A metal detection device and method

Abstract

The application belongs to the technical field of metal detection, and discloses a metal detection device and method. The application adopts multiple sets of resonant sensing units to form a multi-channel inductive sensor array, forms multiple effective detection channels, expands the detection coverage, and acquires multi-channel response information with spatial differences. In addition, in the application, each set of resonant sensing units is combined with an LC resonant loop and a double-channel inductive sensor chip, thereby improving the sensitivity of metal detection. Furthermore, the application also builds a metal detection classification model based on asynchronous event space-time graph multi-branch fusion, fuses CNN time domain features, GRU asynchronous event time sequence features, GCN double-node space interaction features and statistical features, can simultaneously utilize local disturbance, time dynamics, channel space correlation and stable statistical information of the double-channel inductive signal, and improves the metal recognition accuracy and stability in a complex environment.

Description

Technical Field

[0001] This invention belongs to the field of metal detection technology, and specifically relates to a metal detection device and method. Background Technology

[0002] In coal mine production, raw coal is typically transported continuously using equipment such as belt conveyors, transfer conveyors, and crushers. During mining, support, transfer, and crushing operations, metallic foreign objects such as anchor bolts, large steel pieces, and tool fragments may become mixed into the coal flow. Once these metallic foreign objects enter the conveying system with the coal flow, they can easily scratch or tear the conveyor belt, causing problems such as belt misalignment and downtime for maintenance. When metallic foreign objects enter crushers or transfer conveyors, they can also cause equipment jamming, damage to cutting teeth, and even safety accidents. Current methods for detecting metallic foreign objects in coal mine transportation primarily rely on manual inspection or magnetic separation for iron removal.

[0003] Manual inspection relies primarily on workers observing the coal flow and identifying obvious foreign objects. This method is greatly affected by personnel experience, ambient lighting, coal flow obstruction, and conveying speed, making it difficult to promptly detect metallic foreign objects covered by coal lumps. Magnetic separation for iron removal typically uses permanent magnets or electromagnetic separators to adsorb ferromagnetic metals in the coal flow, effectively removing some ferromagnetic foreign objects. However, its ability to handle non-ferromagnetic metals, weakly magnetic metals, or metal targets deeply embedded in the coal flow is limited. Furthermore, this method focuses more on the adsorption and removal of metallic foreign objects, making it difficult to achieve stable detection and identification of them.

[0004] Therefore, in view of the current problems of limited coverage, insufficient detection sensitivity, weak anti-interference ability, and a large number of false alarms and missed alarms in the continuous transportation of metal foreign objects in coal mine conveyor belts, this invention aims to propose a metal detection device and method suitable for coal mine transportation scenarios, so as to improve the stability and reliability of online detection of metal foreign objects in coal mine transportation.

[0005] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0006] The purpose of this invention is to provide a metal detection device that improves the sensitivity of metal detection, expands the detection coverage, and acquires multi-channel response information with spatial differences, while ensuring the stability and real-time performance of data acquisition.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A metal detection device includes multiple sets of resonant sensing units, multiple slave microcontrollers, a master microcontroller, and a host computer; Each resonant sensing unit includes a dual-channel inductor sensor chip and two LC resonant circuits; each LC resonant circuit consists of a set of detection coils and a fixed capacitor. Each LC resonant circuit is used to collect the inductance change caused by a metallic target approaching the detection area; Each dual-channel inductor sensor chip is connected to two LC resonant circuits of the same set of resonant sensing units and is used to convert inductance changes into two digital response signals, DATA0 and DATA1. Multiple sets of resonant sensing units are arranged horizontally at equal intervals along the detection area to form a linear array of inductive sensors. Each slave microcontroller establishes a communication connection with a corresponding resonant sensing unit and reads the two digital response signals DATA0 and DATA1 output by the corresponding resonant sensing unit. The master microcontroller is connected to multiple slave microcontrollers and controls the data acquisition of the multiple slave microcontrollers through polling; the host computer is connected to the slave microcontrollers for communication and is used to receive dual-channel digital response signals DATA0 and DATA1.

[0008] Furthermore, based on the aforementioned metal detection device, this invention also proposes a corresponding metal detection method, which is implemented based on the aforementioned metal detection device, and its specific technical solution is as follows: A metal detection method includes the following steps: Step 1. Obtain the raw inductance detection data, preprocess the raw data, and construct the training dataset; First, the missing value interpolation, moving mean filtering, median filtering, local baseline normalization, and sliding window division are performed on the dual-channel digital response signals of DATA0 and DATA1 to obtain window-level dual-channel detection samples. Then, based on the window-level dual-channel detection samples, further model training data is constructed. The model training data includes the CNN input feature matrix, asynchronous event feature sequence, graph node feature matrix, and statistical feature vector. The CNN input feature matrix includes the response of the DATA0 channel, the response of the DATA1 channel, the temporal difference of DATA0, the temporal difference of DATA1, and the spatial difference between DATA1 and DATA0; The asynchronous event feature sequence is obtained by expanding the two channels DATA0 and DATA1 according to the sampling time and channel order; The graph node feature matrix uses the two detection channels DATA0 and DATA1 as graph nodes; Step 2. Build a metal detection classification model based on asynchronous event spatiotemporal graph multi-branch fusion, including a multi-branch feature extraction backbone network, a multi-branch feature fusion network, and a classifier; The multi-branch feature extraction backbone network adopts a four-branch structure, namely, the CNN temporal feature extraction branch, the asynchronous event GRU temporal feature extraction branch, the two-node graph convolutional spatial feature extraction branch, and the statistical feature extraction branch; A multi-branch feature fusion network is used to concatenate the CNN temporal features, asynchronous event GRU temporal features, two-node graph convolutional spatial features, and statistical enhancement features output by each branch to obtain fused features; The fused features are input into the classifier, and a binary classification is performed on the presence or absence of metal using a fully connected network and a sigmoid function; Step 3. Train the built model based on the training data, and use the trained model to achieve metal detection.

