Magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion

Abstract

The present invention provides a magnetic anomaly identification method based on multi-sequence feature extraction and multi-dimensional feature fusion, belonging to the field of magnetic target detection technology. The method comprises: controlling a pre-mounted platform to cyclically move according to a preset program set that matches an initial layout diagram of real and false targets; activating a magnetic sensor to collect surrounding magnetic field data in real time; employing wavelet transform and sparse representation theory to perform multi-scale decomposition on the raw magnetic anomaly data under each preset program, extracting feature information at different scales; constructing a multi-feature fusion deep learning model; performing multi-sequence feature extraction and multi-dimensional feature fusion on the feature information under each preset program, obtaining a magnetic anomaly identification result; performing result placement analysis on all magnetic anomaly identification results based on the initial layout diagram, obtaining a final identification result, and outputting it. This method reduces the false alarm rate during the magnetic anomaly target identification process.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetic target detection, and in particular to a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion. Background Art

[0002] Magnetic target detection is based on the identification and positioning of the geomagnetic field disturbance anomalies caused by the magnetic target itself. The detection platform is equipped with a magnetometer to collect magnetic field data, and integrates intelligent recognition modules and related algorithms to detect, identify and locate magnetic anomalies based on magnetic field changes, distinguish magnetic anomalies, and then determine them as magnetic targets.

[0003] A magnetic anomaly plot is used in magnetic exploration to show the deviation of the measured magnetic field intensity from the normal magnetic field intensity along a specific survey line or path. This graphical representation can intuitively reveal the location, shape, occurrence, and possible distribution of magnetic objects. A magnetic anomaly plot typically includes the following key elements:

[0004] Horizontal coordinate: represents the geographical location or distance of the measurement point along the survey line. It can be a straight-line distance, longitude and latitude, or other position indicators in a geographic reference system.

[0005] The vertical axis represents the magnetic anomaly value, which is the measured magnetic field intensity minus the theoretical magnetic field intensity (normal field) for the area. The unit is usually nanotesla (nT). Positive values ​​indicate an increase in the magnetic field, while negative values ​​indicate a decrease.

[0006] Curve morphology: The shape of the magnetic anomaly curve can reflect the characteristics of the magnetic object. For example, simple magnetic objects such as cylinders, plates, or spheres will produce specific patterns on the curve, such as peaks, depressions, or complex double-peak structures.

[0007] like Figure 1 The figure shows a partial screenshot of a magnetic anomaly curve under ideal conditions, with a good signal-to-noise ratio and a close-range magnetometer test. In aerial surveys, the magnetometer is far from the target and, due to the inherent dynamic noise of the carrier, the target signal is submerged in the larger geological background, making it difficult to identify. Based on modern signal processing methods, the collected total field data undergoes magnetic compensation, denoising, geomagnetic field correction, and filtering to produce a magnetic anomaly curve. However, due to various environmental interferences, platform dynamic noise, and data processing methods, a high number of false targets often appear, significantly affecting the accuracy of magnetic target identification and resulting in an excessively high false alarm rate for magnetic anomaly targets.

[0008] Therefore, the present invention proposes a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion. Summary of the Invention

[0009] The present invention provides a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion to solve the above-mentioned technical problems.

[0010] The present invention provides a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion, comprising:

[0011] Step 1: After the real target and the decoy target are placed, the pre-mounted platform is controlled to cyclically move according to a preset program set that matches the initial placement diagram of the real target and the decoy target. At the same time, during the movement of the pre-mounted platform, a magnetic sensor is activated to collect surrounding magnetic field data in real time, and the raw magnetic field data based on each preset program is stored separately, wherein the preset program is related to the pitch, roll, turn, acceleration, descent, lift and uniform speed state of the pre-mounted platform, and the state combination order under different preset programs is different;

[0012] Step 2: Using wavelet transform and sparse representation theory, perform multi-scale decomposition on the original magnetic anomaly data under each preset program and extract feature information at different scales;

[0013] Step 3: Construct a multi-feature fusion deep learning model, and perform multi-sequence feature extraction and multi-dimensional feature fusion on the feature information under each preset program to obtain the magnetic anomaly recognition result;

[0014] Step 4: Perform result placement analysis on all magnetic anomaly recognition results based on the initial placement diagram, obtain the final recognition results and output them.

[0015] Preferably, before controlling the movement of the pre-loading platform, the method further includes:

[0016] The motion states involved are moved continuously to construct the first motion sequence;

[0017] Perform discontinuous motion on the motion states involved to construct a second motion sequence;

[0018] Based on the first motion sequence and the second motion sequence, a cyclic motion is formed;

[0019] Assigning a corresponding preset program to each motion sequence to obtain a preset program set based on the cyclic motion;

[0020] There is a pre-set motion trajectory program for each motion state, and the motion trajectory program is related to the initial placement diagram.

[0021] Preferably, the motion states involved are continuously moved to construct a first motion sequence, including:

[0022] Based on the set standard motion curves under different motion states, and the motion speed of each motion point in the standard motion curve, the environmental noise interference coefficient under each motion state is obtained;

[0023] Sort all environmental noise interference coefficients from large to small to obtain a first sub-order;

[0024] Obtaining a first vector of a starting point and a second vector of an ending point of a standard motion curve under different set motion states, and separately calculating the first vector and the second vector to obtain a third vector, wherein the third vector includes: a vector direction and a vector magnitude;

[0025] Sort the vector sizes in all motion states from large to small to obtain the second sub-sequence;

[0026] The vector directions in all motion states are sorted clockwise with the y-axis as the reference line to obtain the third sub-order;

[0027] All the first points are placed as first origins and overlapped, all the first vectors are globally calculated to obtain a fourth vector, a first coefficient is assigned to each first vector according to an angular difference and a length difference between the fourth vector and each first vector, and all the first coefficients are sorted from largest to smallest to obtain a fourth sub-sequence;

[0028] All tail points are used as second origins and overlapped, all second vectors are globally calculated to obtain a fifth vector, a second coefficient is assigned to each second vector according to an angular difference and a length difference between the fifth vector and each second vector, and all second coefficients are sorted from largest to smallest to obtain a fifth sub-sequence;

[0029] A first motion sequence is constructed based on the first sub-sequence, the second sub-sequence, the third sub-sequence, the fourth sub-sequence, and the fifth sub-sequence.

