Magnetic anomaly signal detection method based on orthogonal basis function physically consistent network

Abstract

The application discloses a magnetic anomaly signal detection method based on a physical consistency network of an orthogonal basis function, and relates to the field of magnetic anomaly detection. The method sets a fixed physical projection layer in the neural network, which is determined by an orthogonal basis function target physical subspace matrix; a signal fragment to be detected is input into the network, and a target existence probability and a target anomaly component are output; the target anomaly component is physically decomposed by using the physical projection layer to obtain a physically consistent component and a physical residual component, and a physical consistency index is calculated according to the physically consistent component and the physical residual component; based on the physical consistency index and the target existence probability, a final detection confidence is generated in a conflict gating mode, and the confidence is automatically inhibited when the confidence is inconsistent. The OBF physical subspace is used as a structural constraint in the network, the physical interpretability is retained, the blind confidence of the network is inhibited through the conflict gating mechanism, and the detection robustness is improved under the condition of low signal-to-noise ratio and complex interference.

Description

Technical Field

[0001] This invention relates to the field of magnetic anomaly detection, and more specifically to a method for detecting magnetic anomaly signals based on orthogonal basis function physical consistency networks. Background Technology

[0002] Magnetic anomaly detection utilizes local magnetic field disturbances caused by magnetic or conductive metal targets to achieve target detection. In scenarios such as underwater buried targets, unexploded ordnance detection, pipeline detection, and identification of moving magnetic targets, the target magnetic anomaly signal is usually weak and easily affected by gradual changes in the geomagnetic background, sensor noise, platform motion interference, environmental electromagnetic noise, and non-target anomaly disturbances.

[0003] Traditional threshold detection and constant false alarm rate (CFAR) detection methods rely on statistical distribution assumptions. When the background exhibits non-stationary characteristics, they are difficult to adapt stably, resulting in a significant decrease in detection performance.

[0004] Orthogonal basis function (OBF) detection methods do not rely on statistical distribution assumptions and utilize the projection of the target magnetic anomaly's physical subspace for detection, exhibiting better adaptability to non-stationary backgrounds. Specifically, the OBF detection method constructs orthogonal basis functions of the target anomaly waveform based on a magnetic dipole model, projects the signal to be detected onto the target magnetic anomaly's physical subspace, and uses the projection energy for detection decision. This method has clear physical meaning and good interpretability. However, the OBF detection method is sensitive to target attitude, burial depth, and sensor relative motion parameters. Under conditions of low signal-to-noise ratio, attitude changes, and complex interference, the robustness of detection relying solely on projection of a fixed basis function subspace is insufficient.

[0005] On the other hand, ordinary neural networks can automatically extract complex nonlinear features from signals through end-to-end learning, demonstrating excellent performance in pattern recognition tasks. However, when ordinary neural networks are directly applied to the detection of magnetic anomaly signals, their output lacks constraints on the physical morphology of magnetic anomalies, making it easy to misjudge random spikes, plateaus, or gradual changes in the background as target anomalies, resulting in a high false detection rate.

[0006] Therefore, while traditional OBF detection methods possess physical interpretability, they lack robustness under low signal-to-noise ratio and complex interference conditions. Although ordinary neural networks have strong feature learning capabilities, their outputs lack physical constraints, making them prone to false detections. How to improve detection robustness and reduce false detection rates while maintaining physical interpretability remains a pressing issue. Summary of the Invention

[0007] This invention proposes a magnetic anomaly signal detection method based on orthogonal basis function physical consistency network. Its purpose is to solve the problem that existing magnetic anomaly signal detection methods are difficult to maintain physical interpretability while ensuring detection robustness. The method aims to improve detection robustness and reduce false detection rate under low signal-to-noise ratio and complex interference conditions while maintaining physical interpretability.

