable for processing multi-dimensional geomagnetic signals; 2. Sigmoid activation layer: Adding a Sigmoid activation layer after the convolution layer can highlight important features, suppress minor features, and improve the feature representation ability of the network; 3. Two-dimensional rotation transformation: By performing rotation transformation in a two-dimensional plane, the network is made invariant to the direction change of the signal, thus enhancing the robustness of the model; 4. Maximum pooling dimensionality reduction: Use maximum pooling to perform two-dimensional dimensionality reduction operations, reducing computational complexity while retaining important features; 5. Batch normalization: Batch normalization of data to accelerate network convergence and improve training stability; 6. Feature mapping layer design: Set the feature mapping layer to perform feature mapping on the target area and highlight key features; 7. Downsampling and high-density feature enhancement: Downsampling and high-density feature enhancement are performed in the target area to improve the accuracy and richness of feature representation; 8. Fully connected layer classification: The final classification is performed through the fully connected layer to achieve high-precision feature classification.

[0053] The mathematical expression of the fully convolutional neural network is as follows: For the input feature map X, the output feature map of the lth layer The calculation formula is , in, For the The convolution kernel parameters of the layer, For the The output feature map of the layer is used as the input of the first layer. is the bias term, represents the convolution operation, is the activation function. For a 1×1×1 convolution operation, its mathematical expression is: , in, represents the spatial position of the feature map, c represents the channel index, and C is the number of input channels. For the Sigmoid activation function, its mathematical expression is: , For the batch normalization operation, its mathematical expression is: , , in, and are the mean and variance within the batch, respectively. Avoid division by zero for small constants, and are learnable parameters.

[0054] like Figure 7 As shown, the present invention also provides a fusion calculation method based on RBF neural network, which is used to further improve the accuracy of feature classification. The method includes the following steps: 1. Use the data enhancement results as the input of the RBF neural network; 2. Nonlinear transformation through radial basis function; 3. Combine the probability value and the classification result to get the final classification output.

[0055] The structure of RBF neural network includes: input layer, hidden layer and output layer. A transition layer is added between the input layer and the hidden layer to optimize the calculation process.

[0056] The mathematical expression of RBF neural network is as follows: For the input vector X, the output of the hidden layer node is: , in, For the The centers of hidden layer nodes, For extended parameters.

[0057] The calculation formula of the output layer is: , in, is the weight, For bias.

[0058] In a preferred embodiment of the present invention, the parameters of the RBF neural network are set as follows: 1. Number of hidden layer nodes: 1.5 times the input dimension to ensure sufficient expressive power; 2. Center point selection: Use K-means clustering algorithm to determine the center point location; 3. Extension parameters: determined by the average distance method, ,in is the maximum distance, m is the number of hidden layer nodes; 4. Weight initialization: use He initialization method.

[0059] The training of RBF neural network adopts a two-stage strategy: 1. The first stage: determine the hidden layer parameters (center point and expansion parameters); 2. The second stage: train the output layer weights and use the gradient descent method to minimize the mean square error.

[0060] Through the fusion calculation of RBF neural network, the probability value can be organically combined with the classification result to improve the accuracy and stability of classification. Experiments show that compared with the simple use of DNN classification, RBF fusion calculation can increase the classification accuracy by 3-5 percentage points.

[0061] The present invention also provides a processing method for multi-dimensional geomagnetic signals, which can fully utilize the multi-dimensional information of geomagnetic signals and improve the accuracy of feature extraction and classification.

[0062] In this embodiment, the geomagnetic signal includes the following dimensions: 1. x-dimensional information: the component of the geomagnetic field in the x-axis direction; 2. Y-dimension information: the component of the geomagnetic field in the y-axis direction; 3. z-dimension information: the component of the geomagnetic field in the z-axis direction; 4. xy dimension information: composite information of x and y dimensions; 5.yz dimension information: composite information of y and z dimensions; 6. zx dimension information: composite information of z and x dimensions.

[0063] The processing methods of multi-dimensional information include the following three modes: 1. Three-dimensional data mode: collect information in three dimensions: x, y, and z simultaneously; 2. Composite dimension mode: collect information of three composite dimensions: xy, yz, and zx; 3. Hybrid mode: collect multiple dimensions of the above dimensions at the same time.