[0030] Preferably, the feature information includes: time domain features, image features and frequency domain features.

[0031] Preferably, a multi-feature fusion deep learning model is constructed, including:

[0032] The historical magnetic anomaly data is subjected to feature extraction, which is divided into historical time domain features and historical frequency domain features. The features of historical magnetic anomaly data at different scales are modeled using a recurrent neural network and processed using a convolutional backbone network to obtain the first feature domain data.

[0033] The image features of the time-frequency images of historical magnetic anomaly data are processed by a two-dimensional convolutional backbone neural network to obtain the second feature domain data;

[0034] Based on the interpolation method, the feature domain data under the same historical magnetic anomaly data are aligned in time and space dimensions and mapped to a unified feature space. Combined with the attention system, the focus on important areas and important features is enhanced to achieve data fusion.

[0035] The fused data is subjected to neural network target recognition, an adaptive optimization algorithm is added, an adaptive weight allocation mechanism and a dynamic learning rate adjustment strategy are introduced, and the network parameters are automatically optimized according to the difficulty of the data and the model performance during training until the set standards are met, thereby obtaining a multi-feature fusion deep learning model.

[0036] Preferably, before aligning the characteristic domain data of the same historical magnetic anomaly data in time and space dimensions based on the interpolation method, the method further includes:

[0037] Sort the first feature domain data by time and lock the missing time points;

[0038] If the number of missing time points is 0, then the first feature domain data remains unchanged;

[0039] If there are multiple missing time points, the first data closest to each missing time point, the second closest first data, and the third closest first data are locked;

[0040] Counting a first number of missing time points between the first data and the second data, a second number of missing time points between the second data and the third data, and at the same time, counting a first consecutive frequency of missing data in the time-sorted data;

[0041] Based on the average of the first data, the second data, and the third data, the data corresponding to the missing time point is calculated and interpolated.

[0042] Preferably, before performing result placement analysis on all magnetic anomaly identification results based on the initial placement diagram, the following steps are included:

[0043] capturing an actual motion process parameter set and a standard motion process parameter set of the pre-mounted platform based on each motion state under different preset programs to determine motion differences;

[0044] Obtaining a first difference between a second vector of a first motion state and a first vector of a second motion state, and a second difference between the first vector and the second vector of the second motion state, in two adjacent motion states under each preset program;

[0045] constructing an instantaneous function corresponding to the motion state based on the motion difference, the first difference, and the second difference under the same motion state, and setting a jamming coefficient for the corresponding motion combination state based on the instantaneous function, wherein the motion combination state includes two different motion states;

[0046] Assigning the jam coefficient to each motion state corresponding to the motion combination state;

[0047] The assigned results are compared with the state combination sequence under each preset program, and the magnetic anomaly identification results under the corresponding preset program are supplemented and analyzed.

[0048] Preferably, all magnetic anomaly identification results are analyzed based on the initial layout diagram, including:

[0049] For the locked positions of the true targets and false targets in the initial placement diagram, the magnetic anomaly recognition results under each sequential movement are placed in one-to-one correspondence with the corresponding locked positions to obtain an initial recognition matrix, and the initial recognition matrix is ​​standardized according to the relative position relationship;

[0050] Covering line segments of the initial placement diagram according to each motion state under each preset program, and simultaneously, position locking of the real targets and false targets in the initial placement diagram;

[0051] According to the locked target under the connecting line segment of two adjacent covering line segments and combined with the jamming coefficient under the corresponding motion combination state, the standardized result of the locked position of the corresponding locked target is adjusted, and the average value of each column vector is calculated to obtain the final recognition result.

[0052] Compared with the prior art, the present invention has the following advantages:

[0053] By allowing the pre-mounted platform to move and collect data according to different preset programs, it can obtain magnetic field data from multiple perspectives and in multiple motion states, enriching the data dimension and providing a more comprehensive data foundation for the subsequent accurate identification of magnetic anomaly targets. The multi-feature fusion deep learning model can fully explore the characteristics of different aspects of magnetic anomaly data. Multi-sequence feature extraction and multi-dimensional feature fusion provide the model with a more comprehensive and in-depth understanding of magnetic anomaly data, improving the accuracy and reliability of magnetic anomaly target identification.

[0054] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0055] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0057] Figure 1 is a schematic diagram of a magnetic anomaly curve under ideal conditions in an embodiment of the present invention;

[0058] Figure 2 Flowchart of a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion in an embodiment of the present invention;

[0059] Figure 3 This is a structural diagram of missing continuous frequencies in an embodiment of the present invention;

[0060] Figure 4 This is a structural diagram of the multi-feature fusion deep learning model in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0062] The present invention provides a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion, such as Figure 2 Shown, including:

[0063] Step 1: After the real target and the decoy target are placed, the pre-mounted platform is controlled to cyclically move according to a preset program set that matches the initial placement diagram of the real target and the decoy target. At the same time, during the movement of the pre-mounted platform, a magnetic sensor is activated to collect surrounding magnetic field data in real time, and the raw magnetic field data based on each preset program is stored separately, wherein the preset program is related to the pitch, roll, turn, acceleration, descent, lift and uniform speed state of the pre-mounted platform, and the state combination order under different preset programs is different;

[0064] Step 2: Using wavelet transform and sparse representation theory, perform multi-scale decomposition on the original magnetic anomaly data under each preset program and extract feature information at different scales;

[0065] Step 3: Construct a multi-feature fusion deep learning model, and perform multi-sequence feature extraction and multi-dimensional feature fusion on the feature information under each preset program to obtain the magnetic anomaly recognition result;

[0066] Step 4: Perform result placement analysis on all magnetic anomaly recognition results based on the initial placement diagram, obtain the final recognition results and output them.