[0008] The technical solution of this invention is as follows:

[0009] A method for detecting magnetic anomaly signals based on orthogonal basis function physical consistency networks includes:

[0010] Construct a neural network that includes a shared temporal encoder, a target probability output branch, and a target anomaly component output branch;

[0011] A fixed, non-trainable physical projection layer is set inside the neural network, and the physical projection layer is determined by the orthogonal basis function objective physical subspace matrix;

[0012] The signal segment to be detected is obtained based on the magnetic field data collected by the magnetic sensor. The signal segment to be detected is input into the neural network. After the shared timing encoder extracts the timing features, the target existence probability is output by the target probability output branch, and the target abnormal component is output by the target abnormal component output branch.

[0013] The target anomaly component is physically decomposed using the physical projection layer to obtain physically consistent components that can be explained by the orthogonal basis function target physical subspace and physically residual components that cannot be explained by the orthogonal basis function target physical subspace. A physical consistency index is calculated based on the physically consistent components and the physically residual components. The physical consistency index includes at least the projection energy ratio and angular consistency. The projection energy ratio represents the proportion of the energy of the physically consistent component in the energy of the target anomaly component, and the angular consistency represents the degree of directional similarity between the target anomaly component and the physically consistent component.

[0014] Based on the probability of the existence of the physical consistency index and the target, a final detection confidence score is generated using a conflict gating method. When the physical consistency index indicates that the abnormal component of the target is inconsistent with the physical subspace of the orthogonal basis function target, suppression is applied to the final detection confidence score.

[0015] As a further improvement to the magnetic anomaly signal detection method based on orthogonal basis function physical consistency networks, the orthogonal basis function target physical subspace matrix is ​​constructed as follows: Discrete sampling is performed on the target and survey line parameter spaces to obtain multiple sets of parameter combinations; each set of parameter combinations is substituted into the magnetic anomaly target response model to generate corresponding candidate anomaly waveforms; all candidate anomaly waveforms are arranged column-wise to form a candidate waveform matrix; singular value decomposition is performed on the candidate waveform matrix, and the number of basis functions is determined according to the cumulative energy threshold. Take the front Using left singular vectors as orthogonal basis functions, an orthogonal basis function objective physical subspace matrix is ​​formed. ;Depend on Determine the physical projection matrix The physical projection layer is through accomplish.

[0016] As a further improvement to the magnetic anomaly signal detection method based on orthogonal basis function physical consistency networks, the physical decomposition is performed in the following manner:

[0017] ,

[0018] ,

[0019] Among them, for the first A signal segment to be detected, The target abnormal component, For the physical consistency component, For the physical residual components, Let be the physical projection matrix.

[0020] As a further improvement to the magnetic anomaly signal detection method based on orthogonal basis function physical consistency networks, for the first... A signal segment to be detected, the projection energy ratio Consistency with the stated angle Calculate using the following formulas respectively:

[0021] ,

[0022] ,

[0023] in, The target abnormal component, For the physical consistency component, Describing the L2 norm, Represents the vector dot product. To prevent division by zero of positive numbers.

[0024] As a further improvement to the magnetic anomaly signal detection method based on orthogonal basis function physical consistency networks, for the first... A signal segment to be detected, the final detection confidence level Calculate using the following formula:

[0025] ,

[0026] in, The probability that the target exists. The projection energy ratio, OBF physical consistency score, This is the conflict penalty coefficient. This represents an exponential function.

[0027] As a further improvement to the magnetic anomaly signal detection method based on orthogonal basis function physical consistency networks, for the first... The OBF physical consistency score is used to detect a signal segment. Calculate using the following formula:

[0028] ,

[0029] in, The projection energy ratio, For the sake of the aforementioned angle consistency, For the physical residual components, The target abnormal component, Describing the L2 norm, To prevent division by zero of positive numbers, For the Sigmoid function, , , and These are calibration parameters.

[0030] As a further improvement to the magnetic anomaly signal detection method based on orthogonal basis function physical consistency networks, the total training loss function of the neural network... for:

[0031] ,

[0032] in, The binary cross-entropy loss is the probability of the target's existence. For loss of angular consistency, For the projected energy separation loss, To punish losses in the event of conflict To smooth out the loss, to These are the weighting coefficients.