[0064] For multi-dimensional data, the present invention adopts the following feature calculation method: 1. Mean calculation: , in, represents the d-th dimension value of the ith data point, represents the mean of the dth dimension, and N represents the total number of data points.

[0065] 2. Standard deviation calculation: , in, Represents the standard deviation of the dth dimension.

[0066] 3. Information entropy calculation: , in, represents the information entropy of the d-th dimension, Represents the probability distribution of the dth dimension. 4. Eigenvalue calculation: , in, Represents the eigenvalue of the d-th dimension.

[0067] Through the above method, rich feature information can be extracted from multi-dimensional geomagnetic signals, providing more comprehensive input for subsequent deep learning processing and further improving the effects of feature enhancement and classification.

[0068] The method of the present invention has been verified in actual geomagnetic signal processing. Figure 8 The comparison of application effects is shown. The verification experiment uses geomagnetic signals collected in three different environments: 1. Low noise environment: indoor, laboratory environment with little electromagnetic interference; 2. Medium noise environment: urban streets, environments with certain electromagnetic interference; 3. High noise environment: industrial areas and environments with severe electromagnetic interference.

[0069] The verification results show that the method of the present invention exhibits excellent performance in three environments: 1. Low-noise environment: The signal-to-noise ratio is improved by 40%, and the feature classification accuracy reaches 98.5%; 2. Medium noise environment: The signal-to-noise ratio is improved by 55%, and the feature classification accuracy reaches 95.2%; 3. High noise environment: The signal-to-noise ratio is improved by 70%, and the feature classification accuracy reaches 92.1%.

[0070] Compared with traditional methods, the present invention has significant advantages in processing geomagnetic signals in high-noise environments, and the feature classification accuracy is improved by 15-20 percentage points, which shows that the method of the present invention has strong noise resistance and adaptability.

[0071] In addition, the method of the present invention also shows good stability under different data levels: 1.160,000 data: processing time is about 3 seconds, and classification accuracy is 98.0%; 2.10,000 data sets: processing time is about 0.8 seconds, and classification accuracy is 96.5%; 3.1600 data volumes: processing time is about 0.2 seconds, and classification accuracy is 94.0%.

[0072] This shows that the method of the present invention can adapt to different data scales and processing speed requirements while ensuring high accuracy.

[0073] The present invention provides a deep learning enhancement method for geomagnetic signal data features. The method achieves high-quality enhancement of geomagnetic signal features through adaptive preprocessing, multi-mode segmentation processing, three-subset progressive learning and feature verification mechanism. Compared with traditional methods, the present invention has the following significant advantages: 1. Adaptive processing capability: Able to automatically adjust processing strategies according to signal characteristics to adapt to signals of different quality and characteristics; 2. Efficient feature extraction: Improve the efficiency and quality of feature extraction through multiple algorithms and optimized structures; 3. Progressive learning architecture: adopts three-subset partitioning and progressive training strategy to improve model performance and generalization ability; 4. Verification feedback mechanism: Ensure the reliability of processing results through feature distribution comparison and verification feedback; 5. Multi-dimensional collaborative processing: Able to process multi-dimensional geomagnetic signals to improve the comprehensiveness and accuracy of feature representation.

[0074] The method of the present invention has broad application prospects in many fields such as geomagnetic navigation, geological exploration, space positioning, military reconnaissance, etc. It can provide higher quality feature representation and classification results for these applications and promote the development and application of related technologies.