[0067] In this embodiment, the pre-mounted platform utilizes a carrier capable of multiple motion modes. Equipped with a high-precision attitude control and motion control system, the platform's motion is controlled by a set of preset programs that match the initial layout of the real and decoy targets. These preset programs relate to the platform's pitch (up and down rotation around the horizontal axis), roll (left and right rolling around the vertical axis), turn (rotation around the vertical axis), acceleration, descent, and ascent states. The order in which these states are combined varies between programs. For example, for an aerial detection scenario, program 1 might have the platform first fly horizontally at a constant speed (no pitch or roll), then perform a small pitch ascent, and then turn; while program 2 might first roll a certain angle, then accelerate to descend, and so on. This allows the magnetic sensor to collect magnetic field data from various angles and motion states, ensuring data accuracy.

[0068] In this embodiment, magnetic sensors are distributed at various locations on the platform to comprehensively collect magnetic field information from the surrounding space. These sensors include Hall effect sensors, giant magnetoresistive sensors, and fluxgate sensors. For example, a high-precision fluxgate sensor is used, collecting data at a rate of 100 times per second. The collected data includes magnetic field information of real targets, false targets, and the surrounding environment. Raw magnetic field data based on each preset program is stored separately for subsequent classification and analysis.

[0069] In this embodiment, a dedicated target placement device is installed within the area where the platform moves. This device can flexibly arrange various types of decoys and real targets according to experimental requirements. Decoys can be non-magnetic or weakly magnetic objects of various shapes and materials, acting as interference objects; real targets are objects with specific magnetic characteristics, such as ferromagnetic metal blocks and magnetic weapons and equipment.

[0070] In this embodiment, after the real target and the decoy target are placed, since their positions are determined, a placement result diagram, namely, an initial placement diagram, will be obtained.

[0071] In this embodiment, during the operation of the control platform, due to the presence of certain unusual states during motion, such as pitch, roll, turn, acceleration, descent, and elevation, sudden changes in state can affect the magnetic field to a certain extent. As the platform accelerates, decelerates, and turns, its internal motors, circuits, and other components generate electromagnetic signals. For example, in drones, high-speed motor operation and the operation of the electronic speed controller generate electromagnetic noise. This noise can be superimposed on the magnetic field signal of the magnetic target, causing the measurement results to deviate from the true value. Changes in motion, particularly during acceleration, elevation, and descent, can cause the platform to generate varying degrees of mechanical vibration. This vibration can cause the magnetometer and magnetic sensor mounted on the platform to shift or jitter, altering the sensor's measurement direction and position, and thus affecting the accuracy of the measurement results. For example, on a vehicle-mounted platform, the jolting vibrations of traversing uneven roads can interfere with the sensor's stable measurement of the magnetic target's magnetic field. Attitude changes such as pitch and roll can alter the relative position and angle between the sensor and the magnetic target. When the platform pitches, the magnetic field components of the magnetic target measured by the sensor change. Without accurate attitude compensation, measurement errors can result. It should be noted that the preset programs in each motion state are pre-set. For example, the acceleration state is to move forward in the horizontal direction at an acceleration of 0.1 square meters / meter, and the moving time is 2 seconds.

[0072] Therefore, it is necessary to determine the cyclic motion through the combination design of the states. For example, there are 6 combination designs. At this time, controlling the movement of the carrying platform in accordance with combination 1, combination 2, combination 3, combination 4, combination 5 and combination 6 in sequence is a cyclic motion until the carrying platform completes the execution of combination 6.

[0073] In this embodiment, the raw magnetic anomaly data segment collected by program A in step 1 is assumed to be a discrete signal sequence of length 1000. After a three-layer decomposition using the Daubechies-4 wavelet basis, three sets of high-frequency detail coefficients and one set of low-frequency approximation coefficients are obtained. Then, using the orthogonal matching pursuit algorithm, 100 key coefficients, for example, are selected from these coefficients. These key coefficients represent the main feature information of the magnetic anomaly data segment at different scales, thus achieving multi-scale feature extraction of the data. Through wavelet transform and sparse representation theory, feature information at different scales is extracted from the raw magnetic anomaly data, eliminating redundant information and highlighting key features. This enables subsequent models to process data more efficiently and improves the accuracy of magnetic anomaly target identification.

[0074] In this embodiment, step 2 specifically includes:

[0075] The original magnetic anomaly data is normalized to ensure the comparability of features at different scales and eliminate the dimension effect.

[0076] Apply low-pass filter and other technologies to remove high-frequency noise and retain the essential characteristics of the signal.

[0077] Using wavelet transform, the magnetic data is decomposed into components of different scales. Each scale corresponds to a different frequency bandwidth, and features at different time scales can be extracted.

[0078] Using empirical mode decomposition (EMD), the signal is adaptively decomposed into a series of intrinsic mode functions (IMFs), each of which represents a fluctuation mode on a time scale.

[0079] Based on the subsequent processing of EMD, the frequency characteristics of each IMF are analyzed in combination with Hilbert transform to further refine the multi-scale analysis.

[0080] Statistical quantities such as mean, variance, maximum, minimum, kurtosis, and skewness are calculated at different scales to extract the central tendency, dispersion, and distribution form of magnetic data.

[0081] Through spectrum analysis of Fourier transform and wavelet transform, time-frequency diagram is drawn, energy is distinguished by different colors, RGB image is generated, and corresponding distribution features are extracted.

[0082] In this embodiment, Figure 4 The figure shows a multi-feature fusion deep learning model. This model designs a novel deep neural network architecture that combines a convolutional neural network (CNN) with a recurrent neural network (RNN). CNN captures local spatial features, while RNN captures temporal or spatial correlations in magnetic data sequences, improving the model's ability to understand complex target features. An attention-based feature fusion method is used to fuse time-domain features, frequency-domain features, and image features. The fused features are then subjected to neural network recognition, outputting the recognition results.

[0083] In this embodiment, during the result placement and analysis process, the magnetic data needs to be proportionally adjusted according to the positional relationship to facilitate comprehensive processing of the magnetic data at the same position and ensure the accuracy of the measurement and acquisition results.