[0033] As a further improvement to the magnetic anomaly signal detection method based on orthogonal basis function physical consistency networks, the conflict penalty loss... Calculate using the following formula:

[0034] ,

[0035] in, This represents the batch sample size. For the first The probability of the target existing in each sample. For the first The proportion of projected energy for each sample.

[0036] As a further improvement to the magnetic anomaly signal detection method based on orthogonal basis function physical consistency networks, the angle consistency loss... and the projected energy separation loss Calculate using the following formulas respectively:

[0037] ,

[0038] ,

[0039] in, For the batch sample size, For sample labels, target samples Background samples ; For the first The proportion of projected energy for each sample, For the first The consistency of the angles of each sample; and These are the projection energy interval thresholds for the target sample and the background sample, respectively. .

[0040] As a further improvement to the magnetic anomaly signal detection method based on orthogonal basis function physical consistency networks, the smoothing loss... Calculate using the following formula:

[0041] ,

[0042] in, For the batch sample size, For the first The target anomaly component of each sample. It is a first-order or second-order difference operator. This represents the L1 norm.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] (1) This invention sets up a fixed, non-trainable physical projection layer inside the neural network to perform forced physical decomposition on the target anomaly components output by the network, obtaining physical consistency components and physical residual components, and calculating a physical consistency index based on this. Then, the physical consistency index is applied to the target existence probability using a conflict gating method to generate the final detection confidence. This mechanism transforms the OBF physical subspace from a traditional external independent detector into a structural constraint within the network, which not only preserves the physical interpretability of OBF detection but also solves the problem of ordinary neural network output lacking physical constraints and easily generating blindly confident false detections. When the target anomaly components output by the network are inconsistent with the OBF physical subspace, the physical consistency index will automatically suppress the final detection confidence, thereby effectively reducing the false detection rate caused by random spikes, plateaus, and non-target anomalies, and significantly improving detection robustness under low signal-to-noise ratio and complex interference conditions.

[0045] (2) In calculating the final detection confidence, this invention employs an exponential conflict gating mechanism. By introducing an exponential decay term with the projected energy ratio as the independent variable into the confidence formula, an exponential linkage is formed between the target existence probability and the projected energy ratio. When the target existence probability is high but the projected energy ratio is low, the exponential decay term decreases rapidly, effectively suppressing detection results with high confidence but physical inconsistency, thereby significantly distinguishing between real target anomalies and false alarms caused by network overconfidence.

[0046] (3) The present invention uses a joint loss function consisting of angle consistency loss, projection energy separation loss, conflict penalty loss and smoothing loss to train the network.

[0047] Among them, the projection energy separation loss applies different projection energy interval thresholds to the target sample and the background sample, causing the abnormal components of the target sample to gather in the OBF physical subspace while the abnormal components of the background sample are pushed away from the OBF physical subspace. Combined with the angle consistency loss to constrain the target waveform direction and the smoothing loss to suppress random spikes, a three-dimensional constraint framework of direction alignment, energy separation, decision conflict, and waveform continuity is formed, enabling the network to learn feature representations that are both in line with physical laws and have discriminative ability.

[0048] The conflict penalty loss cross-couples the probability of target presence with the proportion of projected energy to achieve conditional physical constraints, meaning that enhanced physical constraints are only applied when the network determines that a target exists. This differs from the traditional physical information network approach where physical loss and classification loss are parallel and physical constraints apply indiscriminately to all samples. This allows the network to learn the inherent connection that high-probability outputs must be supported by physical consistency during the training phase. Attached Figure Description

[0049] Figure 1A schematic diagram illustrating the determination of the physical projection layer and the physical decomposition based on the physical projection layer.

[0050] Figure 2 This is a schematic diagram of the architecture of the neural network OBF-PCNet. Detailed Implementation

[0051] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0052] I. Implementation Plan

[0053] This embodiment provides a method for detecting magnetic anomaly signals based on orthogonal basis function physical consistency networks, which processes magnetic field data collected by a magnetic sensor. Specifically, the magnetic field data collected by the magnetic sensor is acquired, and the magnetic field data is processed by DC removal, detrending, filtering, standardization, and sliding window slicing to obtain the signal segment to be detected.