[0009] The present invention has the following advantages: As described above, this invention relates to a metal detection device and method. The device comprises multiple sets of resonant sensing units, multiple slave microcontrollers, a master microcontroller, and a host computer. This invention uses multiple sets of resonant sensing units to form a linear array of inductive sensors, thereby creating multiple effective detection channels, expanding the detection coverage, and acquiring multi-channel response information with spatial differences. Furthermore, each resonant sensing unit combines a dual-channel inductive sensor chip with an LC resonant circuit, converting minute inductance changes caused by metal targets into high-resolution digital response signals, thus improving the sensitivity of metal detection. In addition, this invention employs a data acquisition structure combining distributed acquisition by slave microcontrollers and polling scheduling by the master microcontroller. The master microcontroller uniformly schedules the data acquisition tasks of multiple slave microcontrollers, avoiding data conflicts caused by simultaneous uploads from multiple acquisition nodes, ensuring a stable temporal order for multi-channel detection data, and reducing the data acquisition burden on a single controller, thereby improving the stability and real-time performance of multi-channel data acquisition. Furthermore, the method of this invention constructs a metal detection classification model based on asynchronous event spatiotemporal graph multi-branch fusion. In terms of model preprocessing, this invention does not hard-delete local abrupt signals during data processing. Instead, it uses local baseline normalization to preserve transient response features caused by abnormal metal targets or coil states, which helps reduce the risk of falsely rejecting true metal responses. Regarding model structure, it integrates CNN temporal features, GRU asynchronous event temporal features, GCN dual-node spatial interaction features, and statistical features. This allows for the simultaneous utilization of local perturbations, temporal dynamics, channel spatial correlation, and stable statistical information of dual-channel inductor signals, improving the accuracy and stability of metal identification in complex environments, exhibiting strong anti-interference capabilities, and reducing false negatives and false positives. Attached Figure Description

[0010] Figure 1This is a schematic diagram of the control system structure of the metal detection device in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the detection coil in the metal detection device of Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the electrical and physical principle of the metal detection device in Embodiment 1 of the present invention; Figure 4 This is a network structure diagram of the metal detection model in Embodiment 2 of the present invention; Figure 5 This is a graph showing the loss function of the deep learning model for the metal detection model in Embodiment 2 of the present invention. Figure 6 This is a graph showing the performance evaluation results of the metal detection model in Embodiment 2 of the present invention; Figure 7 A schematic diagram of the dual-channel waveform and probability output for metal detection using data from the first set of resonant sensing units; Figure 8 A schematic diagram of the dual-channel waveform and probability output for metal detection using data from the second set of resonant sensing units; Figure 9 A schematic diagram of the dual-channel waveform and probability output for metal detection using data from the third set of resonant sensing units; Figure 10 A schematic diagram of the dual-channel waveform and probability output for metal detection using data from the fourth set of resonant sensing units; Figure 11 A schematic diagram of the dual-channel waveform and probability output for metal detection using data from the fifth group of resonant sensing units; Figure 12 A schematic diagram of the dual-channel waveform and probability output for metal detection using data from the sixth group of resonant sensing units; Figure 13 A schematic diagram of the dual-channel waveform and probability output for metal detection using data from the seventh group of resonant sensing units; Figure 14 A schematic diagram of the dual-channel waveform and probability output for metal detection using data from the eighth group of resonant sensing units; 1-PCB board, 2-detection coil, 3-dual-channel inductive sensor chip. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: To address the problems of insufficient detection sensitivity, inadequate utilization of channel information, unstable threshold determination, numerous false alarms and missed alarms, and difficulty in adapting to multi-channel continuous detection scenarios in complex environments, this invention proposes a metal detection device and method. Specifically, the device utilizes an LC resonant circuit and a dual-channel inductive sensor chip to collect inductance changes caused by metal targets, thereby improving the detection sensitivity of metal disturbances; it expands the detection coverage area by using a linear array of inductive sensors formed by multiple sets of resonant sensing units; it achieves continuous and stable acquisition of multi-channel data by polling and scheduling multiple slave microcontrollers through a master microcontroller; it fully utilizes the differences and correlations between channels through dual-channel responses (DATA0 and DATA1), temporal difference, spatial difference, and graph space feature extraction; and it reduces reliance on a single threshold determination through signal preprocessing, sliding window construction, and metal detection classification model construction, thus effectively reducing false alarms and missed alarms.

[0012] Example 1 like Figure 1 As shown, the metal detection device in this embodiment includes multiple sets of resonant sensing units, multiple slave microcontrollers, a master microcontroller, and a host computer.

[0013] Each resonant sensing unit includes a dual-channel inductor sensor chip and two LC resonant circuits, each of which consists of a detection coil and a fixed capacitor.

[0014] Each LC resonant circuit is used to collect the inductance change caused by a metallic target approaching the detection area.

[0015] Specifically, when an alternating current flows through the detection coil, an alternating magnetic field can be generated around the coil.

[0016] When a metal target enters the detection area, induced eddy currents are generated inside the metal target. These induced eddy currents affect the magnetic field distribution of the coil in the reverse direction, causing a change in the equivalent inductance of the detection coil, which in turn causes a change in the LC resonant frequency.

[0017] The dual-channel inductor sensor chip is used to acquire changes in resonant frequency and output the corresponding digital response signal.

[0018] Each dual-channel inductor sensor chip is connected to two LC resonant circuits of the same set of resonant sensing units, and is used to convert inductance changes into two digital response signals, DATA0 and DATA1.

[0019] In this embodiment, the dual-channel inductor sensor chip 3 can be an LDC1612 inductor-to-digital converter chip.

[0020] Each LDC1612 inductor-to-digital converter chip is connected to two LC resonant circuits of the same group of resonant sensing units, and outputs dual-channel inductor detection data through the DATA0 and DATA1 registers respectively.

[0021] The DATA0 and DATA1 channels of the same LDC1612 inductor-to-digital converter chip together form a dual-channel detection unit, which is used to acquire dual-channel inductor response data within the same detection area.

[0022] Multiple sets of resonant sensing units are arranged horizontally and equidistantly along the detection area to form a linear array of inductive sensors.

[0023] In a preferred embodiment, the linear array of inductive sensors includes eight LDC1612 inductive-to-digital converter chips, each of which is connected to two sets of detection coils, forming a total of 16 effective detection channels.

[0024] In this embodiment, multiple LDC1612 inductor-to-digital converter chips are arranged in an array, which can expand the detection coverage and thus obtain multi-channel inductor response data with spatial differences.

[0025] Each slave microcontroller establishes a communication connection with a corresponding resonant sensing unit and reads the two digital response signals DATA0 and DATA1 output by the corresponding resonant sensing unit.

[0026] The slave microcontrollers are used to control the resonant sensing unit to complete data acquisition. Each slave microcontroller is connected to the corresponding LDC1612 inductor-to-digital converter chip of the unit via the I²C bus to read the digital detection data output from the register.

[0027] In a preferred embodiment, the slave microcontroller is an STM32F446RET6. Each slave microcontroller is responsible for reading, buffering, and uploading data from one LDC1612 inductor-to-digital converter chip.

