Intelligent optimization method for sorting mechanism of waste wood-metal impurities coupled with physical characteristics

Abstract

The present application relates to the technical field of intelligent optimization, and particularly relates to a kind of waste wood metal impurity physical feature coupling sorting mechanism intelligent optimization method.The method comprises the following steps: the electromagnetic induction attenuation spectrum and spectrum reflection characteristics of metal impurities in waste wood stream are collected by high-frequency pulse eddy current sensor array and multispectral imaging unit, and three-dimensional feature tensor is constructed in combination with laser profile scanning data of material conveying belt;Nonlinear response feature matrix is generated based on three-dimensional feature tensor;Nonlinear response feature matrix is input into pre-trained sorting parameter mapping model, which outputs dynamic synergistic control parameter set for the current material state;According to dynamic synergistic control parameter set, control instructions are issued to the actuators of eddy current separator, eddy current sorting machine and air separation unit through industrial bus protocol, and the sorting parameter mapping model is iteratively corrected based on real-time sorting purity feedback data.The present application can continuously improve the rejection effect of multi-class metal impurities.

Description

Technical Field

[0001] This invention relates to the field of intelligent optimization technology, and in particular to an intelligent optimization method for a sorting mechanism that couples the physical characteristics of metal impurities in waste wood. Background Technology

[0002] With the rapid development of China's resource recycling industry and high-end environmental protection equipment industry, the resource recycling and reuse of waste wood has become an important link in solid waste recycling and green manufacturing upgrades, and is widely used in fields such as waste furniture recycling, construction waste wood sorting, and recycled board preparation. During the recycling process of waste wood, the material contains a large number of metal impurities such as iron filings, copper and aluminum fragments, and stainless steel micro-shavings. If these impurities cannot be completely removed, they will directly cause serious wear on subsequent installation tools, production line jams, and even fires caused by metal friction, posing significant safety hazards. At the same time, they significantly reduce the purity and quality of the recycled wood.

[0003] Existing technologies lack a separation mechanism that combines the physical characteristics of multiple metal impurities with multi-physics field coupling, resulting in blind design of separation parameters and poor adaptability to operating conditions. Waste wood contains various types of metal impurities such as iron, copper, aluminum, and stainless steel. The magnetic, electrical, spectral, and density physical properties of different metals vary significantly. Furthermore, uneven wood matrix thickness and humidity fluctuations strongly interfere with metal identification signals, making it impossible to systematically distinguish the differences in physical response of metals of different particle sizes and materials. Relying solely on a single fixed magnetic field strength and fixed eddy current frequency for separation operations cannot match the dynamic changes in wood layer thickness and material humidity. Consequently, the design of magnetic field parameters, eddy current frequency, and wind separation parameters lacks a theoretical basis, leading to significant fluctuations in separation accuracy and efficiency under complex operating conditions, and poor simultaneous removal of multiple types of metal impurities. Summary of the Invention

[0004] Based on this, it is necessary for the present invention to provide an intelligent optimization method for sorting mechanisms that couples the physical characteristics of metal impurities in waste wood, in order to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, a smart optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities includes the following steps: Step S1: Simultaneously collect the electromagnetic induction attenuation spectrum and spectral reflection characteristics of metal impurities in the waste wood flow through a high-frequency pulsed eddy current sensor array and a multispectral imaging unit, and combine them with the laser profile scanning data of the material conveyor belt to construct a three-dimensional feature tensor containing metal particle size, material type and spatial coordinates. Step S2: Based on the three-dimensional feature tensor, the energy dissipation characteristics and phase shift characteristics of metal impurities in time-varying magnetic field and eddy current field are extracted using a multi-physics field coupling analytical engine to generate a nonlinear response feature matrix characterizing the degree of coupling between metal physical properties and wood matrix interference. Step S3: Input the nonlinear response feature matrix into the pre-trained sorting parameter mapping model, which outputs a set of dynamic coordinated control parameters for the magnetic field gradient, eddy current frequency, and airflow sorting cavity wind speed for the current material state. Step S4: Based on the dynamic collaborative control parameter set, control commands are sent to the actuators of the iron remover, eddy current separator and air classifier via the industrial bus protocol, and the sorting parameter mapping model is iteratively corrected based on real-time sorting purity feedback data.

[0006] The beneficial effects of this invention are: The intelligent optimization method for sorting metal impurities in waste wood proposed in this invention, which couples physical characteristics, offers advantages over existing technologies. By relying on a high-frequency pulsed eddy current sensor array, a multispectral imaging unit, and a laser contour scanning device, it simultaneously collects electromagnetic induction attenuation spectrum, spectral reflectance characteristics, and material contour data of the waste wood material flow. This data is then integrated to construct a three-dimensional feature tensor containing metal particle size, material type, and spatial coordinates. This step achieves simultaneous multi-sensor data acquisition, capturing the inherent characteristics of metal impurities and the state of the material matrix from electromagnetic, optical, and spatial dimensions. It fully preserves the differentiated response information of metals of different materials and particle sizes, reduces signal interference caused by fluctuations in wood processing conditions, and provides multi-dimensional, complete, and interference-resistant original data support for subsequent multi-physics feature analysis and sorting parameter control. By leveraging a multi-physics coupling analysis engine to perform deep analysis of the three-dimensional feature tensor, the energy dissipation and phase shift characteristics of metallic impurities in time-varying magnetic and eddy current fields are extracted. This generates a nonlinear response feature matrix that characterizes the physical properties of the metal and the degree of interference coupling with the wood matrix. This step deeply explores the coupling relationship between metallic impurities and the wood matrix, quantifies the differences in electromagnetic response of different metal materials, distinguishes between real metal signals and matrix interference signals, and forms a nonlinear feature matrix that comprehensively reflects the complex working conditions on site. This provides quantifiable feature basis for subsequent dynamic matching of sorting parameters and adaptation to varying material conditions. The nonlinear response feature matrix is ​​input into a pre-trained sorting parameter mapping model. Based on the model, a set of coordinated control parameters for magnetic field gradient, eddy current frequency, and airflow sorting chamber wind speed is dynamically output. This step automatically generates multi-dimensional coordinated control parameters based on real-time working condition characteristics, realizing the linkage adaptation of three core sorting parameters: magnetic field, eddy current, and wind force. This overcomes the limitations of traditional fixed parameter modes, matches corresponding sorting conditions for different metallic impurity characteristics, and significantly improves the synchronous adaptation and sorting capability for multiple types of metallic impurities. By transmitting dynamic control parameters to each sorting actuator through the industrial bus protocol, and simultaneously optimizing the sorting parameter mapping model based on real-time sorting purity feedback data, this step achieves a closed-loop control of dynamic parameter transmission, real-time equipment execution, effect data feedback, and model iterative optimization. This allows the sorting strategy to continuously adapt to the changing waste wood conditions on site, continuously optimize the multi-physics field collaborative sorting logic, continuously improve the removal effect of various types of metal impurities, ensure the purity of recycled wood, and avoid production line wear and fire safety hazards. Attached Figure Description

[0007] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the steps of the intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities in the present invention. Figure 2 for Figure 1A detailed flowchart of step S1; Figure 3 for Figure 1 A detailed flowchart of step S2. Detailed Implementation

[0008] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0009] To achieve the above objectives, please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of the intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities. The present invention provides an intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities. In this example, the intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities includes the following steps: Step S1: Simultaneously collect the electromagnetic induction attenuation spectrum and spectral reflection characteristics of metal impurities in the waste wood flow through a high-frequency pulsed eddy current sensor array and a multispectral imaging unit, and combine them with the laser profile scanning data of the material conveyor belt to construct a three-dimensional feature tensor containing metal particle size, material type and spatial coordinates. In this embodiment, in the upstream area of ​​the waste wood conveyor belt, a high-frequency pulsed eddy current sensor array is installed across the width of the conveyor belt with 32 probes per row. The center-to-center distance between adjacent probes is fixed at 15 mm, ensuring coverage of the entire 1200 mm width of the conveyor belt. Each probe emits an excitation signal with a pulse width of 50 microseconds at a frequency of 120 kHz, and the pulse repetition frequency is set to 2000 Hz. When the waste wood passes under the sensor array at a speed of 1.5 m / s, each probe sequentially collects the eddy current attenuation signal induced by metal impurities, with a sampling rate of 1 million times per second and a recording duration of 300 microseconds. The peak amplitude and attenuation time constant are extracted from the signal of each probe. The peak amplitude is defined as the voltage amplitude at 5 microseconds after the pulse ends, and the attenuation time constant is obtained by exponentially fitting the signal envelope. The calculations show that 64 data points from 32 probes are stitched together to form a complete eddy current decay time series signal. The synchronously triggered multispectral imaging unit consists of two hyperspectral cameras: the first covers a wavelength range of 400 nm to 900 nm with a spectral resolution of 5 nm, and the second covers a wavelength range of 900 nm to 1700 nm with a spectral resolution of 10 nm. The two cameras operate in a synchronous exposure mode triggered by an encoder, with an exposure time of 2 milliseconds. Each scan generates a row of diffuse reflectance spectral data of 1200 pixels. The spectral feature extraction algorithm performs a second derivative transformation on the spectral curve of each pixel, locating the absorption peak of wood cellulose near 1450 nm and the characteristic peak of the metal oxide film near 850 nm on the second derivative curve. The depth and full width at half maximum (FWHM) values ​​of the two absorption peaks are extracted to form a four-dimensional material discrimination feature vector. A laser contour scanner is installed 300 mm downstream of the multispectral imaging unit, acquiring surface contour data of the material layer at a rate of 2000 frames per second, with each frame containing 1200 ranging points and a measurement accuracy of 0.1 mm. The eddy current decay time series signal was extended to 1200 data points through linear interpolation and aligned with the material discrimination feature vector and thickness distribution map. Each voxel contains the metal grain size value. The material type and spatial coordinates are used to generate a three-dimensional feature tensor with dimensions of 1200×L×Z, where L is the total length of the conveyor belt in the forward direction and Z is the maximum depth of the material layer.

[0010] Step S2: Based on the three-dimensional feature tensor, the energy dissipation characteristics and phase shift characteristics of metal impurities in time-varying magnetic field and eddy current field are extracted using a multi-physics field coupling analytical engine to generate a nonlinear response feature matrix characterizing the degree of coupling between metal physical properties and wood matrix interference. In this embodiment, the metal grain size and material type of each voxel are analyzed from the three-dimensional feature tensor, and the skin depth of the metal impurities under a high-frequency alternating magnetic field is calculated. Where ω = 2π × 120000 radians per second, the permeability of ferrite μ = 200 Henry per meter, the permeability of copper and aluminum μ = 1 Henry per meter, the permeability of stainless steel μ = 100 Henry per meter, the electrical conductivity of ferrite σ = 1 × 10^7 Siemens per meter, and the electrical conductivity of copper and aluminum σ = 5.8 × 10^7 Siemens per meter. 7 Siemens stainless steel conductivity per meter σ = 1.4 × 10⁻⁶ 6 Siemens per meter. The skin effect parameter is defined as the ratio of the metal grain size to the skin depth. The Lorentz force response parameter is calculated by multiplying the eddy current decay time constant by the peak amplitude. B = 0.2 Tesla. Decoupling is performed using a wood dielectric constant interference model. Wood dielectric constant... M represents the percentage of moisture content, and the corrected skin depth. Correcting Lorentz force The decoupled parameters were input into a finite element multiphysics coupled model. The solution domain was a space 200 mm above, 50 mm below, and 100 mm on each side of the conveyor belt. The mesh element side length was 2 mm, with a total of 2.5 million meshes. The transient solution time step was 0.001 seconds, with a total duration of 2 seconds, simulating the motion trajectory of metal particles under the combined action of a magnetic field, eddy current field, and airflow field. The energy dissipation characteristics of each particle were extracted from the simulation results. and phase shift characteristics The energy dissipation features and phase shift features are concatenated using tensors. The key feature dimensions are enhanced by an attention mechanism network consisting of three fully connected layers. The first layer has 64 neurons, the second layer has 32 neurons, and the third layer has 2 neurons, which output attention weights. The weighted weights are then concatenated with the metal particle size values ​​to generate a nonlinear response feature matrix with a dimension of N×3.