[0054] This method comprises three main parts: network construction, training, and inference, totaling seven steps. The following is a detailed explanation.

[0055] (I) Network Construction

[0056] Step 1: Construct the OBF target physical subspace and physical projection layer.

[0057] like Figure 1 As shown, based on the magnetic anomaly target response model, the parameter space of the target and the survey line is discretely sampled. Each set of parameters is substituted into the magnetic anomaly target response model to generate the corresponding candidate anomaly waveform. The magnetic anomaly target response model can adopt the magnetic dipole model, and the anomaly component in the observation direction can be expressed as:

[0058] ,

[0059] In the above formula, This is a set of target and survey line parameters, including target burial depth, lateral offset, target magnetic moment direction, survey line velocity, and target scale factor. The target magnetic moment vector; The position vector changes as the sensor moves along the measurement line; The unit vector in the direction of observation; The vacuum permeability; Indicates the sampling time; It is the identity matrix; This represents the L2 norm.

[0060] right Discrete sampling is performed to obtain the parameter set. ,in Each group Substitute into the above formula for magnetic anomaly components, and take samples within the same sampling window. By sampling points, a candidate abnormal waveform with the same length as the signal segment to be detected is obtained. All candidate abnormal waveforms are arranged into a candidate waveform matrix by columns. .

[0061] right Perform singular value decomposition: The number of basis functions is determined according to the cumulative energy threshold. :

[0062] ,

[0063] in The first singular value decomposition result arranged in descending order One singular value; The cumulative energy threshold is preferably between 0.98 and 0.995. (The preceding value is used.) Using 6 to 12 left singular vectors as orthogonal basis functions, the OBF target physical subspace matrix is ​​formed. ,satisfy .

[0064] Depend on Determine the physical projection matrix This matrix will serve as a fixed, non-trainable physical projection layer within the network.

[0065] This approach enables the OBF subspace to cover target response variations caused by different burial depths, attitudes, and offsets, while avoiding excessive basis functions that could lead to background noise entering the target subspace.

[0066] Step 2: Construct the OBF-PCNet network.

[0067] Construct an orthogonal basis function physical consistency network OBF-PCNet, the architecture of which is as follows: Figure 2 As shown. The network includes a shared temporal encoder, a target probability output branch, a target anomaly component output branch, and a physical projection layer implemented by the physical projection matrix described in step 1.

[0068] Shared timing encoder This is used to extract temporal features from input signal segments. Following the input layer, a one-dimensional convolutional layer Conv1D(16, 7), batch normalization, and ReLU activation are sequentially applied. Then, three multi-scale residual blocks are set, with output channels of 16, 32, and 64, and dilation coefficients of 1, 2, and 4, respectively. Each residual block contains two parallel one-dimensional convolutional branches with kernel sizes of 3 and 7. The outputs of the two branches are concatenated along the channel dimension and then fused using a 1×1 convolution. Residual summation occurs after branch concatenation and 1×1 channel fusion. The first residual block has 16 input and 16 output channels, with each branch outputting 8 channels; the residual branches use identity mapping. The second residual block outputs 32 channels from 16, with each branch outputting 16 channels; the residual branches use 1×1 convolutions to align the 16 channels to 32 channels. The third residual block outputs 64 channels from 32, with each branch outputting 32 channels; the residual branches use 1×1 convolutions to align the 32 channels to 64 channels.

[0069] The target probability output branch outputs the target existence probability through global average pooling and a fully connected layer. ,in For the shared timing encoder to the first A signal segment to be detected Extracted temporal features, Let the target probability output function be... This is the Sigmoid function.

[0070] The target anomalous component output branch outputs the target anomalous component through a one-dimensional convolutional decoding layer that maintains a constant length. , With the length of the input signal segment Consistent, This is the function for outputting abnormal components.