[0028] The dual-channel inductance detection data collected by the slave microcontroller is uploaded to the host computer via TTL serial port. The uploaded data includes at least the detection channel number, sampling time sequence number, DATA0 response value and DATA1 response value.

[0029] The host computer sorts, saves, and displays multi-channel data based on the channel number and sampling time sequence.

[0030] The master microcontroller is connected to multiple slave microcontrollers and controls the data acquisition of the multiple slave microcontrollers through polling, controlling the acquisition order of the multiple slave microcontrollers to achieve coordinated acquisition of the linear array inductive sensor array.

[0031] The master microcontroller also uses STM32F446RET6. The master microcontroller is used to uniformly schedule multiple slave microcontrollers and control the start, stop and acquisition sequence of multiple slave microcontrollers through polling.

[0032] In this embodiment, the master control microcontroller sends control commands to each slave control microcontroller according to a preset sampling period, so that each slave control microcontroller sequentially completes the coordinated acquisition task of the linear array inductive sensor array.

[0033] By using a master control polling scheduling method, interference from multiple coils working simultaneously can be effectively avoided.

[0034] The host computer communicates with the slave microcontroller to receive dual-channel digital response signals DATA0 and DATA1.

[0035] The host computer is used to receive, parse, and save the dual-channel inductance detection data uploaded by the slave microcontroller, and can realize functions such as data acquisition, sampling control, data storage, and model inference.

[0036] like Figure 2 As shown, in this embodiment, the detection coil adopts a double-layer PCB wiring structure.

[0037] That is, two detection coils 2 with the same structure are set on the positive and negative sides of the PCB board 1 (shown in red and blue respectively), and the two detection coils on the positive and negative sides of the PCB board are connected in series to form the entire detection coil.

[0038] like Figure 2 The left side shown illustrates a set of detection coils, which consists of a detection coil 2 located on the positive side of the PCB board and a detection coil located on the negative side of the PCB board. Similarly, the set of detection coils on the right side has the same structure.

[0039] Since the detection coil on the negative side of the PCB board (indicated in blue) has the same structural form as the detection coil on the positive side of the PCB board, this embodiment only shows the structure of the detection coil on the positive side of the PCB board.

[0040] like Figure 2 As shown, each detection coil adopts a stretched rectangular planar spiral coil structure.

[0041] The long side of the stretched rectangular planar helical coil is larger than the short side, which is used to form an asymmetric magnetic field distribution along the long axis of each detection coil segment, so that the metal target moves along the detection direction and generates a continuously changing inductive response.

[0042] The structure of the stretched rectangular planar helical coil is as follows: Starting from the inner starting point of each detection coil (e.g., point A), proceed clockwise (e.g., ...). Figure 2 As shown, you can also wind the coil from the inside out in a counter-clockwise direction to form a multi-turn rectangular coil (e.g., a 13-turn coil).

[0043] The long side of each rectangular coil is the same as the major axis of each detection coil segment. The inner starting point of each detection coil segment (i.e.,...) Figure 2 The lower end of the coil (with adjacent turns of coil arranged closely together) is defined as the densely wound end.

[0044] Conversely, the other end opposite to the inner starting point of each detection coil (i.e. Figure 2 The upper middle part is the sparsely wound end.

[0045] It should be noted that the inner starting point of each detection coil segment here is only for the convenience of describing the stretched rectangular planar spiral coil structure. In reality, this inner starting point is the connection point connecting the two detection coil segments on the positive and negative sides of the PCB.

[0046] Each detection coil consists of two sections connected on the positive and negative sides of the PCB. The ends of one section of the detection coil on the positive side of the PCB and the end of one section of the detection coil on the negative side of the PCB together form the two connection points of the entire detection coil.

[0047] Each detection coil is connected to a specific LC resonant circuit through these two connection points.

[0048] By stretching a rectangular structure to form an asymmetric magnetic field distributed along the long axis of the coil, the coupled magnetic flux of the metal target changes continuously with its position as it moves along the detection direction, thereby enhancing the directionality and stability of the inductive response.

[0049] In this embodiment, the detection coil is made of copper foil material, and each section of the detection coil has 13 turns and a line width of 1.5 mm. It also adopts a double-layer PCB wiring structure to improve the magnetic field coverage and inductive response sensitivity of the detection coil.

[0050] In the linear array of inductive sensors, the overall arrangement of each resonant sensing unit is perpendicular to the material transport direction, and the long axis of each detection coil is parallel to the material transport direction.

[0051] During transportation, the material first passes through the densely wound end of each detection coil segment, and then through the sparsely wound end.

[0052] In this embodiment, the detection coil adopts a double-layer structure with the positive and negative sides of the PCB connected in series, which can increase the equivalent number of turns and the sensing range within a limited area, thereby improving the magnetic field coverage and detection sensitivity.

[0053] In addition, each detection coil adopts a stretched rectangular planar spiral structure, and the long axis is consistent with the material transport direction, so that the metal target can continuously pass through the detection area as the material moves, forming a continuously changing inductive response signal.

[0054] The detection coil is configured with densely wound ends and sparsely wound ends, which can form an asymmetric magnetic field distribution. This causes the metal target to produce a directional and transitional response change when it passes through the coil, which helps to improve the stability of metal target detection under continuous motion.

[0055] To describe the equivalent inductance of a stretched rectangular planar helical coil, the equivalent average diameter method is used to model the coil structure.

[0056] Let the maximum external dimension of the outermost conductor of the coil along its length be... The maximum external dimension along the width direction is The innermost conductor has a length dimension of [missing value]. The dimension along the width direction is .

[0057] Then the equivalent average diameter It can be represented as: .

[0058] Fill rate for: .in, Indicates the external equivalent size. Indicates the internal equivalent size. Used to characterize coil area utilization and inter-turn coupling strength.

[0059] The inductance value can be expressed using the modified Wheeler formula as follows: ; in Indicates the coil inductance value. Represents the permeability of free space. Indicates the number of coil turns. Indicates the equivalent average diameter. The value represents the fill rate, and 2.34 and 2.75 are empirical fitting coefficients.

[0060] When a metallic target approaches the detection coil, the electromagnetic coupling between the target and the coil changes, causing the equivalent inductance of the detection coil to vary with the position of the metallic target. Changes are recorded as .

[0061] The detection coil and the fixed capacitor C form an LC resonant circuit, with the corresponding resonant frequency being... It can be represented as: ; in The capacitance value is the value of the capacitor connected in parallel with the inductor coil. Indicates the location of the metal target. The equivalent inductance value of the detection coil is measured at the time. Indicates the location of the metal target. The resonant frequency of the LC resonant circuit.