[0084] The beneficial effect of this technical solution is that by allowing the pre-mounted platform to move and collect data according to different preset programs, it can obtain magnetic field data from multiple perspectives and motion states, enriching the data dimension and providing a more comprehensive data foundation for the subsequent accurate identification of magnetic anomaly targets. The multi-feature fusion deep learning model can fully exploit the characteristics of different aspects of magnetic anomaly data. Multi-sequence feature extraction and multi-dimensional feature fusion provide the model with a more comprehensive and in-depth understanding of magnetic anomaly data, improving the accuracy and reliability of magnetic anomaly target identification and reducing the false alarm rate during the magnetic anomaly target identification process.

[0085] The present invention provides a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion, which, before controlling the movement of the pre-loading platform, further comprises:

[0086] The motion states involved are moved continuously to construct the first motion sequence;

[0087] Perform discontinuous motion on the motion states involved to construct a second motion sequence;

[0088] Based on the first motion sequence and the second motion sequence, a cyclic motion is formed;

[0089] Assigning a corresponding preset program to each motion sequence to obtain a preset program set based on the cyclic motion;

[0090] There is a pre-set motion trajectory program for each motion state, and the motion trajectory program is related to the initial placement diagram.

[0091] In this embodiment, the first motion sequence includes five sequences, and the non-continuous motion refers to controlling the pre-loaded platform to operate separately according to each motion state, which serves as a basic reference.

[0092] In this embodiment, continuous motion refers to the motion states being changed to another state immediately after running for a set time interval. It should be noted that there are only the mentioned motion states.

[0093] The beneficial effects of the above technical solution are: forming a cyclic motion through continuous motion and discontinuous motion, and realizing reasonable control of the pre-loading platform through program configuration.

[0094] The present invention provides a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion, which continuously moves the motion states involved and constructs a first motion sequence, including:

[0095] Based on the set standard motion curves under different motion states, and the motion speed of each motion point in the standard motion curve, the environmental noise interference coefficient under each motion state is obtained;

[0096] Sort all environmental noise interference coefficients from large to small to obtain a first sub-order;

[0097] Obtaining a first vector of a starting point and a second vector of an ending point of a standard motion curve under different set motion states, and separately calculating the first vector and the second vector to obtain a third vector, wherein the third vector includes: a vector direction and a vector magnitude;

[0098] Sort the vector sizes in all motion states from large to small to obtain the second sub-sequence;

[0099] The vector directions in all motion states are sorted clockwise with the y-axis as the reference line to obtain the third sub-order;

[0100] All the first points are placed as first origins and overlapped, all the first vectors are globally calculated to obtain a fourth vector, a first coefficient is assigned to each first vector according to an angular difference and a length difference between the fourth vector and each first vector, and all the first coefficients are sorted from largest to smallest to obtain a fourth sub-sequence;

[0101] All tail points are used as second origins and overlapped, all second vectors are globally calculated to obtain a fifth vector, a second coefficient is assigned to each second vector according to an angular difference and a length difference between the fifth vector and each second vector, and all second coefficients are sorted from largest to smallest to obtain a fifth sub-sequence;

[0102] A first motion sequence is constructed based on the first sub-sequence, the second sub-sequence, the third sub-sequence, the fourth sub-sequence, and the fifth sub-sequence.

[0103] Preferably, the feature information includes: time domain features, image features and frequency domain features.

[0104] In this embodiment, the environmental noise interference coefficient is calculated as follows:

[0105] ;

[0106] Wherein, max represents the maximum value symbol; Zg represents the environmental noise interference coefficient under the corresponding motion state; N1 represents the number of motion points involved in the corresponding standard motion curve; 、 、 They represent proportional constants, which are preset; represents the derivative of the velocity vi with respect to time ti at the i1th moving point; Represents the motion curvature of the i1th moving point i; fi represents the motion friction force under the i1th motion point;

[0107] In this embodiment, 、 、 It is pre-set to adapt to different motion states, that is, the proportional constants in different motion states are different. For example, in the accelerated motion state: = , , It should be noted that the standard motion curves under different motion states are set, and the programs corresponding to the curves are also set in order to control the carrying platform to move according to the preset programs.

[0108] In this embodiment, under actual motion conditions, factors such as speed change, motion curvature, and motion friction can all contribute to environmental noise interference, but the noise generated by one factor often dominates. For example, when a pre-loaded platform rapidly starts or brakes, the noise generated by the speed change rate may be far greater than the noise generated by other factors. Furthermore, when the platform makes sharp turns or other large curvature movements, the noise generated by the motion curvature may be most prominent. By taking the maximum value, the factor that contributes most to the environmental noise interference in a specific motion condition can be accurately captured, and this factor can be used as the key indicator representing the noise interference level of the entire motion condition, thus preventing the judgment of the overall noise level from being influenced by other relatively minor factors. If the maximum value is not taken, simply adding the three factors together or performing other complex calculations would increase the computational complexity and may not clearly and unambiguously reflect the primary source of interference. Taking the maximum value is relatively simple and direct, and can quickly determine which factor is the most critical noise source in the current motion condition. Furthermore, the obtained maximum value can intuitively reflect the upper limit of the environmental noise interference level, providing a clear numerical representation of the severity of the environmental noise interference in that motion condition, facilitating subsequent comparison and analysis. For example, when comparing noise interference under different exercise programs, comparing the maximum values ​​can quickly determine which exercise program is most affected by noise interference. In reality, the contributions of different factors to noise interference can vary greatly, and the magnitudes of these factors may also vary. If a strategy of averaging the values ​​instead of taking the maximum value is adopted, the truly dominant factors will be obscured, resulting in the calculated environmental noise interference coefficient not accurately reflecting the actual interference situation. Taking the maximum value ensures that the most critical interference factors can be accurately captured in various complex situations, making the calculated environmental noise interference coefficient more effective and representative.

[0109] For example, the code for the acceleration state is as follows:

[0110] / / Initial state of the platform

[0111] let x = 50;

[0112] let y = 50;

[0113] let vx = 0;

[0114] let vy = 0;

[0115] const ax = 0.1; / / horizontal acceleration

[0116] const ay = 0.1; / / Vertical acceleration.