[0011] Step S3: Input the nonlinear response feature matrix into the pre-trained sorting parameter mapping model, which outputs a set of dynamic coordinated control parameters for the magnetic field gradient, eddy current frequency, and airflow sorting cavity wind speed for the current material state. In this embodiment, each row of the nonlinear response feature matrix corresponds to three eigenvalues ​​of a metal particle. The matrix is ​​then normalized using the following formula: The weighted energy dissipation characteristic has a minimum value of 0.01 joules and a maximum value of 0.3 joules; the weighted phase shift characteristic has a minimum value of 0.1 radians and a maximum value of 2.5 radians; and the metal grain size has a minimum value of 0.5 millimeters and a maximum value of 20 millimeters. The normalized matrix is ​​then processed using a radial basis function. Mapped to a high-dimensional feature space, γ=0.1. Kernel principal component analysis is used to extract the top three principal components with a cumulative contribution rate of 85%, resulting in a dimensionality-reduced key feature component matrix. This matrix is ​​input into a pre-trained deep reinforcement learning sorting parameter mapping model, which consists of a policy network and a value network. The policy network has three neurons in its input layer, two hidden layers with 128 and 64 neurons respectively, using ReLU activation function, and three neurons in its output layer with Tanh activation function. The output value is mapped to the range of -1 to 1 and scaled to actual physical parameters: magnetic field gradient scaling ranges from 0.1 Tesla per meter to 0.5 Tesla per meter, eddy current frequency scaling ranges from 80 kHz to 160 kHz, and airflow sorting cavity wind speed scaling ranges from 10 m / s to 20 m / s. The value network has six neurons in its input layer, two hidden layers with 128 and 64 neurons respectively, and one output neuron outputting the expected reward value. The reward signal is the sorting purity. After the model outputs a three-dimensional cooperative action vector, it is based on the micro-metallic impurity concentration data. Perform security threshold correction when When the number exceeds 50 per square meter, the magnetic field gradient is corrected. eddy current frequency Airflow sorting chamber wind speed , The revised parameters constitute a dynamic coordinated control parameter set, including parameters for iron grading and parameters for non-ferrous metal grading.

[0012] Step S4: Based on the dynamic collaborative control parameter set, control commands are sent to the actuators of the iron remover, eddy current separator and air classifier via the industrial bus protocol, and the sorting parameter mapping model is iteratively corrected based on real-time sorting purity feedback data.

[0013] In this embodiment, the dynamic coordinated control parameter set is analyzed as the excitation current of the permanent magnet separator. , =0.8 amperes per tesla per meter; eddy current coil drive frequency duty cycle Centrifugal fan inverter frequency , =3 Hz per meter per second. The three parameters are combined into a 48-bit execution instruction data packet, with an additional 8-bit checksum. The data packet is encapsulated into a manufacturing message specification protocol frame conforming to the IEC 61850-8-1 standard, embedding a precise time protocol timestamp in a reserved field. The master controller broadcasts the protocol frame to the industrial Ethernet switching network, and the intelligent switch forwards the message to the corresponding slave controller based on the virtual LAN tag. The slave controller achieves microsecond-level synchronization with the master clock through a precise time protocol synchronization process. The synchronization process includes the master controller sending a synchronization message carrying a transmission timestamp T1, the slave recording a reception timestamp T2, the slave sending a delay request message carrying T2 and recording a transmission timestamp T3, the master controller recording a reception timestamp T4, and calculating the downlink delay. Uplink delay D Take the average value As a link delay, the slave controller utilizes... The local hardware timestamp unit is compensated and calibrated. Upon synchronization, each slave controller simultaneously parses and executes the command, driving the power module to adjust its output. At the sorting exit, the online metal detection probe collects residual metal concentration data at a rate of 1000 sample points per second, calculating the average value for every 100 sample points. Conversion and sorting purity feedback value This feedback value is sent back to the sorting parameter mapping model every 0.1 seconds and compared with the target purity threshold of 98%. The deviation... .like If the learning rate is greater than 2%, adjust the model's learning rate. The learning rate is adjusted to 0.005 when the learning rate is between 2% and 5%. When the percentage is between 5% and 10%, adjust it to 0.01. When the bias is greater than 10%, the learning rate is adjusted to 0.02. After the bias is restored, the learning rate gradually returns to the base value of 0.001.

[0014] Furthermore, as an embodiment of the present invention, reference is made to... Figure 2 As shown, Figure 1 A detailed flowchart of step S1 is shown below. In this embodiment, step S1 includes the following steps: Step S11: Drive the high-frequency pulse eddy current sensor array to scan the width direction of the conveyor belt with a preset overlap coverage rate, collect the eddy current decay time series signal caused by metal impurities at different depths, and extract the peak amplitude and decay time constant in the eddy current decay time series signal. In this embodiment, on the waste wood recycling production line, a high-frequency pulsed eddy current sensor array is installed across the width of the conveyor belt in a fixed arrangement. The center-to-center distance between adjacent sensor probes is set to 15 mm to ensure an overlap coverage rate of no less than 85%. The sensor array consists of 32 independent probes, each with a built-in excitation coil and receiving coil. The excitation frequency is locked at 120 kHz, the pulse width is set to 50 microseconds, and the repetition frequency is 2000 Hz. When the waste wood layer passes under the sensor array at a constant speed of 1.5 meters per second, each probe sequentially emits a high-frequency pulsed magnetic field into the wood layer, generating transient eddy currents inside the metal impurities. These eddy currents decay exponentially as the pulse is interrupted. The receiving coil continuously collects the voltage signal during the decay process, with a sampling rate set to 1 million times per second and a recording duration limited to 300 microseconds. Two key features are extracted from the collected signal: the peak amplitude is defined as the voltage amplitude at 5 microseconds after the pulse ends, in millivolts; the decay time constant is calculated by exponentially fitting the signal envelope, using the following formula: ,in Let be the voltage at time t. The initial amplitude, The decay time constant is expressed in microseconds. Threshold ranges are set for different metal types: ferrous impurities have a peak amplitude between 80 mV and 200 mV, with a decay time constant between 20 and 40 microseconds; copper and aluminum impurities have a peak amplitude between 150 mV and 350 mV, with a decay time constant between 45 and 70 microseconds; and stainless steel impurities have a peak amplitude between 60 mV and 120 mV, with a decay time constant between 15 and 30 microseconds. Each probe independently outputs a pair of peak amplitude and decay time constant values. These 32 pairs are stitched together in probe order to form a complete eddy current decay time sequence signal. This sequence has a fixed length of 64 data points, covering the metal distribution information across the entire conveyor belt width.

[0015] Step S12: Synchronously trigger the multispectral imaging unit to collect diffuse reflectance spectral data of the material surface, and use the spectral feature extraction algorithm to separate the absorption peak of wood cellulose and the characteristic peak of metal oxide film to generate material discrimination feature vector; In this embodiment, the multispectral imaging unit consists of two hyperspectral cameras, each mounted 400 mm above the material surface directly above the conveyor belt, spaced 600 mm apart along the conveyor belt's forward direction. The first camera covers a wavelength range of 400 nm to 900 nm with a spectral resolution of 5 nm; the second camera covers a wavelength range of 900 nm to 1700 nm with a spectral resolution of 10 nm. Each camera is equipped with a linear scanning lens, with a scanning width corresponding to a conveyor belt width of 1200 mm. Each scan generates one row of pixels, with 1200 pixels per row, meaning each pixel corresponds to a 1 mm wide material area. A synchronization trigger signal is generated by an encoder; each encoder rotation corresponds to a 1 mm movement of the conveyor belt, triggering a simultaneous dual-camera exposure. The exposure time is uniformly set to 2 milliseconds. The acquired raw spectral data undergoes dark current correction and reflectance normalization to obtain the diffuse reflectance spectral curve for each pixel. The spectral feature extraction algorithm first performs a second-order derivative transformation on each spectral curve, calculated using the following formula: Where R(λ) is the reflectance at wavelength λ. Two characteristic absorption peaks are located on the second derivative curve: the cellulose absorption peak appears near wavelength 1450 nm, and the precise wavelength corresponding to its second derivative minimum point is determined by cubic spline interpolation with an allowable deviation range of ±5 nm; the characteristic peak of the metal oxide film appears near wavelength 850 nm, and the precise wavelength corresponding to its second derivative minimum point is also determined by cubic spline interpolation with an allowable deviation range of ±3 nm. The depth values ​​of these two absorption peaks are extracted, defined as the difference between the baseline reflectance and the peak-valley reflectance, expressed as a percentage. Furthermore, the full width at half maximum (FWHM) values ​​of the two absorption peaks are extracted, expressed in nanometers. These four characteristic values ​​are combined into a four-dimensional material discrimination feature vector, corresponding to the cellulose absorption peak depth, the cellulose absorption peak FWHM, the metal oxide film absorption peak depth, and the metal oxide film absorption peak FWHM, respectively. For each pixel, if the characteristic peak depth of the metal oxide film is greater than 8% and the full width at half maximum (FWHM) is less than 20 nm, the pixel is marked as a metal impurity region; if the absorption peak depth of cellulose is greater than 15% and the FWHM is greater than 40 nm, the pixel is marked as a pure wood region; otherwise, it is marked as a mixed region. The marking results of the 1200 pixels along the entire scan line form a one-dimensional material tag array of length 1200, which corresponds one-to-one with the spatial location of the eddy current decay time series signal.

[0016] Step S13: Fuse the eddy current decay time series signal with the material discrimination feature vector, and combine it with the material layer thickness distribution map obtained by the laser profile scanner to generate a three-dimensional feature tensor containing metal particle size, material type and spatial coordinates through a three-dimensional spatial reconstruction algorithm.

[0017] In this embodiment, a laser profile scanner is installed 300 mm downstream of the multispectral imaging unit. The scanner's laser line is projected onto the conveyor belt surface, with the laser line direction perpendicular to the conveyor belt's forward direction, covering the entire width of the conveyor belt. The scanner acquires material layer surface profile data at a rate of 2000 frames per second. Each frame contains 1200 measurement points, each corresponding to a horizontal position coordinate and a height value at that position. The height measurement accuracy is 0.1 mm, and the measurement range is set from 0 mm to 200 mm. The continuously acquired profile data is arranged in chronological order. Combined with the forward displacement calculated from the conveyor belt speed, a material layer thickness distribution map is constructed. This map is a two-dimensional matrix. The number of rows in the matrix corresponds to the position step size in the conveyor belt's forward direction, with a step size of 1 mm. The number of columns in the matrix corresponds to the 1200 measurement points in the conveyor belt width direction. The matrix element value is the material layer thickness at that position, in millimeters. The three-dimensional spatial reconstruction algorithm uses the eddy current attenuation time series signal, material discrimination feature vector, and material layer thickness distribution map as inputs. First, the 64 data points in the eddy current decay time series signal are expanded to 1200 data points through linear interpolation to align with the column numbers of the material label array and thickness distribution map. Then, a fusion calculation is performed on each column of data: for column index i, the eddy current characteristic value... Take the peak amplitude and decay time constant corresponding to this column, and use the material label. Extract the feature peak depth and thickness values ​​of the metal oxide film from the material discrimination feature vector. The average thickness of the column in the thickness distribution map is taken. Each voxel of the three-dimensional feature tensor is defined as a feature vector at spatial coordinates (x, y, z), where x is the coordinate in the width direction of the conveyor belt, y is the coordinate in the forward direction of the conveyor belt, and z is the coordinate in the depth direction of the material layer. The voxel size is set to 1 mm in the x direction, 1 mm in the y direction, and 1 mm in the z direction. The metal grain size value of each voxel is calculated by the ratio of the eddy current decay time constant to the peak amplitude, using the following formula: Where D is the estimated metal particle size in millimeters, and K is the calibration coefficient, with a value of 0.05 mm·mV / µs. The material category is determined by the material label array: the material category of the volumetric material in the metal impurity region is assigned a value of 1, the material category of the volumetric material in the pure wood region is assigned a value of 0, and the material category of the volumetric material in the mixed region is assigned a value of 2. The spatial coordinates are obtained by multiplying the index position of the voxel in the 3D matrix by the voxel size in the corresponding direction. The final generated 3D feature tensor has dimensions of 1200×L×Z, where L is the total length of the conveyor belt in the forward direction, Z is the depth layer number after dividing the maximum thickness of the material layer by the voxel size, and each voxel contains three feature values: metal particle size, material category, and spatial coordinates.