[0062] As attached Figure 3 As shown, when the position of the metal target changes, the equivalent inductance of the detection coil... This change, in turn, causes the resonant frequency to change. The frequency changes. The LDC1612 inductor-to-digital converter chip is used to detect this change in resonant frequency.

[0063] The LC resonant sensing method described above can convert the minute inductance changes caused by the metal target into a digital response signal that can be acquired and processed, providing stable dual-channel input data for subsequent metal target identification.

[0064] The present invention uses a detection coil and a fixed capacitor to form an LC resonant circuit. When a metal target approaches, it causes a change in inductance, which is then converted into dual-channel digital response signals DATA0 and DATA1 by the LDC1612 inductance sensor chip.

[0065] Multiple LDC1612 inductor-to-digital converter chips are arranged horizontally and equidistantly to form a multi-channel detection array. The slave microcontroller collects data through the I²C bus and uploads it to the host computer through the TTL serial port. The master microcontroller coordinates the data acquisition task through polling.

[0066] The device of the present invention, through the above design, can obtain stable, continuous and spatially differentiated multi-channel input data in actual metal detection scenarios, providing a reliable data basis for the identification and judgment of the metal detection method in Example 2.

[0067] Example 2 This embodiment 2 describes a metal detection method, which is based on the same inventive concept as the metal detection device in embodiment 1. The metal detection method in this embodiment includes the following steps: Step 1. Obtain the raw inductance detection data, preprocess the raw data, and construct the training dataset.

[0068] First, the missing value interpolation, moving mean filtering, median filtering, local baseline normalization, and sliding window division are performed on the dual-channel digital response signals of DATA0 and DATA1 to obtain window-level dual-channel detection samples.

[0069] The raw inductance detection data is acquired by a linear array of inductance sensors consisting of multiple resonant sensing units.

[0070] A dual-channel detection unit corresponds to a set of resonant sensing units. Each dual-channel detection unit is defined as having two detection channels, denoted as DATA0 and DATA1 respectively. Let the... The original input signal sequence of the resonant sensing unit is: ;in, Indicates the total number of sampling points. This indicates that channel DATA0 is in the... The response value of each sampling point This indicates that channel DATA1 is in the... The response value of each sampling point.

[0071] Step 1.1. First, perform linear interpolation on the missing sample values ​​of the two channels in the original input signal sequence.

[0072] set up and If the valid sampled values ​​at both ends of the missing interval are the missing sample points, then... interpolation results for: .

[0073] in , They represent and The effective sampled value at that location, Indicates the first One channel.

[0074] in, This indicates the DATA0 channel. This indicates the DATA1 channel.

[0075] Step 1.2. Perform moving mean filtering and median filtering on the linearly interpolated DATA0 and DATA1 channels respectively to reduce random sampling noise and isolated spike noise.

[0076] To avoid the accidental rejection of local abrupt responses caused by abnormal metal targets or coil conditions, instead of hard-deleting local abrupt signals, local baseline normalization is applied to both channels to preserve their transient response characteristics.

[0077] Local baseline normalization is applied to both channels, and its expression is as follows: .

[0078] in Indicates the first Group resonant sensing unit, the first Each channel at the sampling point The response value after local baseline normalization, To prevent extremely small constants with a denominator of zero, for example, take... .

[0079] and They represent respectively with The mean and standard deviation of the signal within a local window centered on the signal.

[0080] Step 1.3. Then, a sliding window is used to construct training samples.

[0081] Window length is The sliding step size is Then the first The starting point of each window and termination point They are respectively: .

[0082] In this embodiment, the window length can be set to, for example, [the following settings are available]. The sliding step size can be set to, for example, .

[0083] Corresponding window-level dual-channel detection samples Represented as: ; Define point-level labels as ,in Indicates the first The resonant sensing unit at the sampling point Does a metallic response exist? It indicates that it exists. Indicates nothing; then the first The proportion of metal points within each window is: .

[0084] Let the threshold for the proportion of positive samples in the window be... Then window label Defined as: .

[0085] in Indicates the first Group resonant sensing unit Whether a sliding window is considered a positive sample.

[0086] In a preferred embodiment, the window positive sample ratio threshold The above method converts point-level metal target or coil anomaly labels into window-level training labels for subsequent model training.

[0087] Then, based on the window-level dual-channel detection samples, further model training data is constructed. The model training data includes the CNN input feature matrix, asynchronous event feature sequence, graph node feature matrix, and statistical feature vector.

[0088] The CNN input feature matrix includes the response of the DATA0 channel, the response of the DATA1 channel, the temporal difference of DATA0, the temporal difference of DATA1, and the spatial difference between DATA1 and DATA0.

[0089] The asynchronous event feature sequence is obtained by expanding the two channels DATA0 and DATA1 according to the sampling time and channel order.

[0090] The graph node feature matrix uses the two detection channels DATA0 and DATA1 as graph nodes.

[0091] The statistical feature vector includes single-channel statistical features, time-difference statistical features, frequency domain energy features, dual-channel spatial difference statistical features, and dual-channel response distribution features.

[0092] I. The specific process of constructing the CNN input feature matrix is ​​as follows: For the preprocessed first The first group of resonant sensing units Sample of each window Calculate the time difference of DATA0 channel, the time difference of DATA1 channel, and the spatial difference of the two channels respectively.

[0093] Set the window inside the first The original sampling locations corresponding to each sampling point are: .

[0094] in, Indicates the first The sampling start point of the nth window, then the nth The formula for calculating the time difference of each channel is as follows: .

[0095] in Indicates the first The time difference results for each channel, abbreviated as: .

[0096] Indicates the first Each channel at the current sampling position The normalized response value after preprocessing. Indicates the first Each channel is at the previous sampling position, i.e., the sampling position. The normalized response value after preprocessing.

[0097] The formula for calculating the spatial difference of the two channels is as follows: ; in Indicates the sampling position of channel DATA1. The normalized response value at that point, This indicates that the DATA0 channel is at the sampling position. The normalized response value at the location.

[0098] Indicates the first The resonant sensing unit at the sampling position The two-channel spatial difference result at that point is abbreviated as .

[0099] Then, construct the CNN input feature matrix for the one-dimensional convolutional branch. : ; in 5 represents the number of input feature channels, and 16 represents the number of sampling points within the window.

[0100] The input feature matrix of this CNN contains the response of the DATA0 channel. DATA1 channel response DATA0 Time Difference Results DATA1 Time Difference Results And the two-channel spatial difference results of DATA and DATA0 .