[0117] In this embodiment, the standard motion curves under different motion states are pre-set and can be used directly, and are stored in a state-curve comparison table, which contains standard motion curves under different motion states. That is, in this motion state, the speed direction and speed magnitude executed at each time point are different, but they are all known and set. For example, the speed direction at the first time point in the acceleration state is the positive 90° direction, and the speed magnitude is 5 cm / s. Therefore, the first vector and the second vector can be directly determined. In the vector calculation process, the third vector, the fourth vector, and the fifth vector are obtained through the vector integration designed in the technical solution. It should be noted that the first vector and the second vector are the motion speed and motion direction of the first motion point and the last motion point in the corresponding curve, the motion speed corresponds to the vector magnitude, and the motion direction corresponds to the vector direction. The third vector is obtained by placing the first vector and the second vector based on the origin in the coordinate system and performing vector addition (vector integration) to obtain the third vector.

[0118] In this embodiment, the clockwise sorting means calculating the clockwise angle between each vector direction and the y-axis, and sorting the clockwise angles from large to small to obtain the third sub-order.

[0119] In this embodiment, the global calculation is to perform vector calculation on all first vectors involved in the first origin after overlapping placement, so as to obtain the fourth vector.

[0120] In this embodiment, the angle difference is the angle between the fourth vector and the first vector; and the length difference is the difference between the speeds of the fourth vector and the first vector.

[0121] In this embodiment, , and the calculation method of the second coefficient is similar to that of the first coefficient, which will not be repeated here.

[0122] The beneficial effect of the above technical solution is: by analyzing the interference coefficient, the size and direction of the vector corresponding to the first point and the last point, the states under different situations are sorted to obtain the first motion sequence under the continuous state.

[0123] The present invention provides a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion, and constructs a multi-feature fusion deep learning model, including:

[0124] The historical magnetic anomaly data is subjected to feature extraction, which is divided into historical time domain features and historical frequency domain features. The features of historical magnetic anomaly data at different scales are modeled using a recurrent neural network and processed using a convolutional backbone network to obtain the first feature domain data.

[0125] The image features of the time-frequency images of historical magnetic anomaly data are processed by a two-dimensional convolutional backbone neural network to obtain the second feature domain data;

[0126] Based on the interpolation method, the feature domain data under the same historical magnetic anomaly data are aligned in time and space dimensions and mapped to a unified feature space. Combined with the attention system, the focus on important areas and important features is enhanced to achieve data fusion.

[0127] The fused data is subjected to neural network target recognition, an adaptive optimization algorithm is added, an adaptive weight allocation mechanism and a dynamic learning rate adjustment strategy are introduced, and the network parameters are automatically optimized according to the difficulty of the data and the model performance during training until the set standards are met, thereby obtaining a multi-feature fusion deep learning model.

[0128] In this embodiment, historical magnetic anomaly data refers to magnetic field data collected by magnetic sensors and other equipment within a specific time period and area in the past, which differs from normal magnetic field conditions. These differences may be caused by the presence of magnetic target objects (such as ferromagnetic ores and magnetic metal products).

[0129] In this embodiment, historical time-domain features are extracted from historical magnetic anomaly data and reflect how magnetic field characteristics change over time. Examples include the rate of change of magnetic field intensity at different times and its periodicity. A certain segment of historical magnetic anomaly data records the magnetic field intensity value every second.

[0130] In this embodiment, the historical frequency domain feature converts historical magnetic anomaly data from the time domain to the frequency domain, revealing characteristics regarding the distribution of magnetic field energy across different frequency components. This feature can reveal information such as the signal strength and dominant frequency of different frequencies in magnetic field variations. A Fourier transform is performed on the historical magnetic anomaly data, which records magnetic field intensity by the second, to obtain its frequency domain representation. It is found that the magnetic field energy is highly concentrated around a frequency of 5 Hz in the data. This energy concentration around 5 Hz, along with the energy distribution at other frequencies, constitutes the historical frequency domain feature.

[0131] In this embodiment, the first feature domain data is a representation of historical magnetic anomaly data within a specific feature space, obtained after recurrent neural network modeling and convolutional backbone network processing. It contains the processed and abstracted time-domain and frequency-domain features of the magnetic anomaly data. After the RNN captures temporal correlations and the convolutional backbone network extracts local features, a new set of data vectors is generated. This set of data vectors constitutes the first feature domain data, which is no longer the original magnetic anomaly data but a more representative representation of the data.

[0132] In this embodiment, a time-frequency image is created by converting time-domain magnetic anomaly data onto a time-frequency plane using a specific transformation (such as a short-time Fourier transform or wavelet transform). The image's horizontal axis represents time, and the vertical axis represents frequency. The image's pixel values ​​reflect information such as the magnetic field energy or amplitude at the corresponding time and frequency. A short-time Fourier transform is performed on magnetic anomaly data collected continuously over a period of time, decomposing the magnetic field data at each time point into different frequencies. These are then plotted on a two-dimensional plane using varying grayscale or color according to the energy of each frequency component. The resulting image is called a time-frequency image.

[0133] In this embodiment, the second feature domain data is a representation of the time-frequency images of historical magnetic anomaly data within a specific feature space, obtained after processing them using a two-dimensional convolutional neural network. This representation primarily reflects the characteristic information of the magnetic anomaly data within the time-frequency images. After extracting features from the time-frequency images using the two-dimensional convolutional neural network, a new set of data feature vectors is generated. These vectors constitute the second feature domain data, which further abstracts and represents the features of the time-frequency images.

[0134] This embodiment uses the Adam optimization algorithm, which combines the concepts of momentum and adaptive learning rates. When training a multi-feature fusion deep learning model, the Adam algorithm automatically adjusts the learning rate based on the gradient changes of each parameter during training, using smaller learning rates for parameters with large gradient changes and larger learning rates for parameters with small gradient changes, thereby accelerating model convergence.

[0135] In this embodiment, in the training of the magnetic anomaly target recognition model, the adaptive weight allocation mechanism will increase the weight values ​​of the connection weights corresponding to the key features that can accurately distinguish between magnetic targets and non-magnetic targets, so that the model pays more attention to these features, and reduce the weights corresponding to those features that contribute less to target recognition.

