AI-based precious metal processing process anomaly detection system

Abstract

The application relates to the technical field of production process monitoring, in particular to an AI-based noble metal processing process anomaly detection system, which comprises an acoustic acquisition module, a crystal grain perception module, an energy modeling module, an impedance fusion module and an anomaly judgment module.In the application, an energy propagation network with a spatial topological relationship is constructed by fusing the acoustic wave frequency domain features and the crystal grain micro-morphology parameters, the dynamic changes of microphysics are mapped into node impedance and energy distribution gradient, multi-dimensional heterogeneous data are converted into deep fusion vectors reflecting the essential characteristics of the process, multi-order feature aggregation and nonlinear space mapping are carried out through the graph structure, the deep coupling correlation between the multi-physical parameters is extracted, the limitations and hysteresis disadvantages of the traditional single hard threshold judgment are broken, adaptive boundary division and physical state category discrimination are realized in the face of complex dynamic working conditions, the false alarm and missed alarm phenomena in state monitoring are effectively eliminated, and the continuous safe operation of the core manufacturing process is comprehensively ensured.

Description

Technical Field

[0001] This invention relates to the field of production process monitoring technology, and in particular to an AI-based anomaly detection system for precious metal processing. Background Technology

[0002] The field of production process monitoring technology is a comprehensive engineering category that involves real-time acquisition and dynamic analysis of equipment operating status, process parameters, and environmental conditions in industrial production and manufacturing processes. Its core aspects include the acquisition of physical quantities at the sensor level, wired communication via fieldbus, real-time database reading and writing, and status chart display. Overall, this field acquires physical environmental parameters by deploying field instruments such as temperature sensors, pressure transmitters, and vibration probes, and converts electrical signals into discrete digital quantities through programmable logic controllers. Then, it uses industrial Ethernet to aggregate the field-acquired parameters to a host computer server for centralized comparison and threshold judgment.

[0003] Among them, the traditional precious metal processing anomaly detection system refers to a monitoring system established for physical phenomena that deviate from standard process specifications, such as temperature exceeding limits, material fracture, or surface defects, that occur in the manufacturing processes of high-value metals such as gold, silver, and platinum during smelting, forging, rolling, and polishing. This system uses thermocouples to collect the actual temperature value of the molten metal in the furnace, uses photoelectric switches to record the time and position coordinates of the material on the conveyor belt, and uses an ultrasonic flaw detector to emit high-frequency sound waves into the metal ingot and receive the acoustic characteristics of the reflected echo. Then, the above physical measurement values ​​are compared with the upper and lower limit fixed values ​​preset in the microcontroller register. When the measured value exceeds the set value boundary, it directly drives the relay to close to connect the power supply circuit of the alarm indicator.

[0004] Traditional monitoring systems rely on a single measuring device to obtain local state parameters, directly comparing discrete physical parameters with preset fixed numerical boundaries. This single-dimensional hard threshold judgment mechanism cannot capture the complex coupling relationships between multiple parameters in the production process, which can easily lead to process fluctuations being misjudged as abnormal faults. Moreover, fixed threshold judgment lacks the ability to adaptively adjust to dynamic operating conditions, making it difficult to achieve in-depth state early warning when facing the evolution of highly complex production microstructures. This can easily lead to missed anomalies and delayed early warnings, causing instability risks in product manufacturing quality. Summary of the Invention

[0005] To address the technical problem that traditional monitoring systems rely on a single measuring device to obtain local state parameters and directly compare discrete physical parameters with preset fixed numerical boundaries, this one-dimensional hard threshold judgment mechanism cannot capture the complex coupling relationships between multiple parameters in the production process. This can easily lead to process fluctuations being misjudged as abnormal faults. Moreover, fixed threshold judgment lacks the ability to adaptively adjust to dynamic operating conditions. When faced with the evolution of highly complex production microstructures, it is difficult to achieve in-depth state early warning, which can easily lead to missed anomalies and delayed early warnings, causing instability risks in product manufacturing quality. This invention provides an AI-based anomaly detection system for precious metal processing.

[0006] On the one hand, an AI-based anomaly detection system for precious metal processing is provided, which includes: The acoustic acquisition module acquires ultrasonic transmission and reflection signals from precious metal processing services and converts them into digital signals. It performs discrete Fourier transform to extract frequency domain amplitude components, calculates the frequency domain amplitude ratio and signal delay span, and constructs an acoustic propagation parameter matrix. The grain sensing module acquires images of the microstructure of noble metals, performs grayscale processing and segments grain boundaries, calculates the ratio of grain outline perimeter to area, measures the angle between the centroids of adjacent grains, and generates grain structure feature data. The energy modeling module, based on the grain structure feature data and the acoustic propagation parameter matrix, sets the grain centroid as a node, calculates the Euclidean distance between adjacent nodes as the side length, extracts the propagation drag coefficient and attenuation rate as the side weights, and constructs an energy propagation network model. The impedance fusion module extracts the energy allocation values ​​of the nodes in the energy propagation network model, uses the difference in energy allocation values ​​of adjacent nodes as the impedance gradient, and performs a weighted summation with the weight parameters to generate a multi-dimensional feature vector. The anomaly detection module extracts the multidimensional feature vector, inputs it into a graph convolutional neural network to perform feature aggregation and nonlinear spatial mapping to obtain classification labels, and outputs anomaly detection results by comparing them with a preset physical state benchmark.

[0007] As a further aspect of the present invention, the acoustic propagation parameter matrix includes acoustic attenuation coefficient, acoustic phase velocity, and ultrasonic group velocity; the grain structure characteristic data includes average grain size, grain boundary density, and shape factor; the energy propagation network model includes network adjacency matrix, node degree matrix, and graph Laplacian matrix; the multidimensional feature vector includes acoustic impedance, energy spectral density, and feature root mean square value; and the anomaly detection results include defect spatial location, damage severity, and remaining workpiece life.