[0101] II. The process of constructing the asynchronous event feature sequence is as follows: Meanwhile, the window-level dual-channel detection samples are first arranged in the order of sampling point time, and then arranged in the order of DATA0 and DATA1 within each sampling point to expand the window-level dual-channel detection samples into an asynchronous event sequence.

[0102] For the The first window The first channel Each sampling point defines the expanded event sequence number. for: ; where 32 refers to arranging the dual-channel data in a window into a string in sequence, and then the number of each data point in this string is q, and the value of q is 1-32.

[0103] in This indicates the DATA0 channel. This indicates the DATA1 channel.

[0104] Since the window length in this embodiment is 16, and each sampling point contains two channels, DATA0 and DATA1, the number of events in a window is 16×2=32 after the dual-channel data is expanded according to the "sampling time order + channel order".

[0105] Then the first Numerical characteristics of an event Represented as: .

[0106] in Indicates the first The group resonant sensing unit in the first The first window Numerical feature vectors of events; Indicates the first The first event corresponding to the Each detection channel is at the sampling position. The normalized response value at the location.

[0107] It represents the global time interval feature, used to characterize the overall temporal order relationship between the current event and the previous event in the expanded asynchronous event sequence; This is a local time interval feature used to characterize the time interval relationship between the current response and the previous response within the same detection channel; Encode the channel position.

[0108] Channel number Input channel embedding layer to obtain channel embedding vector Then embed the channel into the vector. Event numerical characteristics By concatenating the vectors, we obtain the feature vector of the asynchronous event. The formula is as follows: .

[0109] in This indicates a concatenation operation; it combines the feature vectors of all asynchronous events within the window. Arranged according to sampling time and channel order, the asynchronous event feature sequence is obtained. The formula is as follows: .

[0110] III. The process of obtaining the feature matrix of graph nodes is as follows: For the GCN graph convolution branch, a separate graph node feature matrix is ​​constructed. When a metallic target approaches the detection area, the DATA0 and DATA1 channels will produce responses with temporal correlation and spatial differences.

[0111] Therefore, the DATA0 and DATA1 detection channels are respectively used as graph nodes. and Construct a two-node linear array.

[0112] No. At the sampling point, the first The graph node characteristics of each channel node are represented as follows: ; in Indicates the first Each channel at the sampling position The normalized response value at that point, This indicates that the channel is at the sampling position. Time difference results; This indicates the relative offset of the channel polling.

[0113] Indicates the first Channel position encoding for each detection channel. To enable the model to distinguish the response sources of different channels, this invention assigns fixed channel position encodings to DATA0 and DATA1 respectively.

[0114] This yields the graph node feature matrix. ; The superscript 2 indicates that there are two nodes in the diagram, namely the two detection channels DATA0 and DATA1; the superscript 4 indicates that each node has four features, namely the normalized response value, the temporal difference result, the channel polling relative offset, and the channel position encoding.

[0115] The adjacency matrix between the two channel nodes DATA0 and DATA1 is as follows: Diagonal elements indicate self-connection of nodes, while off-diagonal elements indicate a spatial association between DATA0 and DATA1.

[0116] For adjacency matrix Normalization is performed to obtain the normalized adjacency matrix. : ; in, Representing the adjacency matrix The corresponding degree matrix.

[0117] and normalized adjacency matrix Both serve as inputs to the two-node graph convolution spatial feature extraction branch.

[0118] The graph node feature matrix is ​​used to represent the response features of the two detection channels DATA0 and DATA1 within the window, and the normalized adjacency matrix... Used to indicate the connection relationship between the two channels DATA0 and DATA1.

[0119] Both serve as inputs to the two-node graph convolutional spatial feature extraction branch, enabling each channel node to fuse the response information of the other channel when updating its own features, thereby extracting the spatial correlation and interaction features between the two detection channels.

[0120] IV. Extract statistical feature vectors for each window of samples.

[0121] The statistical feature vector includes the single-channel statistical features, time difference statistical features, frequency domain energy features, dual-channel spatial difference statistical features, and dual-channel response distribution features of the two channels DATA0 and DATA1.

[0122] Therefore, the first The first group of resonant sensing units Model input in one window Represented as: .

[0123] Step 2. Build a metal detection classification model based on asynchronous event spatiotemporal graph multi-branch fusion, such as... Figure 4 As shown, the model includes a multi-branch feature extraction backbone network, a multi-branch feature fusion network, and a classifier.

[0124] The multi-branch feature extraction backbone network adopts a four-branch structure, namely, the CNN temporal feature extraction branch, the asynchronous event GRU temporal feature extraction branch, the two-node graph convolutional spatial feature extraction branch, and the statistical feature extraction branch.

[0125] The CNN temporal feature extraction branch is used to extract local perturbation features caused by metal targets or coil anomalies from the CNN input feature matrix.

[0126] The input feature matrix of this CNN is a five-channel input feature matrix consisting of the response of DATA0 channel, the response of DATA1 channel, the time difference of DATA0, the time difference of DATA1, and the spatial difference between DATA1 and DATA0.

[0127] The CNN temporal feature extraction branch receives and is used to extract local temporal perturbation features from the dual-channel inductor signal.

[0128] First receive Initial features are extracted through one-dimensional convolution, batch normalization, and nonlinear activation operations to obtain initial features. : .

[0129] in This represents a one-dimensional convolution operation. This indicates a batch normalization operation. This represents a non-linear activation function.

[0130] Subsequently, the initial feature map is input into a multi-scale one-dimensional convolutional network for deep feature extraction, resulting in deep temporal features. .

[0131] The kernel size of some one-dimensional convolutional layers in the multi-scale one-dimensional convolutional network is 5, which extracts local perturbation features over a wider time range; the kernel size of the remaining one-dimensional convolutional layers is 3, which extracts detailed variation features between adjacent sampling points.

[0132] right Global average pooling and global max pooling are performed separately, and the two are concatenated to obtain the CNN branch features, as shown in the following formula: .

[0133] in, This represents the temporal features of a CNN.

[0134] The asynchronous event GRU temporal feature extraction branch expands the DATA0 and DATA1 dual-channel window samples into an asynchronous event sequence according to the sampling time and channel order. Combined with channel embedding information, it models the temporal relationship, local mutation process, and dual-channel response changes of the metal target passing through the detection area through the gated recurrent unit.

[0135] The asynchronous event GRU time-series feature extraction branch is used to model the temporal sequence, channel polling relationship, and local mutation persistence process in a dual-channel asynchronous event sequence.