[0136] The beneficial effect of this technical solution is that the adaptive weight allocation mechanism increases the weights of features, allowing the model to focus more on them, while reducing the weights of features that contribute less to object recognition. When training a multi-feature fusion deep learning model, a high learning rate is set at the beginning to allow the model to quickly update parameters. As training progresses, if the loss function value decreases slowly or fluctuates, the learning rate is reduced, allowing the model to make more detailed parameter adjustments in areas closer to the optimal solution.

[0137] The present invention provides a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion, which further includes:

[0138] Sort the first feature domain data by time and lock the missing time points;

[0139] If the number of missing time points is 0, then the first feature domain data remains unchanged;

[0140] If there are multiple missing time points, the first data closest to each missing time point, the second closest first data, and the third closest first data are locked;

[0141] Counting a first number of missing time points between the first data and the second data, a second number of missing time points between the second data and the third data, and at the same time, counting a first missing consecutive frequency in the time-sorted data;

[0142] Based on the average of the first data, the second data, and the third data, the data corresponding to the missing time point is calculated and interpolated.

[0143] ;

[0144] Wherein, d1, d2, and d3 represent the values ​​of the first data, the second data, and the third data respectively; e1 represents the set threshold value, which is 0.1; represents the first quantity; represents the second quantity; n represents the total number of median values ​​in the time-sorted data; Lp represents the first missing continuous frequency; pN represents the total number of missing time points involved;

[0145] It should be noted that as long as the missing time points are continuous and the number of consecutive points is greater than or equal to 2, it is considered as a continuous frequency. Figure 3 As shown, there are 10 time points, and the circles represent the non-missing time points, and the squares represent the missing time points. At this time, 6 and 7, 9 and 10 are all missing continuously, the value of Lp is 2, and the value of pN is 5.

[0146] In this embodiment, the exponential decay term The system dynamically adjusts the calculation results based on the relevant parameters of the missing time points, more accurately estimating missing data while maintaining a certain degree of smoothness. This reduces interference with subsequent analysis caused by improper handling of missing data and improves overall data quality. It effectively supplements the missing data in the first feature domain, making the data more complete in the temporal dimension. Complete data is crucial for subsequent analysis and modeling based on this feature domain data (such as input into deep learning models for magnetic anomaly identification). It improves the quality of model input data, thereby enhancing the accuracy and reliability of analytical results.

[0147] In this embodiment, under normal circumstances, there will be a corresponding characteristic value at each time point. Therefore, by sorting the first characteristic domain data (characteristic values) in chronological order, it can be directly determined whether a time point is missing.

[0148] In this embodiment, the first data, second data, and third data involved all refer to corresponding characteristic values.

[0149] The beneficial effect of the above technical solution is: by sorting the data in time, the missing time points can be directly locked, and then based on the three data and combined with the quantity and continuous frequency of missing data, effective and reasonable interpolation can be achieved to ensure the integrity of the data and provide a basis for subsequent analysis.

[0150] The present invention provides a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion, which includes:

[0151] capturing an actual instantaneous process parameter set and a standard instantaneous process parameter set based on each motion state of the pre-mounted platform under different preset programs to determine motion differences;

[0152] Obtaining a first difference between a second vector of a first motion state and a first vector of a second motion state, and a second difference between the first vector and the second vector of the second motion state, in two adjacent motion states under each preset program;

[0153] constructing an instantaneous function corresponding to the motion state based on the motion difference, the first difference, and the second difference under the same motion state, and setting a jamming coefficient for the corresponding motion combination state based on the instantaneous function, wherein the motion combination state includes two different motion states;

[0154] Assigning the jam coefficient to each motion state corresponding to the motion combination state;

[0155] The assigned results are compared with the state combination sequence under each preset program, and the magnetic anomaly identification results under the corresponding preset program are supplemented and analyzed.

[0156] In this embodiment, the first difference and the second difference refer to the differences in speed and direction on the vector.

[0157] In this embodiment, the actual motion process parameter set refers to the operating parameters of the specified motion state at the initial operating moment, including: actual speed, actual direction. In actual flight, the actual speed of the drone measured by the sensor is 4.8m / s, and the flight direction is different from the standard direction. Then, the difference in speed is 5-4.8=0.2m / s, and the difference in speed direction is .

[0158] In this embodiment, the instantaneous function f(x)=g1 (motion difference between the two states)+g2 (first difference)+g3 (second difference).

[0159] The jam coefficient = u1×g1 (motion difference between the two states) + u2×g2 (first difference) + u3×g3 (second difference), and u1, u2, and u3 represent weights, with values ​​of 0.5, 0.3, and 0.2 respectively.

[0160] In this embodiment, g1 (the motion difference between the two states), g2 (the first difference), and g3 (the second difference) are obtained by matching from a dual-state-difference comparison table. This comparison table contains the motion differences, first differences, second differences, and corresponding difference coefficients for different combinations of motion states. These are all pre-stored and can be directly retrieved. By calculating the motion differences and vector differences and setting the stutter coefficient, it was found that some motion state combinations under the preset program have large stutter coefficients. Comparing the magnetic anomaly recognition results without and with the stutter coefficient, it was found that considering the stutter coefficient can identify some misjudgments caused by abnormal platform motion. Without considering the stutter coefficient, the misjudgment rate of magnetic anomaly recognition was 15%; after additional analysis with the stutter coefficient, the misjudgment rate was reduced to 10%.

[0161] The beneficial effects of the above technical solution are: by determining the motion difference, first difference, and second difference of two different motion states, a sequential function is constructed to obtain the jamming coefficient, the jamming coefficient is set and a comparative analysis is performed, and the magnetic anomaly recognition results are supplemented and corrected, thereby improving the accuracy and reliability of magnetic anomaly recognition and reducing the risk of misjudgment due to unstable platform motion.