[0018] Furthermore, as an embodiment of the present invention, reference is made to... Figure 3 As shown, Figure 1 A detailed flowchart of step S2 is shown below. In this embodiment, step S2 includes the following steps: Step S21: Extract the skin effect parameters of metallic impurities under high-frequency alternating magnetic field and the Lorentz force response parameters in eddy current field from the three-dimensional feature tensor, and perform decoupling processing based on the wood dielectric constant interference model. In this embodiment, after extracting the metal grain size and material type values ​​of each voxel from the three-dimensional feature tensor, the skin effect parameters of metal impurities under a high-frequency alternating magnetic field are calculated. Skin depth The calculation formula is: Where ω is the angular frequency of the eddy current separator, set to 2π × 120000 radians per second; μ is the permeability of the metallic impurities, with 200 Henry per meter for ferrous impurities, 1 Henry per meter for copper and aluminum impurities, and 100 Henry per meter for stainless steel impurities; σ is the electrical conductivity of the metallic impurities, with 1 × 10⁻⁶ electrical conductivity for ferrous impurities. 7 For Siemens, the conductivity of copper and aluminum impurities per meter is taken as 5.8 × 10⁷. For Siemens, the conductivity of stainless steel impurities per meter is taken as 1.4 × 10⁷. 6 Siemens per meter. The skin effect parameter for each voxel is defined as the ratio of the metal grain size to the skin depth. A ratio greater than 1 indicates that the metal impurity is in the skin effect-dominant region, while a ratio less than 1 indicates that the metal impurity is in the transmission region. The Lorentz force response parameter is calculated by multiplying the eddy current decay time constant by the peak amplitude, using the following formula: Where B is the magnetic flux density of the eddy current separator's magnetic field, set to 0.2 Tesla. The Lorentz force response parameter for each voxel is in Newtons. The wood dielectric constant interference model is used to remove the interference of the wood matrix on the eddy current signal; for example, the wood dielectric constant... The relationship between moisture content and water content is as follows: The value of M represents the percentage moisture content of the wood, ranging from 5% to 30%, with a coefficient of 2.5 representing the baseline value of the dielectric constant of the waste wood matrix in a dry state. After crushing and screening, the free water content in the lignocellulose structure of waste wood is extremely low. At this point, the relative dielectric constant of the wood matrix mainly depends on the molecular polarization characteristics of cellulose, hemicellulose, and lignin. For example, through dielectric spectrum analysis of a large number of test samples, under test conditions at a frequency of 120 kHz, the relative dielectric constant of dry waste wood is concentrated between 2.3 and 2.7, with the median value of 2.5 taken as the baseline value. The coefficient 0.05 represents the increment coefficient of the wood dielectric constant for each percentage point increase in moisture content. This coefficient comes from the dielectric constant calibration experiment of wood samples with different moisture contents. In the experiment, waste wood samples were placed in a constant temperature and humidity chamber, and the moisture content was gradually increased from 5% to 30%. The relative dielectric constant of the sample at a frequency of 120 kHz was measured at each moisture content node. Measurement results show that for every 1% increase in moisture content, the dielectric constant increases by an average of 0.04 to 0.06; the median value of 0.05 is taken as the slope of the linear fit. The decoupling process involves correcting the skin effect parameters for each voxel using the following formula: ,in The vacuum permittivity is 8.854 × 10⁻⁶. -12 Farads per meter; the Lorentz force response parameters for each voxel are corrected, for example, the correction formula is... The coefficient 0.01 represents the correction factor for the effect of moisture content on the Lorentz force response parameter. This coefficient originates from an eddy current signal attenuation experiment: under the same metal impurity embedding conditions, the amplitude of eddy current attenuation signals collected from wood with different moisture contents was compared. The experiment found that for every 1% increase in moisture content, the amplitude of the eddy current signal attenuated by an average of about 1%. Therefore, the correction factor was set to 0.01 to proportionally reduce the Lorentz force response parameter under the influence of moisture content. The corrected skin effect parameter and Lorentz force response parameter are stored as two two-dimensional matrices, with the matrix dimensions consistent with the first two dimensions of the three-dimensional feature tensor, i.e., 1200×L, where L is the total length of the conveyor belt in the forward direction, in millimeters.

[0019] Step S22: Construct a multiphysics coupling model based on finite element analysis, and use the decoupled skin effect parameters and Lorentz force response parameters as boundary condition inputs to simulate the trajectory evolution of metal particles under the combined action of magnetic field, eddy current field and airflow field to generate simulation results; In this embodiment, the multiphysics coupling model is constructed based on the finite element mesh generation method. The solution domain is set as the working area of ​​the eddy current separator, including the space within a height range of 200 mm above the conveyor belt surface, the space within a depth range of 50 mm below the conveyor belt surface, and the extension areas of 100 mm on each side in the width direction of the conveyor belt. The solution domain is divided into hexahedral mesh elements, with a mesh element side length of 2 mm, and a total number of mesh elements of approximately 3 million. The magnetic field boundary condition is set as the magnetic flux density distribution on the surface of the magnetic roller of the eddy current separator. The magnetic roller has a diameter of 300 mm and a length of 1300 mm, and the surface magnetic flux density is sinusoidally distributed along the axial direction, with a peak magnetic flux density of 0.25 Tesla. The eddy current field boundary condition is set as the decoupled Lorentz force response parameters, with the Lorentz force corresponding to each voxel applied as a volume force to the mesh node where the voxel is located. The airflow boundary condition is set as the wind speed distribution at the outlet of the eddy current separator. The outlet is located 200 mm behind the magnetic roller, with a width of 1200 mm and a height of 100 mm. The wind speed is uniformly distributed along the width of the outlet and is set to 15 m / s. The material motion boundary condition is set as the conveyor belt speed, which is 1.5 m / s and directed along the conveyor belt's forward direction. The initial position of each metal particle is determined by its spatial coordinates in the three-dimensional feature tensor, and the initial velocity is set to the conveyor belt speed. The finite element solution process uses a transient solver with a time step of 0.001 seconds and a total solution time of 2 seconds. Within each time step, the magnetic field equation is solved first. Where A is the magnetic vector potential and J is the current density; then solve the eddy current field equations. Where E is the electric field strength and B is the magnetic flux density; then solve the Navier-Stokes equations for the airflow field. Where ρ is the air density, taken as 1.225 kg / m³, u is the air velocity vector, p is the pressure, and μ is the aerodynamic viscosity, taken as 1.8 × 10⁻⁶. -5 Pascal second, F is the Lorentz force; finally, solve the equations of motion for the particles. Where m is the particle mass and v is the particle velocity vector. For gravity, For airflow drag, For Lorentz force, The magnetic force is used. The position and velocity of all particles are updated at the end of each time step until the solution is complete. The simulation results include the complete trajectory of each metal particle over 2 seconds, with 2000 trajectory points, each containing three-dimensional spatial coordinates and a velocity vector.

[0020] Step S23: Analyze the differences in the separation trajectories of metal particles of different materials in the simulation results, and extract the energy dissipation characteristics that characterize the dispersion of their motion trajectories and the phase shift characteristics that characterize the phase synchronization. In this embodiment, the trajectory data of each metal particle is extracted from the simulation results. The trajectory data consists of a series of three-dimensional coordinate points, with an interval of 0.001 seconds between the coordinate points. For each material category, the trajectory data of all similar metal particles are aggregated, and the energy dissipation characteristics are calculated. The energy dissipation characteristic is defined as the rate of change of particle kinetic energy over time, and the kinetic energy calculation formula is as follows: Where m is the particle mass and v is the particle velocity vector. A first-order difference operation is performed on the kinetic energy time series of each particle to obtain the kinetic energy change rate series. ,in =0.001 seconds. The energy dissipation characteristic value is taken as the standard deviation of the kinetic energy change rate sequence, and the standard deviation is calculated using the following formula: Where N is the total number of time steps, The value represents the mean of the kinetic energy change rate sequence. The energy dissipation characteristic value of ferrous impurities ranges from 0.01 joules to 0.05 joules, that of copper and aluminum impurities ranges from 0.1 joules to 0.3 joules, and that of stainless steel impurities ranges from 0.02 joules to 0.08 joules. The phase shift characteristic is defined as the phase synchronization between the particle trajectory and the rotation period of the magnetic roller in the eddy current separator. The rotation period of the magnetic roller is T = 1 / f, where f is the rotation frequency of the magnetic roller, set to 50 Hz, and the period T = 0.02 seconds. A Fourier transform is performed on the trajectory data of each particle to extract the phase angle component at a frequency of 50 Hz. Simultaneously, the phase angle of the magnetic flux density on the surface of the magnetic roller of the eddy current separator at the same frequency is extracted. Phase shift eigenvalues The units are radians. The phase shift eigenvalues ​​of ferrous impurities range from 0.1 to 0.5 radians, those of copper and aluminum impurities range from 1.5 to 2.5 radians, and those of stainless steel impurities range from 0.8 to 1.2 radians. The energy dissipation eigenvalue and phase shift eigenvalue of each metal particle are combined into a two-dimensional feature vector, and the feature vectors of all particles are stored in groups according to material type.

[0021] Step S24: Tensor-join the energy dissipation features and phase shift features, and use an attention mechanism network to enhance the key feature dimensions to generate a nonlinear response feature matrix that characterizes the coupling degree between the physical properties of the metal and the interference of the wood matrix.

[0022] In this embodiment, the energy dissipation feature value and phase shift feature value of each metal particle are concatenated using a tensor. First, an empty three-dimensional tensor is constructed with dimensions N×2×K, where N is the total number of metal particles, 2 is the number of feature dimensions, and K is the number of material categories (3 in total, corresponding to iron, copper-aluminum, and stainless steel). The feature vector of each particle is then filled into the corresponding slice of its material category. For example, the feature vector of iron particles is filled into the first slice, the feature vector of copper-aluminum particles into the second slice, and the feature vector of stainless steel particles into the third slice. After filling, each element of the three-dimensional tensor is a two-dimensional feature vector. The attention mechanism network consists of three fully connected layers. The first fully connected layer has 64 neurons and uses the ReLU activation function; the second fully connected layer has 32 neurons and uses the ReLU activation function; and the third fully connected layer has 2 neurons and uses the Softmax activation function. The attention weights are calculated by flattening the three-dimensional tensor into a one-dimensional vector with a length of N×2×K. The vector is input into the first fully connected layer, outputting an intermediate feature vector of length 64. This intermediate feature vector is then input into the second fully connected layer, outputting another intermediate feature vector of length 32. Finally, the 32-dimensional vector is input into the third fully connected layer, outputting an attention weight vector of length 2. The two weight values ​​correspond to the weighting coefficients of the energy dissipation feature and the phase shift feature, respectively. The weighting coefficients range from 0 to 1, and the sum of the two weight values ​​is 1. Each two-dimensional feature vector in the original 3D tensor is element-wise multiplied with the attention weight vector to obtain a weighted feature vector. The weighted feature vector is then reorganized into a 3D tensor with the dimensions unchanged. This 3D tensor is summed along the material category dimension to obtain an N×2 2D matrix. The rows of the matrix correspond to the metal particle numbers, and the columns correspond to the weighted energy dissipation feature value and the phase shift feature value. This 2D matrix is ​​then concatenated with the metal particle size values ​​in the original 3D feature tensor by adding a column of particle size values ​​along the column direction, resulting in an N×3 2D matrix. This matrix is ​​the nonlinear response characteristic matrix. Each row of the matrix contains three eigenvalues: weighted energy dissipation characteristic, weighted phase shift characteristic, and metal particle size value. The number of rows N in the nonlinear response characteristic matrix is ​​equal to the total number of metal particles, and the number of columns is fixed at 3.