[0071] The physical projection layer is set after the target anomaly component output branch and before the conflict gating calculation. It belongs to the post-output physical constraint layer and does not affect the intermediate features of the shared temporal encoder. For example... Figure 1 As shown, the physical projection layer utilizes the physical projection matrix. For target anomaly components Perform physical decomposition:

[0072] ,

[0073] ,

[0074] in For OBF physical consistency components, it represents the portion that can be interpreted by the OBF target physical subspace; For OBF physical residual components, represent the portion that cannot be explained by the OBF target physical subspace.

[0075] Based on the physical decomposition, the following physical consistency index is calculated. For the... Each signal segment, OBF projection energy ratio Defined as the proportion of the energy of the physically consistent component to the energy of the target anomalous component:

[0076] ,

[0077] Angle consistency Defined as the angular cosine similarity between the target anomaly component and the physically consistent component:

[0078] ,

[0079] in Represents the vector dot product. To prevent division by zero of positive numbers.

[0080] Conflict factors Defined as the product of the probability of the target's existence and the degree of physical inconsistency:

[0081] .

[0082] The above physical consistency index and conflict factor will serve both the subsequent joint loss training (step 4) and conflict-gated inference (step 7).

[0083] (II) Training Process

[0084] The training process of OBF-PCNet is described in detail below.

[0085] Step 3: Training sample construction.

[0086] The training samples include target samples and background samples. Target samples are simulated segments generated by the magnetic dipole model and superimposed with measured background noise; background samples are segments from targetless areas, platform motion disturbances, geomagnetic slow variations, power frequency interference, and random noise. The signal-to-noise ratio of the training samples covers -10dB to 20dB, and the ratio of target samples to background samples is preferably 1:1 to 1:2. The dataset is divided into training, validation, and test sets, with the preferred ratios being 70%, 15%, and 15%, respectively.

[0087] Step 4: Design the joint loss function.

[0088] The total loss function for network training is composed of the classification loss and four physical consistency losses:

[0089] ,

[0090] in to These are the weighting coefficients. The definitions of each loss term are as follows.

[0091] Classification loss The cross-entropy loss is a binary classification loss that calculates the probability of the existence of the target.

[0092] Angular consistency loss The network output outliers used to constrain the target samples maintain a high degree of directional consistency with the OBF target physical subspace.

[0093] ,

[0094] in For the batch sample size, For sample labels, target samples Background samples .

[0095] Projected energy separation loss This is used to concentrate the projected energy of the target sample into the OBF target physical subspace, while pushing the projected energy of the background sample away from the OBF target physical subspace.

[0096] ,

[0097] in and Let f(x) be the projection energy interval thresholds for the target sample and the background sample, respectively, and satisfy the following conditions: Preferred , .

[0098] Conflict punishment loss To suppress false detections caused by the mismatch between the high-confidence output of the neural network and the physical consistency of OBF, the probability of target existence is cross-coupled with the projection energy ratio:

[0099] .

[0100] Smoothing loss Used to suppress random spike noise from being learned by the network as target anomalous components:

[0101] ,

[0102] in It is a first-order or second-order difference operator. This represents the L1 norm.

[0103] The loss weights can be determined on the validation set, and their values ​​range as follows: , , , .

[0104] Step 5: Training configuration and training process.

[0105] The network is trained using the Adam optimizer, and the initial learning rate can be set to... The batch size can be 32 or 64, and the maximum number of training epochs can be 200. Training stops when the validation set loss does not decrease for 20 consecutive epochs, or the validation set F1 score does not improve for 20 consecutive epochs, and the parameters of the model with the best performance on the validation set are saved.

[0106] (III) Reasoning Process

[0107] Step 6: Signal acquisition and preprocessing.

[0108] Magnetic field data of the area under test is acquired using a fluxgate sensor to obtain the original magnetic anomaly signal sequence. The original signal is then processed by DC removal, detrending, bandpass filtering, normalization, and sliding window slicing to obtain a sequence of length [length missing]. The signal segment to be detected. In underwater towed or moving survey line scenarios, A value of 512 can be chosen to balance the integrity of the target response and the amount of real-time computation. Each signal segment is denoted as .