[0162] The present invention provides a magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion, which performs result placement analysis on all magnetic anomaly recognition results based on an initial placement diagram, including:

[0163] For the locked positions of the true targets and false targets in the initial placement diagram, the magnetic anomaly recognition results under each sequential movement are placed in one-to-one correspondence with the corresponding locked positions to obtain an initial recognition matrix, and the initial recognition matrix is ​​standardized according to the relative position relationship;

[0164] Covering line segments of the initial placement diagram according to each motion state under each preset program, and simultaneously, position locking of the real targets and false targets in the initial placement diagram;

[0165] According to the locked target under the connecting line segment of two adjacent covering line segments and combined with the jamming coefficient under the corresponding motion combination state, the standardized result of the locked position of the corresponding locked target is adjusted, and the average value of each column vector is calculated to obtain the final recognition result.

[0166] In this embodiment, the initial placement diagram shows clear coordinates for both true and false targets. After the pre-loaded platform completes a sequence of motions according to a pre-set program, a set of magnetic anomaly identification results is generated (determining which locations are true targets and which are false targets). These identification results are then mapped one-to-one with the locked positions of the targets in the initial placement diagram. The identification results for each location (e.g., 1 for a true target, 0 for a false target) are placed in the corresponding position of a matrix to form the initial identification matrix.

[0167] Normalize the initial recognition matrix based on relative position. For example, arrange the rows and columns of the matrix according to the coordinate order of the target in the initial placement diagram to ensure that the positions of the matrix elements correspond to the positions of the targets in real space. Also, perform normalization and other operations on the values ​​in the matrix to make different recognition results comparable. For example, if the recognition results use numerical values ​​of varying strength to represent the likelihood of a target, these values ​​can be normalized to the range [0, 1].

[0168] In this embodiment, assume that there are 4 targets in the initial placement diagram, arranged in a 2×2 rectangle, with coordinates of (1,1), (1,2), (2,1), and (2,2). After the pre-loaded platform completes a movement, the recognition result is that the position (1,1) is the true target (marked as 1), the position (1,2) is the false target (marked as 0), the position (2,1) is the false target (marked as 0), and the position (2,2) is the true target (marked as 1). The initial recognition matrix is When performing normalization, if the values ​​are in other ranges, they can be normalized to [0,1] through linear transformation or other methods. Here, the matrix elements are already in the appropriate range, and you only need to ensure that the positions correspond accurately.

[0169] In this embodiment, under each preset program, each motion state of the pre-loaded platform moves within the spatial area corresponding to the initial placement diagram. Based on the platform's motion trajectory and the sensor's detection range, the coverage line segment of each motion state relative to the initial placement diagram is determined. For example, if the platform moves in a straight line and the sensor's detection range is a distance d on either side, then the coverage line segment is the line segment region projected onto the plane of the initial placement diagram, within the distance d on either side of the trajectory. Still using the initial placement diagram of the four targets described above as an example, assume that one motion state of the pre-loaded platform under a preset program is a straight line from left to right, and the sensor's detection range is a distance 0.5 units on either side. The platform's motion trajectory is projected onto the plane of the initial placement diagram as a straight line along the x-axis, passing through the area from x=0.5 to x=2.5. The coverage line segment is then the line segment region within the ranges 0.5 to 2.5 on the x-axis and -0.5 to 0.5 on the y-axis. Simultaneously, the coordinates of the four targets are locked to (1,1), (1,2), (2,1), and (2,2).

[0170] The normalized results are adjusted based on the targets locked under the connecting line segments of two adjacent covered segments and the stutter coefficient of the corresponding motion combination. A large stutter coefficient indicates unstable platform motion in that motion combination, which may affect recognition results. In this case, the normalized results for targets locked in that area are corrected. For example, the confidence level for identifying a true target can be appropriately lowered (e.g., adjusting the confidence level for identifying a true target from 0.8 to 0.6). The average value of each column vector in the initial recognition matrix is ​​calculated. Each column vector represents the recognition results under different motion states or programs at the same column position. Calculating the average value combines multiple results to obtain a more accurate and reliable final recognition result. For example, if a column vector is [0.6, 0.4, 0.8], the average value is calculated as (0.6 + 0.4 + 0.8) / 3 = 0.6. This average value is used to determine whether the position is a true target or a false target (e.g., setting a threshold of 0.5, a value greater than 0.5 is considered a true target).

[0171] Assume that there is a locked target under the connecting line segment of two adjacent covered line segments. Its normalized result in the initial recognition matrix is ​​that it is recognized as a true target (value is 0.8), but the motion combination state corresponding to this position has a large jam coefficient. Adjust its value to 0.6. The initial recognition matrix is , calculating the average of the column vectors, the average of the first column is (0.6 + 0.8) / 2 = 0.7, and the average of the second column is (0.4 + 0.2) / 2 = 0.3. According to the threshold of 0.5, the position corresponding to the first column is ultimately identified as a true target, and the position corresponding to the second column is ultimately identified as a false target.

[0172] This example involved generating an initial recognition matrix, determining covered segments, adjusting the results based on the stutter coefficient, and calculating the final recognition result. The results were compared with those obtained without these processing methods. This method significantly improved recognition accuracy. For example, the accuracy rate was 70% without processing, but increased to 85% after processing.

[0173] The beneficial effects of the above technical solution are: it can systematically integrate the magnetic anomaly identification results and target position information, take into account the influence of the platform motion state's jamming coefficient on the identification results, and through standardized processing and comprehensive calculation, effectively improve the accuracy and reliability of magnetic anomaly target identification, and reduce misjudgments and missed judgments.

[0174] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion is characterized by: include: Step 1: After the real target and the decoy target are placed, the pre-mounted platform is controlled to cyclically move according to a preset program set that matches the initial placement diagram of the real target and the decoy target. At the same time, during the movement of the pre-mounted platform, a magnetic sensor is activated to collect surrounding magnetic field data in real time, and the raw magnetic field data based on each preset program is stored separately, wherein the preset program is related to the pitch, roll, turn, acceleration, descent, lift and uniform speed state of the pre-mounted platform, and the state combination order under different preset programs is different; Step 2: Using wavelet transform and sparse representation theory, perform multi-scale decomposition on the original magnetic anomaly data under each preset program and extract feature information at different scales; Step 3: Construct a multi-feature fusion deep learning model, and perform multi-sequence feature extraction and multi-dimensional feature fusion on the feature information under each preset program to obtain the magnetic anomaly recognition result; Step 4: Perform result placement analysis on all magnetic anomaly recognition results based on the initial placement diagram, obtain the final recognition results and output them.