[0136] First, the asynchronous event feature sequence is received, and then the asynchronous event features are mapped to a unified latent space through a linear mapping layer: ; in Indicates the first Mapping characteristics of an event and These are learnable parameters.

[0137] Subsequently, a lightweight gated loop encoder is used to encode the 32 asynchronous event feature vectors obtained from the window expansion, outputting a set of hidden state vectors. In this embodiment, the window length L=16, and each sampling point contains two channels, DATA0 and DATA1. Therefore, the length of the expanded asynchronous event sequence is 2L=32.

[0138] .

[0139] in Indicates the first The first group of resonant sensing units In each window, the hidden state feature vectors are obtained from 32 asynchronous events after GRU time-series feature extraction.

[0140] It should be noted that the calculation process of the lightweight gated cyclic encoder is relatively conventional and will not be elaborated here.

[0141] The hidden state vector set is subjected to average pooling and max pooling respectively, and then concatenated to obtain the asynchronous event time sequence feature. : .

[0142] The two-node graph convolution spatial feature extraction branch uses DATA0 and DATA1 channels as graph nodes, and constructs an adjacency matrix based on their spatial correspondence in the same dual-channel inductor sensor chip.

[0143] Spatial interaction features between the two channels DATA0 and DATA1 are extracted using a graph convolutional network.

[0144] The two-node graph convolution spatial feature extraction branch is used to model the spatial interaction relationship between the two channel nodes DATA0 and DATA1.

[0145] Let the initial graph node features be: ; No. The formula for calculating layer graph convolution is as follows: ; in Represents a non-linear activation function. Indicates the first The learnable weight matrix of layer graph convolution. The normalized adjacency matrix is ​​represented; the graph convolution output, after global average pooling in both the time and node dimensions, yields the spatial interaction feature vector: ; in, Indicates the number of convolutional layers in the graph. This indicates global average pooling.

[0146] The statistical feature extraction branch is used to extract stability discrimination information such as single-channel statistical features, time difference statistical features, frequency domain energy features, dual-channel spatial difference statistical features, and dual-channel response distribution features from the window samples.

[0147] The characteristics of the dual-channel response distribution include the response center, response width, asymmetry, and gradient energy.

[0148] The statistical feature extraction branch is used to supplement stable discriminative information beyond deep features.

[0149] For each window sample, let the statistical feature vector be... The mapping process is as follows: .

[0150] in, Represents a nonlinear mapping function. and For learnable parameters, This represents a statistically enhanced feature vector.

[0151] A multi-branch feature fusion network is used to concatenate the CNN temporal features, asynchronous event GRU temporal features, two-node graph convolutional spatial features, and statistical enhancement features output from each branch to obtain the fused feature, as shown in the following formula: .

[0152] The fused features are input into the classifier, which performs binary classification on the presence or absence of metal using a fully connected network and a sigmoid function.

[0153] When the probability of metal presence at the window level is not lower than a set threshold, it is determined that a metal target exists in the sampling interval corresponding to the current window; the threshold can be set to 0.85 in this invention. The fused features are processed through a multi-layer fully connected network to obtain a classification score. : .

[0154] Finally, the window-level metal presence probability is obtained using the Sigmoid function: ; in Indicates the first The first group of resonant sensing units The probability of a metallic target existing within a sliding window. When If the probability is not lower than the preset probability threshold of 0.85, it is determined that there is a metal target in the current window.

[0155] Otherwise, when When the probability is below the preset probability threshold of 0.85, the current window is determined to be in a metal-free state.

[0156] Step 3. Train the built model based on the training data, and use the trained model to achieve metal detection.

[0157] In a preferred embodiment, the model is trained using a binary cross-entropy loss function with positive sample weights to alleviate the imbalance between the number of metal-free windows and the number of metal-containing windows.

[0158] In this embodiment, positive samples refer to training samples with window-level labels. The training samples.

[0159] During model training, a binary cross-entropy loss function with positive sample weights is used. Optimize the model: ; in Indicates the total number of samples in the window. Indicates the first Group resonant sensing unit The actual label of the window This represents the probability of the model indicating the presence of metal in the window if it falls under the metal-containing state. Represents the positive sample weight coefficient. .

[0160] In the training phase of the metal detection classification model based on asynchronous event spatiotemporal graph multi-branch fusion, the Adam W optimizer is used to iteratively update the network parameters, and the model performance is evaluated in conjunction with the validation set.

[0161] Depend on Figure 5 As shown, with the increase of training rounds, the training loss and validation loss generally show a downward trend and gradually stabilize, indicating that the model can effectively learn the features in the dual-channel inductance detection data and the training process has good convergence.

[0162] Depend on Figure 6 As shown, the model maintains a high level of accuracy, precision, recall, and F1 score on the validation set, indicating that the model has a good ability to distinguish between metal target windows and non-metal windows, and can reduce false positives while reducing the risk of false negatives.

[0163] During the inference phase, if the probability of metal presence at the window level is not lower than a set threshold, it is determined that a metal target exists in the sampling interval corresponding to that window.

[0164] The host computer merges the sampling intervals corresponding to consecutive or adjacent high-probability windows to form metal detection intervals, and marks the metal detection intervals in the waveform display interface.

[0165] When any group of resonant sensing units has a metal detection range, the output shows a metal result; when none of the resonant sensing units form a metal detection range, the output shows a no-metal result.

[0166] Let the first The set of metal detection intervals in the detection data sequence acquired by the resonant sensing unit is as follows: ; in, Indicates the first The set of metal detection intervals in the data sequence to be detected acquired by the group of resonant sensing units.

[0167] This represents the total number of sliding windows in the data sequence to be detected. This means merging the sampling intervals corresponding to consecutive or adjacent high-probability windows to form a set of metal detection intervals.

[0168] Indicates the first The first group of data sequences to be detected The starting sampling point of a sliding window. Indicates the first The first group of data sequences to be detected The final sampling point of a sliding window.

[0169] when If the interval is not empty, a metal target is determined to exist; if no metal detection interval is generated, a result of no metal is output.

[0170] Overall judgment results Represented as: ; in This indicates that the final output contains metal. This indicates that the final output contains no metal.

[0171] Appendix Figures 7 to 14 The results show the detection results obtained after inputting a total of 8 sets of collected data from the first to the eighth resonant sensing units into the trained metal detection classification model. These include the dual-channel detection waveforms of DATA0 and DATA1, the window-level metal presence probability curve, the set threshold line, and the metal detection interval marked by the model.