[0109] Step 7: Conflict gating confidence calculation and target decision.

[0110] During the detection phase, the signal segment to be detected is... Input the trained OBF-PCNet and obtain the target presence probability through the forward propagation of the network described in step 2. Target anomaly components OBF projection energy ratio and angle consistency .

[0111] according to , OBF physical consistency score is calculated based on the ratio of physical residual energy. :

[0112] ,

[0113] in , , and The calibration parameters were determined using a grid search on the validation set. This is the Sigmoid function.

[0114] The final detection confidence score is generated based on the physical consistency index and the probability of target existence, using a conflict gating method. :

[0115] ,

[0116] in This is the conflict penalty coefficient. This represents an exponential function. The core mechanism of this formula lies in: when the target exists, there is a probability... High and OBF projection energy ratio When the value is high, the conflict penalty item Approaching 1, the final detection confidence level is... and Joint decision; when Higher but At lower levels, the conflict penalty term decays exponentially, and the final confidence level is significantly suppressed, thereby achieving automatic rejection of network outputs that are highly confident but physically inconsistent.

[0117] Final detection confidence level With detection threshold Comparison. When When, it is determined that a magnetic anomaly target exists in the current signal segment; when When this occurs, it is determined to be background or non-target disturbance. Threshold On the validation set, it was determined that the preset false alarm rate was met. Under the given conditions, select the threshold that maximizes the F1 score; preferably, select... .

[0118] For multiple consecutive signal segments that meet the threshold condition, they are merged into the same target anomaly region, and then... The center of the window corresponding to the maximum value is taken as the target center position. The system outputs the target alarm, target center position, target presence probability, OBF physical consistency score, conflict factor, and final detection confidence. Specific Implementation

[0119] The following example, using the detection of magnetic anomalies in underwater buried ferromagnetic targets, illustrates a specific embodiment of the present invention. In this embodiment, the overall execution flow follows steps 1 to 7 as described in the above scheme. It should be noted that the order of the following steps is not a mandatory order during implementation; technicians can adjust them according to the situation, or even execute them in parallel. For example, steps S1 and S3 can be executed in parallel.

[0120] S1: Experimental scenario and data acquisition (corresponding to steps 3 and 6).

[0121] A ferromagnetic target is deployed in the underwater or shallow water test area, located below or near the measurement line. A triaxial fluxgate sensor is used to move at a constant speed along the preset measurement line and collect magnetic field data. The sampling frequency is set to 100Hz, and the moving speed is set to 0.1m / s to 0.2m / s. After acquiring the triaxial magnetic field data, the total field strength anomaly or the single-axis component with the most significant response to the target can be selected as the original magnetic anomaly sequence.

[0122] S2: Signal preprocessing and slicing (corresponding to step 6).

[0123] The original magnetic anomaly sequence was subjected to mean removal, third-order polynomial detrending, and bandpass filtering. The filtering bandwidth was determined based on the measurement line velocity and the target response duration. Subsequently, standardization was performed, and length was used as the standard. The signal segments are divided using a sliding window with a 50% overlap rate. Each segment is labeled as either a target sample or a background sample. Target samples correspond to segments where the survey line passes through the target response area, while background samples correspond to segments far from the target area.

[0124] S3: Construct the OBF orthogonal basis function matrix (corresponding to step 1).

[0125] Candidate magnetic anomaly waveforms are generated based on the magnetic dipole model. The target burial depth is then set. lateral offset Linear velocity measurement The magnetic moment directions are respectively along , , Values ​​are taken from three directions. Candidate waveforms are generated from the above combinations and then uniformly resampled to... Then, perform zero-mean and unity-norm normalization. Form a matrix from all candidate waveforms. ,right Perform singular value decomposition. Set a cumulative energy threshold. Determine the number of basis functions as described in step 1. In this embodiment ,get Thus, the physical projection matrix is ​​determined. .