[0023] Furthermore, step S22 includes the following steps: Step S221: Establish the coupled control equations for the electromagnetic field module described by Maxwell's equations, the fluid field module described by Navier-Stokes equations, and the particle motion module described by the discrete phase model. Based on the coupled control equations, construct a multiphysics coupled model based on finite element analysis. In this embodiment, the construction of the multiphysics coupling model is based on the coupling control equations. The electromagnetic field module is described by Maxwell's equations, specifically expressed by Ampere's circuital law. Where H is the magnetic field strength vector, J is the conduction current density vector, and D is the electric displacement vector; Faraday's law of electromagnetic induction Where E is the electric field intensity vector and B is the magnetic flux density vector; Gauss's law of magnetism Gauss's electric law ,in Let be the free charge volume density. The fluid field module is described by the Navier-Stokes equations, expressed as: Where ρ is the fluid density, u is the velocity vector, p is the pressure, μ is the dynamic viscosity, g is the gravitational acceleration vector, and F is the volume force vector. The particle motion module is described by a discrete phase model, and the force equilibrium equation for the particles is: ,in For particle mass, For particle velocity vectors, For drag force, For gravity, For Lorentz force, The force is magnetic. The three modules are connected via coupling terms: the Lorentz force generated by the electromagnetic field module. As the volume force input to the fluid field module and particle motion module; the fluid field module calculates the airflow velocity. As drag force The input, where Let A be the drag coefficient and A be the particle's frontal area. The solution domain is set to the working area of ​​the eddy current separator, with a spatial range of 2000 mm in length, 1400 mm in width, and 300 mm in height along the conveyor belt's forward direction. Tetrahedral elements are used for mesh generation, with a maximum edge length of 5 mm and a minimum edge length of 1 mm. The mesh is refined near the magnetic roller surface and air outlet, reducing the edge length of elements to 2 mm in the refined areas. The total number of mesh elements is approximately 2.5 million. The solver is set to a transient solver with a fixed time step of 0.0005 seconds and a total solution time of 1.5 seconds. The convergence criterion is a residual less than 1 × 10⁻⁶. -6 .

[0024] Step S222: Import the actual physical property parameters of the waste wood, including density, moisture content and dielectric constant, and use them as variable medium property inputs in the fluid field module; In this embodiment, the actual physical property parameters of the waste wood are obtained through offline measurement. Density measurement uses the drying and weighing method, where the waste wood sample is dried to constant weight in a 105°C oven, the dried mass is weighed, and the volume is measured using the water displacement method to calculate the density value. The density range of the waste wood is 0.4 g / cm³ to 0.8 g / cm³, and is set to 0.65 g / cm³ based on actual batch measurement results. Moisture content measurement uses a resistance moisture meter, where the moisture meter probe is inserted into the wood layer to read the moisture content value. The moisture content range of the waste wood is 8% to 25%, and is set to 15% based on actual on-site measurements. Dielectric constant measurement uses the parallel plate capacitance method, where the wood sample is placed between parallel plate electrodes, the capacitance value is measured, and the result is calculated using the formula... Calculate the relative permittivity, where C is the capacitance value with wood present. This represents the capacitance value without wood. The measured relative permittivity of waste wood at 120 kHz is 3.2. These three physical property parameters are used as variable medium property inputs in the fluid field module. Specifically, in the material property definition of the fluid field module, the air density is replaced with the equivalent density of the wood-air mixture. The formula for calculating the equivalent density is... ,in For wood density, The air density is taken as 1.225 kg per cubic meter. This is the wood volume fraction, calculated based on the material layer thickness and conveyor belt width. The value ranges from 0.3 to 0.7. The formula for calculating the equivalent dynamic viscosity is as follows: ,in The aerodynamic viscosity is taken as 1.8 × 10⁻⁶. -5 Pascal second. The formula for calculating the equivalent dielectric constant is: ,in Let be the relative permittivity of wood. The relative permittivity of air is set to 1. These equivalent parameters are independently assigned in each grid cell of the fluid field module to achieve spatial distribution characterization of the wood medium properties.

[0025] Step S223: Based on the metal grain size and spatial coordinate data in the three-dimensional feature tensor, initialize multiple discrete metal particle phases with different electrical conductivity and magnetic permeability in the model computation domain; In this embodiment, discrete metal particle phases are initialized within the model computation domain based on the metal particle size and spatial coordinate data in the three-dimensional feature tensor. Each voxel in the three-dimensional feature tensor contains a metal particle size value D and spatial coordinates (x, y, z), where x is the coordinate in the width direction of the conveyor belt, y is the coordinate in the forward direction of the conveyor belt, and z is the coordinate in the depth direction of the material layer. All voxels in the three-dimensional feature tensor are traversed, and voxels with material category 1 or 2 are selected, representing the metal portion in the metal impurity region and the mixed region. For each selected voxel, a discrete metal particle object is created. The diameter of the particle object is set to the metal particle size value D of that voxel, with D ranging from 0.5 mm to 20 mm. The density of the particle object is set according to the material category: 7.87 g / cm³ for iron particles, 2.7 g / cm³ for copper-aluminum particles, and 7.93 g / cm³ for stainless steel particles. The electrical conductivity and magnetic permeability of the particle object are set according to the material category: 1 × 10⁻⁶ for iron particles. 7 Siemens sets the permeability to 200 Henry per meter; the conductivity of copper-aluminum particles is set to 5.8 × 10⁻⁶. 7 Siemens sets the permeability to 1 Henry per meter; the conductivity of stainless steel particles is set to 1.4 × 10⁻⁶. 6 The permeability is set to 100 Henry per meter. The initial position of each particle is set to its spatial coordinates, and its initial velocity is set to the conveyor belt velocity vector, with a magnitude of 1.5 m / s and a direction along the positive y-axis. The initial angular velocity of each particle is set to zero. All created discrete metal particle objects form a particle swarm, with the total number of particles equal to the number of metal voxels in the 3D feature tensor, typically ranging from 500 to 2000. Each particle object is assigned a unique identifier, incrementing from 1. The particle swarm serves as input to the discrete phase model and participates in solving the particle motion equations in subsequent solutions.

[0026] Step S224: The decoupled skin effect parameters and Lorentz force response parameters are used as boundary condition inputs, and a multiphysics coupling model is solved based on discrete metal particle phases to simulate the evolution of the motion trajectory of metal particles under the combined action of magnetic field, eddy current field and airflow field, so as to generate simulation results.

[0027] In this embodiment, the decoupled skin effect parameters and Lorentz force response parameters are used as boundary conditions input to the multiphysics coupling model. The skin effect parameters are used to correct the skin depth. Input in the form of each metal particle, corresponding to... The value is calculated based on the material type of the particle and the moisture content of the wood, using the following formula: Where ω = 2π × 120000 radians per second, and μ and σ are the magnetic permeability and electrical conductivity of the particle. Let be the relative permittivity of wood. is the vacuum permittivity. The value is used to correct the eddy current distribution inside the particle. In the electromagnetic field module, the current density J inside the particle is corrected to... , where r is the distance from a point inside the particle to the particle surface. Lorentz force response parameters are used to correct for the Lorentz force. Input in the form of each metal particle, corresponding to... The value is determined by the eddy current decay time constant of the particle. Peak amplitude The magnetic flux density B is calculated using the following formula: , where M is the percentage of moisture content in the wood. The values ​​are applied as volume forces to the mesh nodes at the particle's location. After setting the boundary conditions, the transient solver is started. Within each time step, the solver sequentially solves the electromagnetic field equations, fluid field equations, and particle motion equations, exchanging data through coupling terms. The solution process lasts 1.5 seconds and consists of 3000 time steps. After each time step, the 3D spatial coordinates and velocity vector of each metal particle are recorded. After the solution is complete, the coordinate data of all particles across all time steps are compiled into a motion trajectory dataset, with each particle's trajectory containing 3000 coordinate points. The simulation results are output as particle trajectory files, with each line containing eight fields: particle identifier, timestamp, x-coordinate, y-coordinate, z-coordinate, x-direction velocity, y-direction velocity, and z-direction velocity. These simulation results are used for trajectory analysis and feature extraction in subsequent steps.

[0028] Furthermore, the initialization of multiple discrete metal particle phases with different electrical and magnetic conductivity within the model computation domain in step S223 includes the following steps: The three-dimensional feature tensor is analyzed to identify the distribution areas of metal impurities of different particle sizes and materials in large iron, copper and aluminum fragments and iron powder, so as to generate identification results. In this embodiment, the three-dimensional feature tensor is a data cube with dimensions of 1200×L×Z, where L is the total length of the conveyor belt in the forward direction, and Z is the maximum depth of the material layer. The parsing operation begins by traversing each voxel of the cube, with the voxel index (i,j,k) corresponding to the i-th column in the width direction, the j-th row in the forward direction, and the k-th layer in the depth direction. Each voxel stores three feature values: metal particle size D, material category C, and spatial coordinates (x,y,z). The material category C takes values ​​of 0, 1, and 2, representing pure wood, metal impurity, and mixed regions, respectively. The identification process first filters out voxels with a material category of 0, retaining only voxels with C equal to 1 or 2. For the retained voxels, a secondary classification is performed based on the value of the metal particle size D: when D is greater than 10 mm, the voxel is marked as a large iron distribution area; when D is between 1 mm and 10 mm, the voxel is marked as a copper-aluminum fragment distribution area; and when D is less than 1 mm, the voxel is marked as an iron powder distribution area. After classification, connected component merging is performed on spatially adjacent voxels of the same category. The rule for determining connected components is that the absolute value of the index difference between two voxels in the x, y, or z directions does not exceed 1. Each connected component generates a recognition result entry, which includes the center coordinates of the connected component, the bounding box size, the number of voxels contained, and the average metal grain size. For example, a large iron block connected component consisting of 50 voxels has center coordinates of (x=345 mm, y=678 mm, z=12 mm), bounding box sizes of 8 mm in the x direction, 10 mm in the y direction, and 4 mm in the z direction, and an average metal grain size of 14.5 mm. All recognition result entries are sorted from largest to smallest connected component volume to form a recognition result list. The list length is equal to the total number of connected components, typically ranging from 200 to 800.

[0029] Based on the identification results, the standard electrical conductivity and magnetic permeability parameters of the corresponding material are retrieved from the preset metal physical property library, and the random distribution of particle size of the particle group is generated based on the particle size distribution data. In this embodiment, the pre-set metal physical property library is stored in tabular form, containing standard electrical conductivity and standard magnetic permeability values ​​for three metal materials. The standard electrical conductivity of iron is set to 1×10⁻⁶. 7 Siemens sets the standard permeability at 200 Henry per meter. The standard conductivity for copper and aluminum is set at 5.8 × 10⁻⁶. 7 Siemens sets the standard permeability at 1 Henry per meter. The standard conductivity for stainless steel is set at 1.4 × 10⁻⁶. 6Siemens sets the standard permeability to 100 Henry per meter. Based on the identification result list, each connected component entry is traversed, and its material category label is read. If the label is "large piece of iron," the standard electrical conductivity and standard magnetic permeability of iron are retrieved from the attribute library; if the label is "copper / aluminum fragments," the standard parameters for copper and aluminum are retrieved; if the label is "iron powder," the standard parameters for iron are retrieved. Simultaneously, the average metal grain size value of the connected component is read and recorded as... .based on and the number of voxels within the connected domain This generates a random distribution of particle size. The random distribution adopts a normal distribution model, and the mean of the normal distribution is... Set as Standard deviation Set as Multiply by 0.2. Generate N random particle size values ​​using the inverse transform sampling method. Specifically, generate N random numbers that follow a uniform distribution. The value ranges from 0 to 1, and then the inverse function of the cumulative distribution function of the normal distribution is used. Calculate the corresponding particle size value The generated random particle size values ​​need to be truncated, with the lower limit set to [value missing]. Multiply by 0.5, and set the upper limit to [value]. Multiply by 1.5, and resample random values ​​that exceed the cutoff range until they fall within the range. A set of random particle size distribution values ​​is generated for each connected component, and the random particle size distribution values ​​of all connected components are summarized into a global particle size list. The length of the list is the sum of the number of voxels in all connected components, typically 500 to 2000.