[0172] From the appendix Figures 7 to 14 It can be seen that when there is a metal target in the collected data, the dual-channel inductor response of DATA0 and DATA1 exhibits local perturbation in the corresponding sampling interval. As a result, the window-level probability of metal presence in the model output increases accordingly, and after exceeding the set threshold, it forms continuous or adjacent high-probability windows.

[0173] This invention preprocesses dual-channel data, divides it into sliding windows, and constructs features. The data is then input into a metal detection classification model composed of CNN branches, GRU branches, graph convolution branches, and statistical feature branches. This model outputs a window-level probability of metal presence and generates detection results based on a set threshold. This invention fully utilizes the temporal perturbations, channel differences, and spatial correlation information of the dual-channel inductive signals to improve the stability and accuracy of metal detection.

[0174] Example 3 This embodiment 3 describes a computer device that includes a memory and one or more processors.

[0175] Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the metal detection method in Embodiment 2 above. Here, the computer device refers to the host computer in Embodiment 1 above.

[0176] After receiving the multi-channel inductance detection data, the host computer first organizes the data according to the channel number and sampling time, forming 8 sets of dual-channel data sequences DATA0 and DATA1. Then, it performs missing value interpolation, moving mean filtering, median filtering, and local baseline normalization on the 8 sets of dual-channel data respectively. Next, it constructs window samples according to the set window length and sliding step size, calls the metal detection classification model based on asynchronous event spatiotemporal graph multi-branch fusion, and outputs the window-level metal presence probability.

[0177] When the probability of metal presence at the window level is not lower than a set threshold, the sampling interval corresponding to that window is determined to contain a metal target. The host computer merges the sampling intervals corresponding to consecutive or adjacent high-probability windows into a metal detection interval and marks the metal interval detected by the model in red on the waveform graph. When the probability of metal presence in all windows is lower than the set threshold and no metal detection interval is formed, the system outputs a result indicating no metal.

[0178] The device of this invention can convert the resonant frequency changes caused by metal targets into continuous dual-channel digital response signals, forming DATA0 and DATA1 dual-channel detection data. After preprocessing by a host computer, the two detection data can be further constructed into a CNN input feature matrix, an asynchronous event feature sequence, a graph node feature matrix, and a statistical feature vector, which are then input into a metal detection classification model based on multi-branch fusion of asynchronous event spatiotemporal graphs. Thus, hardware acquisition and the deep learning classification model form a complete data closed loop, realizing a complete metal detection process from metal disturbance signal acquisition and model recognition to detection result output.

[0179] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0180] Example 4 This embodiment 4 describes a computer-readable storage medium storing a program that, when executed by a processor, is used to implement the steps of the metal detection method in embodiment 2 above.

[0181] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0182] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A metal detection method, implemented based on a metal detection device, characterized in that, The metal detection device includes: Multiple sets of resonant sensing units, multiple slave microcontrollers, master microcontroller and host computer; Each resonant sensing unit includes a dual-channel inductor sensor chip and two LC resonant circuits; each LC resonant circuit consists of a set of detection coils and a fixed capacitor. Each LC resonant circuit is used to collect the inductance change caused by a metallic target approaching the detection area; Each dual-channel inductor sensor chip is connected to two LC resonant circuits of the same set of resonant sensing units and is used to convert inductance changes into two digital response signals, DATA0 and DATA1. Multiple sets of resonant sensing units are arranged horizontally at equal intervals along the detection area to form a linear array of inductive sensors. Each slave microcontroller establishes a communication connection with a corresponding resonant sensing unit and reads the two digital response signals DATA0 and DATA1 output by the corresponding resonant sensing unit. The master microcontroller is connected to multiple slave microcontrollers and controls the data acquisition of the multiple slave microcontrollers through polling; the host computer is connected to the slave microcontrollers for receiving dual-channel digital response signals DATA0 and DATA1. The metal detection method includes the following steps: Step 1. Obtain the raw inductance detection data, preprocess the raw data, and construct the training dataset; First, the missing value interpolation, moving mean filtering, median filtering, local baseline normalization, and sliding window division are performed on the dual-channel digital response signals of DATA0 and DATA1 to obtain window-level dual-channel detection samples. Then, based on the window-level dual-channel detection samples, further model training data is constructed, including CNN input feature matrix, asynchronous event feature sequence, graph node feature matrix, and statistical feature vector; The statistical feature vector includes single-channel statistical features, time-difference statistical features, frequency domain energy features, dual-channel spatial difference statistical features, and dual-channel response distribution features; The CNN input feature matrix includes the response of the DATA0 channel, the response of the DATA1 channel, the temporal difference of DATA0, the temporal difference of DATA1, and the spatial difference between DATA1 and DATA0. The asynchronous event feature sequence is obtained by expanding the two channels DATA0 and DATA1 according to the sampling time and channel order; The graph node feature matrix uses the two detection channels DATA0 and DATA1 as graph nodes; Step 2. Build a metal detection classification model based on asynchronous event spatiotemporal graph multi-branch fusion. The metal detection classification model includes a multi-branch feature extraction backbone network, a multi-branch feature fusion network, and a classifier. The multi-branch feature extraction backbone network adopts a four-branch structure, namely, the CNN temporal feature extraction branch, the asynchronous event GRU temporal feature extraction branch, the two-node graph convolutional spatial feature extraction branch, and the statistical feature extraction branch; A multi-branch feature fusion network is used to concatenate the CNN temporal features, asynchronous event GRU temporal features, two-node graph convolutional spatial features, and statistical enhancement features output by each branch to obtain fused features; The fused features are input into the classifier, and a binary classification is performed on the presence or absence of metal using a fully connected network and a sigmoid function; Step 3. Train the built model based on the training data, and use the trained model to achieve metal detection.

2. The metal detection method according to claim 1, characterized in that, The dual-channel inductor sensor chip uses the LDC1612 inductor-to-digital converter chip; Each LDC1612 inductor-to-digital converter chip is connected to two LC resonant circuits of the same group of resonant sensing units, and outputs dual-channel inductor detection data through two registers respectively.

3. The metal detection method according to claim 1, characterized in that, The detection coil adopts a double-layer PCB wiring structure, that is, two detection coils with the same structure are set on the positive and negative sides of the PCB board, and the two detection coils on the positive and negative sides of the PCB board are connected in series to form the entire detection coil. Each detection coil adopts a stretched rectangular planar spiral coil structure; The long side of the stretched rectangular planar spiral coil is larger than the short side, which is used to form an asymmetric magnetic field distribution in the long axis direction of each detection coil segment, so that the metal target will generate a continuously changing inductive response when it moves along the detection direction.