[0126] S4: Establish the OBF-PCNet network (corresponding to step 2).

[0127] An OBF-PCNet network is constructed. The input is a magnetic anomaly signal segment of length 512. The shared temporal encoder is constructed according to the structure described in step 2: Conv1D(16, 7), batch normalization, ReLU activation, and three multi-scale residual convolutional blocks, with output channels of 16, 32, and 64, and dilation coefficients of 1, 2, and 4, respectively. The target probability output branch is output through global average pooling and a fully connected layer. The target anomaly component branch is decoded using one-dimensional convolution, outputting a length of 512. Set a fixed physical projection layer within the network. ,right Physical decomposition is performed to obtain physically consistent components. and physical residual components And further calculation , and conflict factors .

[0128] S5: Training data construction and network training (corresponding to steps 3, 4, and 5).

[0129] A training dataset was constructed using measured target segments, targetless background segments, and simulated segments generated by a magnetic dipole model and superimposed with measured background noise. The signal-to-noise ratio of the training samples was set to -10dB to 20dB, and the ratio of target samples to background samples was set to 1:1. The proportions of the training set, validation set, and test set were set to 70%, 15%, and 15%, respectively. , , , , , The Adam optimizer is used, with an initial learning rate of... The batch size is 64, and the maximum number of training epochs is 200. Training stops when the validation set loss does not decrease for 20 consecutive epochs or the validation set F1 score does not improve for 20 consecutive epochs, and the parameters of the model with the best performance on the validation set are saved.

[0130] S6: Calculation of conflict gating confidence and determination of threshold (corresponding to step 7).

[0131] Input the validation set into the trained OBF-PCNet and calculate... , , , and final detection confidence Set the conflict penalty coefficient. Calibration parameters , , and The detection threshold is determined using a grid search on the validation set. The validation set is used to determine the threshold that maximizes the F1 score, provided the false alarm rate does not exceed 5%. If the optimal threshold of 0.60 is found in the validation set, the detection phase is fixed. .

[0132] S7: Online detection and target area output (corresponding to step 7).

[0133] During online detection, the real-time magnetic anomaly signal is preprocessed and sliced ​​according to the S2 method, then input into the trained OBF-PCNet to calculate the final detection confidence. .when When, it is determined that a magnetic anomaly target exists in the corresponding segment; when When the condition is met, it is determined to be either background or non-target disturbance. For multiple consecutive segments that meet the threshold condition, they are merged into the same target anomaly region, and then... The center of the window corresponding to the maximum value is taken as the target center position. The system outputs the target alarm, target center position, target presence probability, OBF physical consistency score, conflict factor, and final detection confidence.

[0134] S8: Results Analysis and Interpretation.

[0135] If a certain segment higher and , If the value is also high, it indicates that the neural network detection result is consistent with the OBF target physical subspace, and the alarm reliability is high; if Higher but A lower confidence level indicates a conflict between the network's discrimination result and the physical consistency of the OBF. Ultimately, the confidence level will be reduced by the exponential penalty term, thereby suppressing false alarms caused by random noise, platform disturbances, or non-target anomalies.

[0136] It should be noted that, as will be apparent to those skilled in the art, the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics thereof. The scope of the present invention is defined by the claims rather than the foregoing description.