[0030] By utilizing the random distribution of particle size and standard physical property parameters, discrete phase particles representing different metallic impurities are randomly generated within the computational domain. Each discrete phase particle carries its unique electrical conductivity, magnetic permeability, and initial spatial coordinate properties. In this embodiment, discrete phase particles representing different metallic impurities are randomly generated within the computational domain. The spatial range of the computational domain is set to 0 mm to 1200 mm in the x-direction, 0 mm to L mm in the y-direction, and 0 mm to Z mm in the z-direction, where L is the total length of the conveyor belt in the forward direction and Z is the maximum depth of the material layer. Each discrete phase particle corresponds to a particle size value in the global particle size list. The total number of particles is equal to the length of the global particle size list. For each particle, its initial spatial coordinates are first generated. The generation rules for the initial coordinates are as follows: the x-coordinate is uniformly and randomly selected within the range of 0 mm to 1200 mm; the y-coordinate is uniformly and randomly selected within the range of 0 mm to L mm; and the z-coordinate is uniformly and randomly selected within the range of 0 mm to Z mm. After the initial coordinates of each particle are generated, it is assigned electrical conductivity and magnetic permeability attributes. The electrical conductivity attribute is based on the material category of the connected domain to which the particle belongs, and a standard value is retrieved from the preset attribute library, but a random perturbation of ±5% is added. The perturbation method is to multiply the standard electrical conductivity by a certain factor. ,in The values ​​are uniformly distributed random numbers ranging from -1 to 1. The magnetic permeability attribute is also subject to a ±5% random perturbation, using the same perturbation method. The particle size value for each particle is directly taken from the corresponding value in the global particle size list. Each particle is also assigned a unique identifier, incrementing from 1. All particle attribute information is organized into a particle attribute table. Each row in the table contains seven fields: identifier, x-coordinate, y-coordinate, z-coordinate, particle size, electrical conductivity, and magnetic permeability. The number of rows in the particle attribute table equals the total number of particles. For example, an iron particle might have an identifier of 157, an x-coordinate of 345.2 mm, a y-coordinate of 678.5 mm, a z-coordinate of 12.3 mm, a particle size of 14.5 mm, and an electrical conductivity of 1.03 × 10⁻⁶. 7 Siemens per meter, permeability 195 Henry per meter.

[0031] The set of discrete phase particles is defined as the initial particle phase boundary condition of the multiphysics coupling model.

[0032] In this embodiment, the set of discrete phase particles is defined as the initial particle phase boundary condition of the multiphysics coupling model. This boundary condition definition includes two parts: the initial state of the particle phase and the physical property parameters of the particle phase. The initial state part includes the initial position coordinates, initial velocity, and initial angular velocity of each particle. The initial position coordinates are directly taken from the x, y, and z coordinate values ​​in the particle property table. The initial velocity is set as the conveyor belt velocity vector, with a magnitude of 1.5 meters per second and a direction along the positive y-axis; that is, the velocity component of each particle is set as... =0 meters per second =1.5 meters per second =0 meters per second. The initial angular velocity is set to zero vector, i.e. =0 radians per second =0 radians per second =0 radians per second. The physical properties of the granular phase include the diameter, density, electrical conductivity, and magnetic permeability of each particle. The diameter is taken from the particle size value in the particle property table. The density is set according to the material type: 7870 kg / m³ for iron particles, 2700 kg / m³ for copper and aluminum particles, and 7930 kg / m³ for stainless steel particles. The electrical conductivity and magnetic permeability are directly taken from the corresponding values ​​in the particle property table. The initial state and physical properties of all particles are combined into an initial granular phase dataset. The dataset is formatted as follows: each row contains 20 values, in the following order: identifier, x-coordinate, y-coordinate, z-coordinate, ... , , , , , The initial particle phase dataset is used as the boundary condition input to the multiphysics coupled model. It includes parameters such as diameter, density, electrical conductivity, magnetic permeability, x-direction particle size distribution coefficient, y-direction particle size distribution coefficient, z-direction particle size distribution coefficient, shape factor, surface roughness coefficient, and agglomeration coefficient. The particle size distribution coefficient is set to 1, the shape factor to 1, the surface roughness coefficient to 0.5, and the agglomeration coefficient to 0. At the beginning of the first time step of the model solution, all particles are placed in their initial positions within the computational domain and participate in the iterative calculation of the particle motion equations in subsequent solutions.

[0033] Furthermore, step S3 includes the following steps: Step S31: Normalize the nonlinear response feature matrix and map it to a high-dimensional feature space. Use kernel principal component analysis to extract the key feature components that dominate the sorting efficiency. In this embodiment, the nonlinear response feature matrix is ​​a two-dimensional matrix with dimensions N×3, where N is the total number of metal particles, and the three columns correspond to the weighted energy dissipation feature, the weighted phase shift feature, and the metal particle size value, respectively. Normalization is performed independently for each column, and the normalization formula is as follows: Where x is the original value in that column, The minimum value in this column. This represents the maximum value in this column. The minimum value of the weighted energy dissipation feature column is set to 0.01 joules, and the maximum value is set to 0.3 joules; the minimum value of the weighted phase shift feature column is set to 0.1 radians, and the maximum value is set to 2.5 radians; the minimum value of the metal grain size value column is set to 0.5 millimeters, and the maximum value is set to 20 millimeters. After normalization, all values ​​fall within the range of 0 to 1. The normalized matrix is ​​mapped to a high-dimensional feature space through a radial basis function kernel function, the expression of which is: ,in and Let be the feature vectors of two samples, and γ be the kernel parameter, set to 0.1. Kernel principal component analysis (KPCA) calculates the eigenvalues ​​and eigenvectors of the covariance matrix in a high-dimensional feature space. The covariance matrix has dimensions N×N, and each element consists of a kernel function value. The eigenvalues ​​are sorted from largest to smallest, and the eigenvectors corresponding to the top few eigenvalues ​​with a cumulative contribution rate reaching 85% are selected as the principal component directions. The formula for calculating the cumulative contribution rate is... ,in It is the sum of the first k eigenvalues. The sum of all eigenvalues. Typically, the cumulative contribution of the first three eigenvalues ​​reaches 85%, so the first three principal component directions are extracted. The normalized feature matrix is ​​projected onto these three principal component directions to obtain the dimension-reduced key feature component matrix, with a dimension of N×3. Each row corresponds to the three key feature component values ​​of a metal particle. These three key feature components represent a comprehensive measure of the metal impurity's sensitivity to magnetic field gradients, its response to eddy current frequency, and its dependence on airflow velocity during the sorting process. Their values ​​directly correlate with the sorting efficiency.

[0034] Step S32: Input the key feature components into a sorting parameter mapping model based on deep reinforcement learning. The model uses sorting purity as the reward signal and outputs a cooperative action vector of magnetic field gradient, eddy frequency and airflow sorting cavity wind speed for the current material state. In this embodiment, the deep reinforcement learning sorting parameter mapping model consists of a policy network and a value network. The input layer of the policy network receives each row of the key feature component matrix, representing the three key feature component values ​​for each metal particle, and has three neurons. The policy network contains two hidden layers: the first hidden layer has 128 neurons with the ReLU activation function, and the second hidden layer has 64 neurons with the ReLU activation function. The output layer has three neurons, corresponding to the three action parameters: magnetic field gradient, eddy current frequency, and airflow sorting cavity wind speed. The output layer activation function is the Tanh function, mapping the output values ​​to the range of -1 to 1. The input layer of the value network also receives three key feature component values, plus the three action parameter values ​​output by the policy network, for a total of six inputs, and has six neurons. The value network contains two hidden layers: the first hidden layer has 128 neurons with the ReLU activation function, and the second hidden layer has 64 neurons with the ReLU activation function. The output layer has one neuron, outputting the expected reward value of the state-action pair. The reward signal is defined as the sorting purity, and the formula for calculating sorting purity is: ,in The number of correctly sorted metal particles. The total number of metal particles is represented by the value. The sorting purity ranges from 0% to 100%. During training, the policy network outputs action parameters based on the current key feature components. These action parameters are scaled and converted into actual physical parameters: the magnetic field gradient is scaled from 0.1 Tesla per meter to 0.5 Tesla per meter, the eddy current frequency from 80 kHz to 160 kHz, and the airflow separation chamber wind speed from 10 m / s to 20 m / s. After executing the action, the sorting purity value is obtained through simulation or actual sorting experiments. This value is used as a reward signal and fed back to the value network, which updates its parameters. The policy network adjusts its action output based on the evaluation results of the value network. The training iterations are set to 10,000 rounds, with each round processing the key feature components of a batch of 256 metal particles. After training, the policy network can directly output three action parameter values ​​based on the three input key feature component values, forming a three-dimensional cooperative action vector. The vector elements are the magnetic field gradient value, the eddy current frequency value, and the airflow separation chamber wind speed value, respectively.

[0035] Step S33: Based on the concentration data of micro-metal impurities detected in the current material flow, the safety threshold of the cooperative action vector is corrected to generate a dynamic cooperative control parameter set, which includes graded processing parameters for ferrous and non-ferrous metals.

[0036] In this embodiment, the micro-metal impurity concentration data is acquired in real time by a high-frequency pulsed eddy current sensor array. The sensor array scans the material layer on the conveyor belt at a rate of 2000 times per second, detecting the number of metal particles with a diameter less than 3 mm. The micro-metal impurity concentration is defined as the number of micro-metal particles per square meter of material layer, expressed in units of particles per square meter. The concentration data is updated every 0.5 seconds, and the updated concentration value is recorded as follows: The safety threshold is set at 50 micro-metal particles per square meter. When the concentration exceeds 50 units per square meter, the safety threshold correction process is triggered. The correction process first calculates the multiple by which the concentration exceeds the standard. Then, three corrections are made to the cooperative action vector: the magnetic field gradient correction formula is as follows. ,in This represents the original magnetic field gradient value. This is the corrected magnetic field gradient value, with an upper limit of correction. The coefficient 0.1 corresponds to the correction coefficient for the magnetic field gradient. In a laboratory setting, a small-scale experimental sorting line was constructed. Wood samples with known concentrations of 30, 50, 70, and 90 micro-metal particles per square meter were fed into the line, and the rejection rate of iron impurities under different magnetic field gradients was recorded. Experimental data showed that when the concentration of micro-metal impurities increased from the baseline value of 50 particles per square meter to 90 particles per square meter, the magnetic field gradient needed to be reduced by approximately 8% to 12% to ensure that the iron rejection rate did not decrease by more than 5 percentage points. Taking the median value of 10% as the correction ratio, i.e., reducing by 10% for every doubling of the standard, the coefficient was set to 0.1; the eddy current frequency correction formula is... ,in This is the original eddy current frequency value. The corrected eddy current frequency value, with an upper limit of correction. The coefficient 0.05 corresponds to 120% of the eddy current frequency correction factor. On the same test sorting line, sorting experiments on copper and aluminum fragments showed that increased concentration of micro-metallic impurities leads to insufficient repulsive force on fine copper wires in the eddy current field, resulting in decreased sorting efficiency. By comparing the sorting effects at different eddy current frequencies, it was found that for every doubling of the concentration, increasing the eddy current frequency by 4% to 6% can restore the copper and aluminum sorting efficiency to over 95% of the normal level. Taking the median value of 5% as the correction ratio, the coefficient is set to 0.05; the airflow separation chamber wind speed correction formula is... ,in This is the original wind speed value. The corrected wind speed value, with the upper limit of correction being... The coefficient 0.08 corresponds to a correction factor for the airflow separation chamber velocity, representing 130% of the total value. In airflow separation experiments, as the concentration of micro-metallic impurities increases, fine iron powder and copper shavings are easily encased by wood chips, requiring a stronger airflow to remove them. Measured data shows that for every doubling of the concentration, increasing the airflow velocity by 7% to 9% can improve the success rate of removing micro-metallic impurities to over 90%. The median value of 8% was chosen as the correction ratio, hence the coefficient was set to 0.08. The three corrected parameters form a dynamic collaborative control parameter set, which is divided into two parts: iron grading parameters and non-ferrous metal grading parameters. The iron grading parameters include the magnetic field gradient value and the airflow separation chamber velocity value. The magnetic field gradient value controls the magnetic field strength of the permanent magnet separator and the electromagnetic separator, while the airflow separation chamber velocity value controls the blowing force of the pneumatic baffle. The non-ferrous metal grading parameters include eddy current frequency and airflow separation chamber velocity. The eddy current frequency controls the excitation frequency of the eddy current separator, while the airflow separation chamber velocity controls the collection airflow velocity after copper and aluminum impurities are ejected. The dynamic collaborative control parameter set is updated every 0.5 seconds and sent to the actuator of the sorting equipment in real time, enabling online dynamic adjustment of the sorting parameters.