2. The magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion according to claim 1 is characterized in that: Before controlling the motion of the pre-loading platform, it also includes: The motion states involved are moved continuously to construct the first motion sequence; Perform discontinuous motion on the motion states involved to construct a second motion sequence; Based on the first motion sequence and the second motion sequence, a cyclic motion is formed; Assigning a corresponding preset program to each motion sequence to obtain a preset program set based on the cyclic motion; There is a pre-set motion trajectory program for each motion state, and the motion trajectory program is related to the initial placement diagram.

3. The magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion according to claim 2 is characterized in that: The motion states involved are moved continuously to construct the first motion sequence, including: Based on the set standard motion curves under different motion states, and the motion speed of each motion point in the standard motion curve, the environmental noise interference coefficient under each motion state is obtained; Sort all environmental noise interference coefficients from large to small to obtain a first sub-order; Obtaining a first vector of a starting point and a second vector of an ending point of a standard motion curve under different set motion states, and separately calculating the first vector and the second vector to obtain a third vector, wherein the third vector includes: a vector direction and a vector magnitude; Sort the vector sizes in all motion states from large to small to obtain the second sub-sequence; The vector directions in all motion states are sorted clockwise with the y-axis as the reference line to obtain the third sub-order; All the first points are placed as first origins and overlapped, all the first vectors are globally calculated to obtain a fourth vector, a first coefficient is assigned to each first vector according to an angular difference and a length difference between the fourth vector and each first vector, and all the first coefficients are sorted from largest to smallest to obtain a fourth sub-sequence; All tail points are used as second origins and overlapped, all second vectors are globally calculated to obtain a fifth vector, a second coefficient is assigned to each second vector according to an angular difference and a length difference between the fifth vector and each second vector, and all second coefficients are sorted from largest to smallest to obtain a fifth sub-sequence; A first motion sequence is constructed based on the first sub-sequence, the second sub-sequence, the third sub-sequence, the fourth sub-sequence, and the fifth sub-sequence.

4. The magnetic anomaly identification method based on multi-sequence feature extraction and multi-dimensional feature fusion according to claim 1 is characterized in that: The feature information includes: time domain features, image features and frequency domain features.

5. The magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion according to claim 1 is characterized in that: Build a multi-feature fusion deep learning model, including: The historical magnetic anomaly data is subjected to feature extraction, which is divided into historical time domain features and historical frequency domain features. The features of historical magnetic anomaly data at different scales are modeled using a recurrent neural network and processed using a convolutional backbone network to obtain the first feature domain data. The image features of the time-frequency images of historical magnetic anomaly data are processed by a two-dimensional convolutional backbone neural network to obtain the second feature domain data; Based on the interpolation method, the feature domain data of the same historical magnetic anomaly data are aligned in time and space dimensions and mapped to a unified feature space. The attention mechanism is combined to enhance the focus on important areas and important features to achieve data fusion. The fused data is subjected to neural network target recognition, an adaptive optimization algorithm is added, an adaptive weight allocation mechanism and a dynamic learning rate adjustment strategy are introduced, and the network parameters are automatically optimized according to the difficulty of the data and the model performance during training until the set standards are met, thereby obtaining a multi-feature fusion deep learning model.

6. The magnetic anomaly recognition method based on multi-sequence feature extraction and multi-dimensional feature fusion according to claim 5 is characterized in that: Before aligning the characteristic domain data under the same historical magnetic anomaly data in time and space dimensions based on the interpolation method, it also includes: Sort the first feature domain data by time and lock the missing time points; If the number of missing time points is 0, then the first feature domain data remains unchanged; If there are multiple missing time points, the first data closest to each missing time point, the second closest first data, and the third closest first data are locked; Counting a first number of missing time points between the first data and the second data, a second number of missing time points between the second data and the third data, and at the same time, counting a continuous frequency of the first missing points in the time-sorted data; Based on the average of the first data, the second data, and the third data, the data corresponding to the missing time point is calculated and interpolated.

7. The magnetic anomaly identification method based on multi-sequence feature extraction and multi-dimensional feature fusion according to claim 1 is characterized in that: Before analyzing the placement of all magnetic anomaly identification results based on the initial placement diagram, the following steps should be performed: capturing an actual motion process parameter set and a standard motion process parameter set of the pre-mounted platform based on each motion state under different preset programs to determine motion differences; Obtaining a first difference between a second vector of a first motion state and a first vector of a second motion state, and a second difference between the first vector and the second vector of the second motion state, in two adjacent motion states under each preset program; constructing an instantaneous function corresponding to the motion state based on the motion difference, the first difference, and the second difference under the same motion state, and setting a jamming coefficient for the corresponding motion combination state based on the instantaneous function, wherein the motion combination state includes two different motion states; Assigning the jam coefficient to each motion state corresponding to the motion combination state; The assigned results are compared with the state combination sequence under each preset program, and the magnetic anomaly identification results under the corresponding preset program are supplemented and analyzed.

8. The magnetic anomaly identification method based on multi-sequence feature extraction and multi-dimensional feature fusion according to claim 7 is characterized in that: All magnetic anomaly identification results are analyzed based on the initial layout diagram, including: For the locked positions of the true targets and false targets in the initial placement diagram, the magnetic anomaly recognition results under each sequential movement are placed in one-to-one correspondence with the corresponding locked positions to obtain an initial recognition matrix, and the initial recognition matrix is ​​standardized according to the relative position relationship; Covering line segments of the initial placement diagram according to each motion state under each preset program, and simultaneously, position locking of the real targets and false targets in the initial placement diagram; According to the locked target under the connecting line segment of two adjacent covering line segments and combined with the jamming coefficient under the corresponding motion combination state, the standardized result of the locked position of the corresponding locked target is adjusted, and the average value of each column vector is calculated to obtain the final recognition result.

Images