[0075] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A deep learning enhancement method for geomagnetic signal data features, characterized in that: include: The acquisition step includes: collecting geomagnetic field signals; The preprocessing step includes: preprocessing the geomagnetic field signal to obtain a time series signal with a variable data amount; The feature extraction step includes: extracting data features from the time series signal to obtain data feature extraction results of geomagnetic signal features; The enhancement step includes: performing deep learning processing on the data feature extraction result of the geomagnetic signal feature to obtain a data enhancement result of the geomagnetic field feature; The pre-processing step specifically includes: The geomagnetic field signal is sequentially subjected to sliding average calculation, normalization and detrending processing, wherein the sliding average calculation is used to reduce high-frequency noise in the geomagnetic field signal, the normalization is used to convert the geomagnetic field signal data into a standard normal distribution, and the detrending is used to eliminate the linear trend in the geomagnetic field signal; The processed geomagnetic field signal is subjected to standardized denoising data processing, and whether it is necessary to increase noise and interference is determined according to the noise and interference intensity; The processed geomagnetic field signal is processed into non-overlapping segments; The noise and interference intensity is obtained by calculating the variance of the geomagnetic field signal; when the variance is less than 0.01, it is determined that the noise and interference do not need to be increased; when the variance is greater than or equal to 0.01, it is determined that the noise and interference need to be increased.

2. The deep learning enhancement method for geomagnetic signal data features according to claim 1, characterized in that: The non-overlapping segmentation process includes: The variance value of the geomagnetic field signal after statistical processing; Calculate the number of valid segments of the geomagnetic field signal with a variance value less than 0.01; When the number of valid segments is greater than or equal to a threshold, performing non-overlapping segment processing on the geomagnetic field signal in an equal time segmentation manner; When the number of valid segments is less than a threshold, the geomagnetic field signal is processed in non-overlapping segments by adopting an equal variance segmentation method.

3. The deep learning enhancement method for geomagnetic signal data features according to claim 1, characterized in that: The data volume of the time series signal obtained in the preprocessing step includes at least one of 160000, 10000 and 1600; wherein: When the data volume is 160,000, it contains a time series signal consisting of 2048 consecutive geomagnetic signal data points as a group; When the data volume is 10,000, it contains a time series signal consisting of 250 consecutive geomagnetic signal data points as a group; When the data volume is 1600, it contains a time series signal with 100 consecutive geomagnetic signal data points as a group.

4. The deep learning enhancement method for geomagnetic signal data features according to claim 1, characterized in that: The data feature extraction in the feature extraction step adopts a random forest algorithm or a deep learning algorithm, and determines whether the timing signal needs to add noise and interference based on the processed feature vector or data distribution histogram.

5. The deep learning enhancement method for geomagnetic signal data features according to claim 1, characterized in that: The enhancement step specifically includes: Divide the data feature extraction results into three subsets, wherein the first subset accounts for 15%, the second subset accounts for 25%, and the remaining data subset is the third subset accounting for 60%; Inputting the second subset into a deep learning neural network for nonlinear classification to obtain a vector of geomagnetic signal features of the second subset; Inputting the first subset into the trained deep learning neural network for nonlinear classification; The third subset is input into a further trained deep learning neural network for nonlinear classification.

6. The deep learning enhancement method for geomagnetic signal data features according to claim 5, characterized in that: The method further comprises: Calculating distribution histograms of the geomagnetic signal features of the three partial subsets after vector classification; Comparing the distribution histogram with the data feature extraction result in the feature extraction step; When the standard deviation between any ratio and the data feature extraction result is greater than or equal to 10%, noise and interference are added to all the data feature extraction results, and a deep learning algorithm is used for data enhancement; When the standard deviation between all ratios and the data feature extraction result is less than 10%, it is determined that no noise or interference needs to be added.

7. The deep learning enhancement method for geomagnetic signal data features according to claim 5, characterized in that: The deep learning neural network performs nonlinear classification including: Classifying the vectors of the geomagnetic signal features into two-class, three-class or multi-class categories; Using deep learning DNN processing to obtain a trained deep learning classification model; The newly input vector is classified using the trained deep learning classification model.

8. The deep learning enhancement method for geomagnetic signal data features according to claim 1, characterized in that: The method further comprises: The data enhancement result of the geomagnetic field characteristics is input into a fully convolutional neural network, wherein the fully convolutional neural network comprises: A 1×1×1 convolution kernel is used to extract data channel features and enhance channel dimension features. A Sigmoid activation layer is added after the convolution layer. Perform rotation transformation in the two-dimensional plane and use maximum pooling for two-dimensional dimensionality reduction; Batch normalize the data and set up the feature mapping layer; Downsampling and high-density feature enhancement are performed in the target area, and classification is performed through a fully connected layer.

Images