[0037] Furthermore, step S4 includes the following steps: Step S41: Parse the dynamic coordinated control parameter set and generate execution instruction data packets corresponding to the excitation current of the permanent magnet separator, the driving frequency of the eddy current coil, and the frequency of the centrifugal fan inverter, respectively. In this embodiment, the dynamic coordinated control parameter set includes two parts: iron grading parameters and non-ferrous metal grading parameters. The analysis operation begins by reading the magnetic field gradient value from the parameter set. The magnetic field gradient value is in Tesla per meter and is denoted as... Excitation current of permanent magnet separator The conversion relationship between the magnetic field gradient and the equipment calibration curve is determined by the equipment calibration curve equation. ,in The conversion factor is 0.8 amperes per tesla per meter. Substitute the values ​​into the equation to calculate the excitation current value, in amperes. Then read the eddy current frequency value, in hertz, denoted as . The driving frequency of the eddy current coil is directly equal to... No conversion is needed, but the frequency value must be converted to the corresponding pulse width modulation duty cycle. The duty cycle calculation formula is as follows: Where 200,000 is the highest operating frequency of the eddy current drive circuit, in Hertz. The calculated duty cycle ranges from 40% to 80%. Then, the airflow velocity in the air sorting chamber is read, in meters per second, and recorded as... The conversion relationship between the frequency of the centrifugal fan inverter and the wind speed is determined by the fan performance curve, and the performance curve equation is: Where K2 is the conversion factor, with a value of 3 Hz per meter per second. The inverter frequency value, in Hertz, is calculated by substituting the values ​​into the equation. The three execution parameter values ​​are combined into an execution command data packet. The packet format consists of three 16-bit unsigned integers arranged sequentially: the first integer is the integer result of multiplying the excitation current value by 100 and rounding it down; the second integer is the integer result of dividing the eddy current coil drive frequency value by 100 and rounding it down; and the third integer is the integer result of multiplying the centrifugal fan inverter frequency value by 10 and rounding it down. The total data packet length is 48 bits, with an additional 8-bit checksum byte. The checksum is calculated by adding the three integers and taking the lower 8 bits. Execution command data packets are generated every 0.5 seconds and immediately enter the transmission queue after each generation, awaiting transmission.

[0038] Step S42: Send the execution instruction data packet to the actuators of the iron remover, eddy current separator and air classifier unit via the real-time Ethernet protocol, and trigger the coordinated action sequence of the actuators. At the same time, start the encoder to provide real-time feedback on the equipment operating status. In this embodiment, the real-time Ethernet protocol adopts the industrial Ethernet communication standard, with a communication cycle set to 1 millisecond, completing one data exchange within each communication cycle. The execution instruction data packet is routed to three target devices via Ethernet switches: the permanent magnet separator's Ethernet interface IP address is set to 192.168.1.10, the eddy current separator's Ethernet interface IP address is set to 192.168.1.20, and the air separator's Ethernet interface IP address is set to 192.168.1.30. When sending the data packet, the target device's MAC address and IP address are written into the Ethernet frame header, the frame type field is set to 0x0800, and the data field carries 48 bits of data and an 8-bit checksum for the execution instruction data packet. After receiving the data packet, each device first checks if the checksum is correct; if the checksum passes, it writes the three execution parameter values ​​into its respective register. The permanent magnet separator's register address 0x01 stores the excitation current value, the eddy current separator's register address 0x02 stores the eddy current coil drive frequency value, and the air classifier's register address 0x03 stores the centrifugal fan inverter frequency value. After writing, all three devices simultaneously trigger their actions: the permanent magnet separator adjusts its excitation current to the target value within 10 milliseconds, the eddy current separator adjusts its eddy current frequency to the target value within 5 milliseconds, and the air classifier adjusts its fan speed to the target value within 20 milliseconds. The start-up time difference of the coordinated action sequence is controlled within 1 millisecond. Simultaneously, the encoders on each device begin collecting real-time operating status data: the permanent magnet separator's encoder monitors the magnetic roller speed, the eddy current separator's encoder monitors the belt offset, and the air classifier's encoder monitors the actual fan speed. The status data is updated every 1 millisecond and transmitted back to the controller via Ethernet protocol. The controller determines whether the equipment is operating normally based on the status data.

[0039] Step S43: Deploy an online metal detection probe at the sorting outlet to collect residual metal concentration data of the sorted material, and use the residual metal concentration data as real-time sorting purity feedback data to send back to the sorting parameter mapping model in step S3. In this embodiment, the online metal detection probe is installed at the sorting outlet. The probe consists of two sets of ring coils: one set is a transmitting coil, and the other is a receiving coil. The two sets of coils are coaxially mounted, with the axis perpendicular to the material flow direction. The transmitting coil continuously emits electromagnetic waves at a fixed frequency of 100 kHz, and the receiving coil detects the attenuation of the electromagnetic waves after passing through the material. When material containing metal impurities passes through the probe, the amplitude of the electromagnetic waves detected by the receiving coil changes, and the change is proportional to the volume and conductivity of the metal impurities. The probe collects 1000 sample points per second, with each sample point representing the amplitude of the output voltage of the receiving coil, in millivolts. After continuously collecting 100 sample points, the average value of these 100 sample points is calculated as the residual metal concentration index for that time period. The residual metal concentration index is in millivolts; a higher value indicates more residual metal. This index is updated every 0.1 seconds, and the updated value is recorded as follows: The residual metal concentration data is transmitted to the controller via a serial communication interface at a baud rate of 115200 bits per second. The data frame format consists of 1 start bit, 8 data bits, 1 stop bit, and no parity bit. The controller receives... Then, it is converted into a sorting purity feedback value, and the conversion formula is as follows: ,in The value represents the background noise amplitude measured by the probe in a material-free state, in millivolts. The typical value is 50 millivolts. The sorting purity feedback value ranges from 0% to 100%, and is updated every 0.1 seconds. This value is then sent back as real-time feedback data to the sorting parameter mapping model in step S3 for online updating of the model parameters.

[0040] Step S44: Compare the real-time sorting purity feedback data with the preset target purity threshold. If the deviation exceeds the allowable range, automatically adjust the learning rate of the sorting parameter mapping model to accelerate the convergence speed of subsequent parameter iterations.

[0041] In this embodiment, the preset target purity threshold is set to 98%, and the allowable deviation range is set to ±2%. Real-time sorting purity feedback value. A comparison is performed every 0.1 seconds with the target purity threshold. The comparison operation first calculates the deviation value. The unit is percentage points. If If the absolute value of the deviation is greater than 2 percentage points, meaning the sorting purity is below 96% or above 100%, then the deviation is considered to exceed the allowable range. When the deviation exceeds the allowable range, the learning rate of the sorting parameter mapping model is automatically adjusted. The adjustment rule for the learning rate is based on the magnitude of the absolute value of the deviation: if... The absolute value is between 2 and 5 percentage points, and the learning rate is adjusted from the base value of 0.001 to 0.005; if If the absolute value is between 5 and 10 percentage points, the learning rate is adjusted to 0.01; If the absolute value of the bias exceeds 10 percentage points, the learning rate is adjusted to 0.02. The adjusted learning rate takes effect immediately and is applied to the next parameter iteration update of the sorting parameter mapping model. The learning rate is continuously adjusted until the bias value reaches a certain threshold. The absolute value of the deviation falls back to within 2 percentage points. When the deviation returns to the allowable range, the learning rate is gradually restored to the base value of 0.001. The restoration process involves halving the learning rate every 0.5 seconds until it reaches 0.001. This dynamic adjustment mechanism ensures that the model can converge quickly when the sorting purity deviates from the target, while maintaining fine-tuning when the sorting purity is stable, avoiding parameter oscillations.

[0042] Furthermore, the step S42, which involves sending execution command data packets to the actuators of the iron remover, eddy current separator, and air classifier via real-time Ethernet protocol, includes the following steps: The execution instruction data packet is encapsulated into a protocol frame corresponding to the manufacturing message specification of IEC 61850-8-1, and a timestamp based on a precise time protocol is embedded in the protocol frame; In this embodiment, the execution instruction data packet consists of three 16-bit unsigned integers, corresponding to the excitation current value of the permanent magnet separator, the drive frequency value of the eddy current coil, and the frequency value of the centrifugal fan inverter, respectively. The total length of the data packet is 48 bits, with an additional 8-bit checksum byte. The encapsulation operation first fills the content of the execution instruction data packet into the application protocol data unit of the manufacturing message specification protocol. The format of the application protocol data unit follows the IEC61850-8-1 standard, where the data type field is set to 0x01, indicating that the message is a write request message; the data length field is set to 0x06, indicating that the subsequent data length is 6 bytes; the data field contains three 16-bit integers and a checksum byte, arranged in big-endian byte order, with the most significant byte first and the least significant byte last. After encapsulating the application protocol data unit, the header information of the manufacturing message specification protocol is added to its exterior. The header information includes a version number field set to 0x01 and a transaction identifier field set to the sequence number of the current message, which increments by 1 for each message frame sent. Next, a timestamp for the Precision Time Protocol (PTP) is added to the outside of the Manufacturing Message Specification Protocol (MCMP) frame. The timestamp consists of two parts: seconds and nanoseconds. The seconds are taken from the current Unix timestamp of the autonomous controller's internal clock, and the nanoseconds are taken from the current nanosecond counter value of the autonomous controller's internal clock. The timestamp resolution is 1 nanosecond. The timestamp is embedded in a reserved field of the protocol frame, which is 10 bytes long. The first 4 bytes store the seconds, and the last 6 bytes store the nanoseconds. The encapsulated protocol frame has a total length of 64 bytes, including a 14-byte Ethernet header, a 20-byte IP header, an 8-byte UDP header, a 10-byte Precision Time Protocol timestamp, an 8-byte Manufacturing Message Specification Protocol (MCMP) header, and a 6-byte Application Protocol Data Unit (API), with a 4-byte frame check sequence appended at the end.

[0043] The main controller broadcasts the encapsulated protocol frames to the industrial Ethernet switching network. The intelligent switches in the network forward the messages to the corresponding slave controllers of the iron remover, eddy current separator and air classifier based on the virtual LAN tag. In this embodiment, the main controller sends the encapsulated protocol frame to the industrial Ethernet switching network via an Ethernet port. The Ethernet port is configured in full-duplex mode with a transmission rate of 100 megabits per second. When the protocol frame is sent, the destination MAC address in the Ethernet header is set to the broadcast address FF-FF-FF-FF-FF-FF, the source MAC address is set to the main controller's MAC address 00-11-22-33-44-55, and the Ethernet type field is set to 0x0800. After the protocol frame enters the switching network, the first-level intelligent switch receives the broadcast frame. The intelligent switch maintains a virtual LAN tag mapping table containing three entries: virtual LAN tag 10 corresponds to the slave controller of the permanent magnet separator, virtual LAN tag 20 corresponds to the slave controller of the eddy current separator, and virtual LAN tag 30 corresponds to the slave controller of the air separator. The intelligent switch determines the VLAN tag based on the lower two bits of the transaction identifier field in the protocol header of the message specification protocol frame. A value of 00 assigns VLAN tag 10, 01 assigns VLAN tag 20, and 10 assigns VLAN tag 30. After assigning the VLAN tag, the intelligent switch inserts a 4-byte VLAN tag field into the Ethernet header of the protocol frame. This tag field contains the tag protocol identifier (0x8100) and tag control information. The priority field of the tag control information is set to 7, the canonical format indicator field is set to 0, and the VLAN identifier field is set to the assigned tag value. The protocol frame with the VLAN tag is then forwarded to the corresponding VLAN port, which is connected to the Ethernet interface of the corresponding slave controller via a fiber optic link.