4. The metal detection method according to claim 3, characterized in that, The structure of the stretched rectangular planar helical coil is as follows: Starting from the inner starting point of each detection coil, the coil is wound from the inside out in a clockwise or counterclockwise direction to form a multi-turn rectangular coil, and the long side of each rectangular coil is the long axis of each detection coil. The end where the inner starting point of each detection coil is located is defined as the densely wound end, where two adjacent turns of the coil are closely arranged. The other end opposite the end where the inner starting point of each detection coil is located is defined as the sparsely wound end.

5. The metal detection method according to claim 4, characterized in that, In the linear array of inductive sensors, the overall arrangement of each resonant sensing unit is perpendicular to the material transport direction, and the major axis of each detection coil is parallel to the material transport direction. During transportation, the material first passes through the densely wound end of each detection coil, and then through the sparsely wound end.

6. The metal detection method according to claim 1, characterized in that, In step 1, the process of acquiring window-level dual-channel detection samples is as follows: Each group of resonant sensing units is defined to include two detection channels, denoted as DATA0 and DATA1 respectively; Let the first The original input signal sequence of the resonant sensing unit for: ;in, Indicates the total number of sampling points. This indicates that channel DATA0 is in the... The response value of each sampling point. This indicates that channel DATA1 is in the... The response value of each sampling point; Step 1.

1. First, perform linear interpolation on the missing sample values ​​of the two channels in the original input signal sequence; let... and If the valid sampled values ​​at both ends of the missing interval are the missing sample points, then... interpolation results for: ； in , They represent and The effective sampled value at that location, Indicates the first One channel; in This indicates the DATA0 channel. Indicates the DATA1 channel; Step 1.

2. Apply moving mean filtering and median filtering to the linearly interpolated DATA0 and DATA1 channels respectively. Then, apply local baseline normalization to both channels, as shown in the following expression: ； in Indicates the first Group resonant sensing unit, the first Each channel at the sampling point The response value after local baseline normalization, To prevent extremely small constants with a denominator of zero; and They represent respectively with The mean and standard deviation of the signal within a local window centered on the signal; Step 1.

3. Then, a sliding window is used to construct training samples; Window length is The sliding step size is Then the first Sampling start point of each window and termination point They are respectively: ； Corresponding window-level dual-channel detection samples Represented as: ； Define point-level labels as ,in Indicates the first The resonant sensing unit at the sampling point Does a metallic response exist? It indicates that it exists. Indicates nothing; then the first The proportion of metal points within each window for: ； Let the threshold for the proportion of positive samples in the window be... Then window label Defined as: ； in Indicates the first Group resonant sensing unit Whether a sliding window is considered a positive sample indicates the presence of metal.

7. The metal detection method according to claim 6, characterized in that, In step 1, the specific construction process of the CNN input feature matrix is ​​as follows: For the preprocessed first The first group of resonant sensing units Sample of each window Calculate the time difference of DATA0 channel, the time difference of DATA1 channel, and the spatial difference of the two channels respectively; Set the window inside the first The original sampling location corresponding to each sampling point for: ； in Indicates the first The sampling start point of the nth window, then the nth The formula for calculating the time difference of each channel is as follows: ； in Indicates the first The time difference results for each channel, abbreviated as: ; Indicates the first Each channel at the current sampling position The normalized response value after preprocessing. Indicates the first Each channel is at the previous sampling position, i.e., the sampling position. The normalized response value after preprocessing; The formula for calculating the spatial difference of the two channels is as follows: ； in This indicates that the DATA1 channel is at the sampling position. The normalized response value at that point, This indicates that the DATA0 channel is at the sampling position. The normalized response value at the location; Indicates the first The resonant sensing unit at the sampling position The two-channel spatial difference result at that point is abbreviated as ; Then, construct the CNN input feature matrix for the one-dimensional convolutional branch. : ； in The superscript 5 indicates the number of input feature channels, and the superscript 16 indicates the number of sampling points within the window.

8. The metal detection method according to claim 6, characterized in that, In step 1, the process of constructing the asynchronous event feature sequence is as follows: First, arrange the sampling points in chronological order, and then arrange them in the order of DATA0 and DATA1 within each sampling point to expand the window-level dual-channel detection samples into an asynchronous event sequence; For the The first window The first channel Each sampling point defines the expanded event sequence number. for: ； in This indicates the DATA0 channel. Indicates the DATA1 channel; Then the first Numerical characteristics of an event Represented as: ； in Indicates the first The group resonant sensing unit in the first The first window Numerical feature vectors of events; Indicates the first Each detection channel is at the sampling position. The normalized response value at the location; It represents the global time interval feature, used to characterize the overall temporal order relationship between the current event and the previous event in the expanded asynchronous event sequence; This is a local time interval feature used to characterize the time interval relationship between the current response and the previous response within the same detection channel; Encode the channel position; Channel number Input channel embedding layer to obtain channel embedding vector Then Event numerical characteristics By concatenating the vectors, we obtain the feature vector of the asynchronous event. The formula is as follows: ； in This indicates a concatenation operation; it combines the feature vectors of all asynchronous events within the window. Arranged according to sampling time and channel order, the asynchronous event feature sequence is obtained. The formula is as follows: .

9. The metal detection method according to claim 6, characterized in that, In step 1, the process of obtaining the graph node feature matrix is ​​as follows: For the GCN graph convolution branch, a separate graph node feature matrix is ​​constructed; when a metal target approaches the detection area, the DATA0 and DATA1 channels will produce responses with temporal correlation and spatial differences; Therefore, the DATA0 and DATA1 detection channels are respectively used as graph nodes. and Construct a two-node linear array; No. At the sampling point, the first The graph node characteristics of each channel node are represented as follows: ； in For the first Each channel at the sampling position The normalized response value at that point, This indicates that the channel is at the sampling position. Time difference results; Indicates the relative offset of the channel polling; For the first Channel position code for each detection channel; This yields the graph node feature matrix. ; The superscript 2 indicates that there are two nodes in the figure, namely the two detection channels DATA0 and DATA1; the superscript 4 indicates that each node has four features, namely the normalized response value, the temporal difference result, the channel polling relative offset, and the channel position encoding. The adjacency matrix between the two channel nodes DATA0 and DATA1 is as follows: Diagonal elements indicate self-connection of nodes, while off-diagonal elements indicate a spatial association between DATA0 and DATA1. For adjacency matrix Normalization is performed to obtain the normalized adjacency matrix. : ； in, Representing the adjacency matrix The corresponding degree matrix; Will and normalized adjacency matrix Both serve as inputs to the two-node graph convolution spatial feature extraction branch.

Images