Claims

1. A method for detecting magnetic anomaly signals based on orthogonal basis function physical consistency networks, comprising: Construct a neural network that includes a shared temporal encoder, a target probability output branch, and a target anomaly component output branch; Its characteristic is that it further includes the following steps: A fixed, non-trainable physical projection layer is set inside the neural network, and the physical projection layer is determined by the orthogonal basis function objective physical subspace matrix; The signal segment to be detected is obtained based on the magnetic field data collected by the magnetic sensor. The signal segment to be detected is input into the neural network. After the shared timing encoder extracts the timing features, the target existence probability is output by the target probability output branch, and the target abnormal component is output by the target abnormal component output branch. The target anomaly component is physically decomposed using the physical projection layer to obtain physically consistent components that can be explained by the orthogonal basis function target physical subspace and physically residual components that cannot be explained by the orthogonal basis function target physical subspace. A physical consistency index is calculated based on the physically consistent components and the physically residual components. The physical consistency index includes at least the projection energy ratio and angular consistency. The projection energy ratio represents the proportion of the energy of the physically consistent component in the energy of the target anomaly component, and the angular consistency represents the degree of directional similarity between the target anomaly component and the physically consistent component. Based on the probability of the existence of the physical consistency index and the target, a final detection confidence score is generated in a conflict gating manner. When the physical consistency index indicates that the abnormal component of the target is inconsistent with the physical subspace of the orthogonal basis function target, suppression is applied to the final detection confidence score. For the A signal segment to be detected, the final detection confidence level Calculate using the following formula: ， in, The probability that the target exists. The projection energy ratio, OBF physical consistency score, This is the conflict penalty coefficient. Represents an exponential function; For the The OBF physical consistency score is used to detect a signal segment. Calculate using the following formula: ， in, The projection energy ratio, For the sake of the aforementioned angle consistency, For the physical residual components, The target abnormal component, Describing the L2 norm, To prevent division by zero of positive numbers, For the Sigmoid function, , , and These are calibration parameters.

2. The magnetic anomaly signal detection method based on orthogonal basis function physical consistency network according to claim 1, characterized in that, The orthogonal basis function target physical subspace matrix is ​​constructed as follows: the target and survey line parameter spaces are discretized to obtain multiple sets of parameter combinations; each set of parameter combinations is substituted into the magnetic anomaly target response model to generate the corresponding candidate anomaly waveform; All candidate abnormal waveforms are arranged into a candidate waveform matrix by columns; singular value decomposition is performed on the candidate waveform matrix, and the number of basis functions is determined according to the cumulative energy threshold. Take the front Using left singular vectors as orthogonal basis functions, an orthogonal basis function objective physical subspace matrix is ​​formed. ;Depend on Determine the physical projection matrix The physical projection layer is through accomplish.

3. The magnetic anomaly signal detection method based on orthogonal basis function physical consistency network according to claim 2, characterized in that, The physical decomposition is performed in the following manner: , , Among them, for the first A signal segment to be detected, The target abnormal component, For the physical consistency component, For the physical residual components, Let be the physical projection matrix.

4. The magnetic anomaly signal detection method based on orthogonal basis function physical consistency network according to claim 1, characterized in that, For the A signal segment to be detected, the projection energy ratio Consistency with the stated angle Calculate using the following formulas respectively: ， ， in, The target abnormal component, For the physical consistency component, Describing the L2 norm, Represents the vector dot product. To prevent division by zero of positive numbers.

5. The magnetic anomaly signal detection method based on orthogonal basis function physical consistency network according to claim 1, characterized in that, The total training loss function of the neural network for: ， in, The binary cross-entropy loss is the probability of the target's existence. For loss of angular consistency, For the projected energy separation loss, To punish losses in the event of conflict To smooth out the loss, to These are the weighting coefficients.

6. The magnetic anomaly signal detection method based on orthogonal basis function physical consistency network according to claim 5, characterized in that, The conflict penalty loss Calculate using the following formula: ， in, For the batch sample size, For the first The probability of the target existing in each sample. For the first The proportion of projected energy for each sample.

7. The magnetic anomaly signal detection method based on orthogonal basis function physical consistency network according to claim 5, characterized in that, The angle consistency loss and the projected energy separation loss Calculate using the following formulas respectively: ， ， in, For the batch sample size, For sample labels, target samples Background samples ; For the first The proportion of projected energy for each sample, For the first The consistency of the angles of each sample; and These are the projection energy interval thresholds for the target sample and the background sample, respectively. .

8. The magnetic anomaly signal detection method based on orthogonal basis function physical consistency network according to claim 5, characterized in that, The smoothing loss Calculate using the following formula: ， in, For the batch sample size, For the first The target anomaly component of each sample. It is a first-order or second-order difference operator. This represents the L1 norm.

Images