[0044] After receiving the protocol frame, each slave controller parses out the timestamp mark and uses the hardware timestamp unit to achieve microsecond-level synchronization with the master clock; In this embodiment, after each slave controller receives a protocol frame with a VLAN tag, it first removes the VLAN tag field from the Ethernet header, restoring the original Ethernet frame format. Then, it extracts the precise time protocol timestamp from the reserved field of the protocol frame, obtaining the seconds and nanosecond values. The slave controller integrates a hardware timestamp unit, which includes a local clock counter with a counting frequency of 125 MHz and a counting period of 8 nanoseconds. Upon receiving the protocol frame, the hardware timestamp unit records the current count value of the local clock, denoted as . Simultaneously, the master clock seconds in the timestamp marker are parsed. and nanoseconds Convert the master clock time to a local clock count value. Calculate the time deviation between the local clock and the master clock. .like If the absolute value of the count exceeds 125 count units, or 1 microsecond, a clock synchronization adjustment is triggered. The adjustment operation increments or decrements the local clock counter. The system performs a counting operation to reduce the deviation between the local clock and the master clock to within one counting unit, or within 8 nanoseconds. After synchronization is complete, the slave controller records the local clock count at the moment of synchronization as the trigger reference time for subsequent command execution. The synchronization process is repeated every time a protocol frame is received, ensuring that the clock deviation between the slave controller and the master controller is always maintained within the microsecond range.

[0045] When the synchronization time arrives, each slave controller simultaneously parses the execution command and drives the corresponding power module to adjust the control quantity output to the actuator.

[0046] In this embodiment, after clock synchronization, each slave controller parses the three parameter values ​​of the execution instruction from the application protocol data unit of the protocol frame. The parsing operation first reads the first 16-bit integer from the data field, divides it by 100 to obtain the permanent magnet separator excitation current value in amperes; reads the second 16-bit integer, multiplies it by 100 to obtain the eddy current coil drive frequency value in hertz; and reads the third 16-bit integer, divides it by 10 to obtain the centrifugal fan inverter frequency value in hertz. After parsing the three parameter values, the slave controller waits for the synchronization moment. The synchronization moment is determined by a local clock counter. When the local clock count reaches the count value corresponding to the timestamp in the protocol frame plus a fixed transmission delay compensation value, the execution operation is triggered. The transmission delay compensation value is obtained by averaging multiple round-trip time measurements, typically 200 microseconds. At the instant the synchronization moment arrives, the slave controller simultaneously sends control signals to the three power modules. After receiving the excitation current value, the power module of the permanent magnet separator adjusts the duty cycle of the DC chopper circuit using pulse width modulation. The duty cycle calculation formula is as follows: ,in The target excitation current value, The maximum output current of the power module is set to 50 amps. After receiving the frequency value, the power module of the eddy current separator adjusts the inverter's switching frequency via a phase-locked loop circuit, ensuring the output AC frequency is precisely equal to the target frequency. The power module of the air classifier receives the inverter's frequency value and adjusts the output voltage and frequency of the three-phase inverter using a space vector pulse width modulation algorithm, driving the asynchronous motor to operate at the target speed. The three power modules start simultaneously at a synchronized moment, with a timing deviation controlled within 1 microsecond, ensuring that the iron remover, eddy current separator, and air classifier complete parameter adjustments at the same time, achieving true coordinated operation.

[0047] Furthermore, the method of achieving microsecond-level synchronization with the master clock using a hardware timestamp unit includes the following steps: The master controller periodically sends synchronization messages carrying a sending timestamp, and the slave controller records the local receiving timestamp of the message; In this embodiment, the main controller integrates a high-precision clock source with a crystal oscillator frequency of 125 MHz and a timing resolution of 8 nanoseconds. The main controller periodically generates synchronization messages at a frequency of 100 times per second, and the synchronization message format follows the event message structure of the Precision Time Protocol. Each time a synchronization message is generated, the main controller reads the current count value of the local clock. The count value includes a second counter and a nanosecond counter. The second counter records the number of seconds since 00:00 on January 1, 1970, and the nanosecond counter records the number of nanoseconds within the current second. The resolution of the nanosecond counter is 1 nanosecond. The main controller writes the read transmission timestamp T1 into the reserved field of the synchronization message. The reserved field is 10 bytes long; the first 4 bytes store the number of seconds, and the last 6 bytes store the number of nanoseconds. The message type field of the synchronization message is set to 0x00, indicating that the message is a synchronization event message. The synchronization message is broadcast to the industrial Ethernet switching network via the Ethernet port, which is configured in full-duplex mode with a transmission rate of 100 megabits per second. Upon receiving the first byte of the synchronization message, the Ethernet physical layer chip of the slave controller immediately triggers the hardware timestamp capture mechanism. The hardware timestamp unit reads the current count value of the local clock and records it as the local receive timestamp T2. The slave controller then receives the entire synchronization message and extracts the transmission timestamp T1 from the reserved field of the message. At this point, the slave controller simultaneously possesses both the transmission timestamp T1 and the local receive timestamp T2, both in nanoseconds with a precision of 1 nanosecond. These two timestamp values ​​are temporarily stored in the slave controller's register, awaiting further processing.

[0048] The slave controller immediately replies with a delay request message carrying its reception timestamp to the master controller, and the master controller records the reception time of the message; In this embodiment, after receiving the synchronization message and recording the local receive timestamp T2, the slave controller immediately constructs a delay request message. The message type field of the delay request message is set to 0x01, indicating that the message is a delay request event message. The slave controller writes the recently recorded local receive timestamp T2 into the reserved field of the delay request message. The reserved field is also 10 bytes long, with the first 4 bytes storing the number of seconds and the last 6 bytes storing the number of nanoseconds. The construction process of the delay request message is completed at the hardware level. When the Ethernet physical layer chip of the slave controller sends the first byte of the delay request message, it triggers the hardware timestamp capture mechanism again to record the current count value of the local clock as the sending timestamp T3 of the delay request message. T3 is also temporarily stored in the slave controller's register. The delay request message is sent to the industrial Ethernet switching network through the Ethernet port and forwarded to the master controller via the intelligent switch. When the master controller's Ethernet physical layer chip receives the first byte of the delay request message, it triggers a hardware timestamp capture mechanism to record the current count value of the master clock as the reception timestamp T4 of the delay request message. The master controller then extracts the slave controller's local reception timestamp T2 from the reserved field of the delay request message. At this point, the master controller has three timestamp values: the synchronization message transmission timestamp T1, the slave reception timestamp T2, and the delay request message reception timestamp T4. The slave controller has two timestamp values: the synchronization message reception timestamp T2 and the delay request message transmission timestamp T3. The master controller stores T1, T2, and T4 in a register, awaiting the next calculation.

[0049] The master controller calculates the link delay of the network transmission path based on the time difference between the two message exchanges, and sends the link delay to the slave controller through subsequent data packets; In this embodiment, after receiving the delay request message and extracting the timestamp T2, the main controller immediately calculates the link delay of the network transmission path. The link delay calculation is based on the delay request-response mechanism of the Precision Time Protocol, and the calculation formula is as follows: T1 is the timestamp for the master controller sending the synchronization message, T2 is the timestamp for the slave controller receiving the synchronization message, T3 is the timestamp for the slave controller sending the delay request message, and T4 is the timestamp for the master controller receiving the delay request message. All timestamps are in nanoseconds. Since the slave controller has not yet sent T3 to the master controller, the master controller first needs to obtain T3 through another message. The master controller constructs a delay response message, setting the message type field to 0x02, indicating that this message is a delay response message. The master controller writes the calculated link delay (Delay) value into the reserved field of the delay response message. The reserved field is 10 bytes long, with the first 4 bytes storing... The integer part of the value is stored in the last 6 bytes. The decimal part of the value is displayed in nanoseconds. The delay response message is sent to the corresponding slave controller via the Ethernet port. Upon receiving the delay response message, the slave controller extracts the link delay from the reserved field. The slave controller simultaneously reads the transmission timestamp T3 of the delay request message from its own register. The slave controller uses these four timestamps T1, T2, T3, and T4 to verify the correctness of the link delay calculation. The verification formula is (T4-T1)-(T3-T2) should equal twice the value. The value has an allowable error range of ±100 nanoseconds. After successful verification, the slave controller will... The value is stored in non-volatile memory for subsequent clock compensation calibration.

[0050] The slave controller uses the received link delay to compensate and calibrate the local hardware timestamp unit, thereby eliminating clock deviations caused by network jitter.

[0051] In this embodiment, the slave controller receives the link delay. After the value is calculated, the compensation and calibration process for the local hardware timestamp unit is immediately initiated. The hardware timestamp unit contains a programmable clock offset register, which defaults to 0. The compensation and calibration operation first calculates the time offset between the master clock and the local clock. The calculation formula is: Where T2 is the local timestamp of the slave controller receiving the synchronization message, and T1 is the timestamp of the master controller sending the synchronization message. This refers to link latency. The unit is nanoseconds; a positive value indicates that the local clock is faster than the master clock, and a negative value indicates that the local clock is slower than the master clock. The slave controller will... The value is written to the clock offset register of the hardware timestamp unit. The write operation is completed via the SPI bus, with a clock frequency of 10 MHz. Each time a timestamp is captured, the hardware timestamp unit automatically subtracts the value in the clock offset register from the captured local clock count. The corrected timestamp is obtained by calculating the value. The time deviation between the corrected timestamp and the master clock is controlled within ±100 nanoseconds. To eliminate residual deviations caused by network jitter, the slave controller performs a compensation calibration process every second, updating the clock offset register value after each calibration. During the calibration process, the slave controller also fine-tunes the crystal oscillator frequency of the local clock by an adjustment amount of [value missing]. The value is divided by the calibration period, which is 10 milliseconds, and the fine-tuning unit is nanoseconds per millisecond. Crystal oscillator frequency fine-tuning is achieved by adjusting the frequency divider coefficient of the phase-locked loop (PLL), with an adjustment step of 1 ppm (parts per million). After 10 consecutive calibrations, the frequency deviation between the local clock and the master clock is controlled within ±0.1 ppm, and the time deviation is controlled within ±50 nanoseconds, completely eliminating the impact of network jitter on clock synchronization accuracy.

[0052] Furthermore, the calculation of the link delay of the network transmission path based on the time difference between the two message exchanges includes the following steps: Define the time when the master controller sends the synchronization message as T1, the time when the slave controller receives it as T2, the time when the slave sends the delay request message as T3, and the time when the master controller receives it as T4. In this embodiment, the high-precision clock source inside the main controller operates at a frequency of 125 MHz, with a timing resolution of 8 nanoseconds. When generating a synchronization message, the main controller reads the current count value of the local clock and defines this moment as T1. The value of T1 consists of two parts: a second counter and a nanosecond counter. The second counter records the number of seconds since 00:00 on January 1, 1970, and the nanosecond counter records the number of nanoseconds within the current second. The resolution of the nanosecond counter is 1 nanosecond. The main controller writes T1 into a reserved field in the synchronization message. The reserved field is 10 bytes long, with the first 4 bytes storing the number of seconds and the last 6 bytes storing the number of nanoseconds. The synchronization message is broadcast to the industrial Ethernet switching network via the Ethernet port. When the Ethernet physical layer chip of the slave controller receives the first byte of the synchronization message, it triggers a hardware timestamp capture mechanism, reads the current count value of the local clock, and defines this moment as T2. T2 also consists of a second counter and a nanosecond counter, with a resolution of 1 nanosecond. After extracting T1 from the synchronization message, the slave controller immediately constructs a delay request message and writes T2 into the reserved field of the delay request message. When sending the first byte of the delay request message, the slave controller triggers the hardware timestamp capture mechanism again, reads the current count value of the local clock, and defines this moment as T3. The numerical structure of T3 is the same as T1 and T2. The delay request message is forwarded to the master controller via the Ethernet switching network. When the master controller's Ethernet physical layer chip receives the first byte of the delay request message, it triggers the hardware timestamp capture mechanism, reads the current count value of the master clock, and defines this moment as T4. The numerical structure of T4 is the same as the previous three timestamps. At this point, all four timestamps T1, T2, T3, and T4 have been recorded. T1 and T4 are stored in the master controller's register, and T2 and T3 are stored in the slave controller's register. The numerical precision of each timestamp is 1 nanosecond, and the time interval between the four timestamps is on the order of microseconds.

[0053] Calculate the difference between the downlink delay (T2-T1) from the master controller to the slave controller and the network jitter correction factor; In this embodiment, after receiving the delay request message and extracting T2, the master controller begins calculating the downlink delay. The downlink delay is defined as the time consumed for a synchronization message to travel from the master controller to the slave controller. When calculating the downlink delay, the difference between T2 and T1 is first calculated, i.e. =T2-T1, in nanoseconds. This includes the propagation time of synchronization messages in the Ethernet cable, the queuing time and processing time in the intelligent switch, and the receive delay of the physical layer chip of the slave controller. Network jitter correction factor. Used to eliminate the uncertainty of switch queuing time. The value of is determined through statistical methods. The main controller records the value of each of the past 100 synchronization message exchanges. Values, calculate these 100 Standard deviation of values Network jitter correction factor Set to 3 times the standard deviation, i.e. The formula for calculating downlink latency is as follows: .like Less than Then Set to 0. Downlink latency is measured in nanoseconds, typically ranging from 1000 to 5000 nanoseconds. The main controller will calculate... The value is temporarily stored in the register, waiting to be averaged together with the uplink delay.

[0054] Calculate the difference between the uplink delay (T4-T3) from the station controller to the master controller and the network jitter correction factor; In this embodiment, after calculating the downlink delay, the master controller continues to calculate the uplink delay. The uplink delay is defined as the time it takes for a delay request message to travel from the slave controller to the master controller. When calculating the uplink delay, the difference between T4 and T3 is calculated first. =T4-T3, in nanoseconds. This includes the propagation time of the delay request message in the Ethernet cable, the queuing time and processing time in the intelligent switch, and the receive delay of the main controller's physical layer chip. Network jitter correction factor. This was also determined using statistical methods. The main controller recorded the delay request message exchanged in the past 100 exchanges. Values, calculate these 100 Standard deviation of values Network jitter correction factor Set to 3 times the standard deviation, i.e. The formula for calculating uplink latency is as follows: .like Less than Then Set to 0. Uplink latency is measured in nanoseconds, typically ranging from 1000 to 5000 nanoseconds. The main controller will calculate... The value is temporarily stored in the register, along with the downlink delay. Together they are used for subsequent average calculations.

[0055] The average of the downlink delay and the uplink delay is taken as the link delay of the network transmission path.

[0056] In this embodiment, the main controller obtains the downlink delay. and uplink latency Then, the average of the two values ​​is calculated as the link delay of the network transmission path. The formula for calculating the link delay is: The unit is nanoseconds. If and If the difference exceeds 1000 nanoseconds, it is determined that there is an anomaly in the network link. The main controller discards the calculation result and re-initiates a round of synchronization message and delay request message exchange process until... and The difference is maintained until it is less than 1000 nanoseconds. Under normal circumstances, and The difference is typically within 200 nanoseconds because the uplink and downlink of industrial Ethernet switching networks are symmetrical. The calculated link delay... The value is written to the reserved field of the delayed response message. The reserved field is 10 bytes long, with the first 4 bytes storing... The integer part of the value is stored in the last 6 bytes. The decimal part of the value is displayed in nanoseconds with a resolution of 1 nanosecond. The delay response message is sent to the corresponding slave controller via the Ethernet port. Upon receiving the delay response message, the slave controller extracts the link delay from the reserved field. The value is recorded and stored in non-volatile memory. The slave controller utilizes... The local hardware timestamp unit is compensated and calibrated, and the time deviation between the calibrated local clock and the master clock is controlled within ±100 nanoseconds. The entire link delay calculation process is repeated 100 times per second to ensure that the link delay value always reflects the latest changes in the current network status.

[0057] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A smart optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities, characterized in that, Includes the following steps: Step S1: Simultaneously collect the electromagnetic induction attenuation spectrum and spectral reflection characteristics of metal impurities in the waste wood flow through a high-frequency pulsed eddy current sensor array and a multispectral imaging unit, and combine them with the laser profile scanning data of the material conveyor belt to construct a three-dimensional feature tensor containing metal particle size, material type and spatial coordinates. Step S2: Based on the three-dimensional feature tensor, the energy dissipation characteristics and phase shift characteristics of metal impurities in time-varying magnetic field and eddy current field are extracted using a multi-physics field coupling analytical engine to generate a nonlinear response feature matrix characterizing the degree of coupling between metal physical properties and wood matrix interference. Step S3: Input the nonlinear response feature matrix into the pre-trained sorting parameter mapping model, which outputs a set of dynamic coordinated control parameters for the magnetic field gradient, eddy current frequency, and airflow sorting chamber wind speed for the current material state. Step S4: Based on the dynamic collaborative control parameter set, control commands are sent to the actuators of the iron remover, eddy current separator and air classifier via the industrial bus protocol, and the sorting parameter mapping model is iteratively corrected based on real-time sorting purity feedback data.

2. The intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities, as described in claim 1, is characterized in that... Step S1 includes the following steps: Step S11: Drive the high-frequency pulse eddy current sensor array to scan the width direction of the conveyor belt with a preset overlap coverage rate, collect the eddy current decay time series signal caused by metal impurities at different depths, and extract the peak amplitude and decay time constant in the eddy current decay time series signal. Step S12: Synchronously trigger the multispectral imaging unit to collect diffuse reflectance spectral data of the material surface, and use the spectral feature extraction algorithm to separate the absorption peak of wood cellulose and the characteristic peak of metal oxide film to generate material discrimination feature vector; Step S13: Fuse the eddy current decay time series signal with the material discrimination feature vector, and combine it with the material layer thickness distribution map obtained by the laser profile scanner to generate a three-dimensional feature tensor containing metal particle size, material type and spatial coordinates through a three-dimensional spatial reconstruction algorithm.

3. The intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities, as described in claim 1, is characterized in that... Step S2 includes the following steps: Step S21: Extract the skin effect parameters of metallic impurities under high-frequency alternating magnetic field and the Lorentz force response parameters in eddy current field from the three-dimensional feature tensor, and perform decoupling processing based on the wood dielectric constant interference model. Step S22: Construct a multiphysics coupling model based on finite element analysis, and use the decoupled skin effect parameters and Lorentz force response parameters as boundary condition inputs to simulate the trajectory evolution of metal particles under the combined action of magnetic field, eddy current field and airflow field to generate simulation results; Step S23: Analyze the differences in the separation trajectories of metal particles of different materials in the simulation results, and extract the energy dissipation characteristics that characterize the dispersion of their motion trajectories and the phase shift characteristics that characterize the phase synchronization. Step S24: Tensor-join the energy dissipation features and phase shift features, and use an attention mechanism network to enhance the key feature dimensions to generate a nonlinear response feature matrix that characterizes the coupling degree between the physical properties of the metal and the interference of the wood matrix.

4. The intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities according to claim 3, characterized in that, Step S22 includes the following steps: Step S221: Establish the coupled control equations for the electromagnetic field module described by Maxwell's equations, the fluid field module described by Navier-Stokes equations, and the particle motion module described by the discrete phase model. Based on the coupled control equations, construct a multiphysics coupled model based on finite element analysis. Step S222: Import the actual physical property parameters of the waste wood, including density, moisture content and dielectric constant, and use them as variable medium property inputs in the fluid field module; Step S223: Based on the metal grain size and spatial coordinate data in the three-dimensional feature tensor, initialize multiple discrete metal particle phases with different electrical conductivity and magnetic permeability in the model computation domain; Step S224: The decoupled skin effect parameters and Lorentz force response parameters are used as boundary condition inputs, and a multiphysics coupling model is solved based on discrete metal particle phases to simulate the evolution of the motion trajectory of metal particles under the combined action of magnetic field, eddy current field and airflow field, so as to generate simulation results.

5. The intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities according to claim 4, characterized in that, The initialization of multiple discrete metal particle phases with different electrical and magnetic conductivity in the model computation domain as described in step S223 includes the following steps: The three-dimensional feature tensor is analyzed to identify the distribution areas of metal impurities of different particle sizes and materials in large iron, copper and aluminum fragments and iron powder, so as to generate identification results. Based on the identification results, the standard electrical conductivity and magnetic permeability parameters of the corresponding material are retrieved from the preset metal physical property library, and the random distribution of particle size of the particle group is generated based on the particle size distribution data. By utilizing the random distribution of particle size and standard physical property parameters, discrete phase particles representing different metallic impurities are randomly generated within the computational domain. Each discrete phase particle carries its unique electrical conductivity, magnetic permeability, and initial spatial coordinate properties. The set of discrete phase particles is defined as the initial particle phase boundary condition of the multiphysics coupling model.

6. The intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities, as described in claim 1, is characterized in that... Step S3 includes the following steps: Step S31: Normalize the nonlinear response feature matrix and map it to a high-dimensional feature space. Use kernel principal component analysis to extract the key feature components that dominate the sorting efficiency. Step S32: Input the key feature components into a sorting parameter mapping model based on deep reinforcement learning. The model uses sorting purity as the reward signal and outputs a cooperative action vector of magnetic field gradient, eddy frequency and airflow sorting cavity wind speed for the current material state. Step S33: Based on the concentration data of micro-metal impurities detected in the current material flow, the safety threshold of the cooperative action vector is corrected to generate a dynamic cooperative control parameter set, which includes graded processing parameters for ferrous and non-ferrous metals.

7. The intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Parse the dynamic coordinated control parameter set and generate execution instruction data packets corresponding to the excitation current of the permanent magnet separator, the driving frequency of the eddy current coil, and the frequency of the centrifugal fan inverter, respectively. Step S42: Send the execution instruction data packet to the actuators of the iron remover, eddy current separator and air classifier unit via the real-time Ethernet protocol, and trigger the coordinated action sequence of the actuators. At the same time, start the encoder to provide real-time feedback on the equipment operating status. Step S43: Deploy an online metal detection probe at the sorting outlet to collect residual metal concentration data of the sorted material, and use the residual metal concentration data as real-time sorting purity feedback data to send back to the sorting parameter mapping model in step S3. Step S44: Compare the real-time sorting purity feedback data with the preset target purity threshold. If the deviation exceeds the allowable range, automatically adjust the learning rate of the sorting parameter mapping model to accelerate the convergence speed of subsequent parameter iterations.

8. The intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities according to claim 7, characterized in that, The step S42, which involves sending execution command data packets to the actuators of the iron separator, eddy current separator, and air classifier via real-time Ethernet protocol, includes the following steps: The execution instruction data packet is encapsulated into a protocol frame corresponding to the manufacturing message specification of IEC 61850-8-1, and a timestamp based on a precise time protocol is embedded in the protocol frame; The main controller broadcasts the encapsulated protocol frames to the industrial Ethernet switching network. The intelligent switches in the network forward the messages to the corresponding slave controllers of the iron remover, eddy current separator and air classifier based on the virtual LAN tag. After receiving the protocol frame, each slave controller parses out the timestamp mark and uses the hardware timestamp unit to achieve microsecond-level synchronization with the master clock; When the synchronization time arrives, each slave controller simultaneously parses the execution command and drives the corresponding power module to adjust the control quantity output to the actuator.

9. The intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities according to claim 8, characterized in that, The method of using a hardware timestamp unit to achieve microsecond-level synchronization with the master clock includes the following steps: The master controller periodically sends synchronization messages carrying a sending timestamp, and the slave controller records the local receiving timestamp of the message; The slave controller immediately replies with a delay request message carrying its reception timestamp to the master controller, and the master controller records the reception time of the message; The master controller calculates the link delay of the network transmission path based on the time difference between the two message exchanges, and sends the link delay to the slave controller through subsequent data packets; The slave controller uses the received link delay to compensate and calibrate the local hardware timestamp unit, thereby eliminating clock deviations caused by network jitter.

10. The intelligent optimization method for sorting waste wood based on the coupling of physical characteristics of metal impurities according to claim 9, characterized in that, The calculation of the link delay of the network transmission path based on the time difference between two message exchanges includes the following steps: Define the time when the master controller sends the synchronization message as T1, the time when the slave controller receives it as T2, the time when the slave sends the delay request message as T3, and the time when the master controller receives it as T4. Calculate the difference between the downlink delay (T2-T1) from the master controller to the slave controller and the network jitter correction factor; Calculate the difference between the uplink delay (T4-T3) from the station controller to the master controller and the network jitter correction factor; The average of the downlink delay and the uplink delay is taken as the link delay of the network transmission path.

Images