[0008] As a further aspect of the present invention, the acoustic acquisition module includes: The signal conversion submodule acquires the ultrasonic transmission signal and reflected echo signal of the precious metal processing service, converts the ultrasonic transmission signal and reflected echo signal of the precious metal processing service into a discrete digital sequence, performs a Fourier transform on the discrete digital sequence to extract the amplitude value, and generates the frequency domain amplitude component. The delay extraction submodule, based on the frequency domain amplitude component, divides the frequency domain amplitude component by the frequency domain reference value to obtain the amplitude division ratio, reads the difference between the transmission and return times and assigns it to the signal delay span, and generates a span parameter vector by combining the amplitude division ratio and the signal delay span; The parameter construction submodule calls the span parameter vector, calculates the product of the span parameter vector and the preset attenuation compensation coefficient value, maps the product to a preset blank multidimensional array according to the row and column dimensions, and establishes the acoustic propagation parameter matrix.

[0009] As a further aspect of the present invention, the grain sensing module includes: The boundary segmentation submodule acquires images of the microstructure of precious metals, converts the images into a single-channel grayscale matrix, extracts the pixel step position separation region within the single-channel grayscale matrix, and generates a set of grain boundary pixels. The morphological parameter extraction submodule calls the grain boundary pixel set, accumulates the edge pixel count to obtain the grain outline perimeter, counts the internal pixel points to obtain the cross-sectional area, divides the grain outline perimeter by the cross-sectional area to obtain the outline ratio value, measures the intersection angle of the centroid connection line of adjacent grains to obtain the centroid connection line angle, and generates morphological space parameter values. The structural feature statistics submodule, based on the morphological space parameter value, accumulates the number of independent grains within the field of view to obtain the number of grains in the region, and writes the number of grains in the region and the morphological space parameter value into the feature record matrix according to the grain index number to generate grain structure feature data.

[0010] As a further aspect of the present invention, the step position separation region of pixels within a single-channel grayscale matrix is ​​extracted by calculating the difference in grayscale values ​​between adjacent pixels in the single-channel grayscale matrix and marking the positions of pixels whose grayscale value difference exceeds a preset threshold as boundary candidate points. The preset threshold ranges from 10% to 20% of the grayscale dynamic range of the single-channel grayscale matrix.

[0011] As a further aspect of the present invention, the energy modeling module includes: The node distance calculation submodule extracts the grain centroid coordinates in the grain structure feature data as spatial distribution nodes, performs linear span association mapping based on the spatial absolute position features of adjacent spatial distribution nodes in a preset coordinate system, and generates node Euclidean distance vectors. The edge weight extraction submodule, based on the node Euclidean distance vector and the acoustic propagation parameter matrix, reads the stagnation value and the decrease ratio at the spatially distributed nodes in the acoustic propagation parameter matrix, and appends the product of the stagnation value and the decrease ratio to the corresponding position of the node Euclidean distance vector to generate the propagation edge weight matrix. The network topology construction submodule calls the propagation edge weight matrix, extracts the non-zero elements in the propagation edge weight matrix as interconnected edges, and connects the spatially distributed nodes sequentially according to the interconnected edges to establish an energy propagation network model.

[0012] As a further aspect of the present invention, extracting the non-zero element positions in the propagation edge weight matrix as interconnected edges refers to traversing all elements of the propagation edge weight matrix, recording the row and column index pairs with non-zero element values ​​as endpoint index pairs of interconnected edges, and arranging the endpoint index pairs in the row priority order in the propagation edge weight matrix to form an interconnected edge index sequence. A directed connection relationship is established between the two spatially distributed nodes corresponding to each endpoint index pair in the interconnected edge index sequence. The set of all directed connection relationships constitutes the edge set of the energy propagation network model.

[0013] As a further aspect of the present invention, the impedance fusion module includes: The node extraction submodule extracts the energy allocation values ​​of the nodes in the energy propagation network model, and performs mapping on the energy allocation values ​​of the nodes in the energy propagation network model according to the topological space sequence, integrating the node energy values ​​into an array range to generate an energy sequence vector; The gradient calculation submodule calls the energy sequence vector, extracts the parameter element pairs of adjacent topological connection states in the energy sequence vector, performs subtraction operation on the parameter element pairs to obtain the node energy range index, and integrates the node energy range index to perform alignment and arrangement to generate the node impedance gradient matrix. The feature fusion submodule, based on the node impedance gradient matrix, calls the corresponding weight parameters of the energy propagation network model, reads the numerical relationship between the weight parameters and the elements of the node impedance gradient matrix, multiplies and accumulates the values ​​of the node impedance gradient matrix elements with the corresponding weight parameters, and generates a multidimensional feature vector.

[0014] As a further aspect of the present invention, the anomaly determination module includes: The feature aggregation submodule extracts the multidimensional feature vector, collects a preset topological space connection matrix, performs multiplication calculations between the elements in the multidimensional feature vector and the elements in the topological space connection matrix, extracts the calculated adjacent node parameters, performs accumulation, and generates a node feature aggregation matrix. The spatial mapping submodule detects the activation function mapping parameters based on the node feature aggregation matrix, performs mapping on the node feature aggregation matrix using the activation function mapping parameters, calculates the deviation of the mapping amount from the preset state baseline value, delineates the classification boundary based on the deviation amount, and generates a state classification label value. The status comparison submodule calls the status classification label value to obtain the physical operating status quantity of the equipment, compares the numerical difference between the status classification label value and the physical operating status quantity of the equipment, determines the risk category when the numerical difference exceeds the safety tolerance limit, and generates the abnormal detection result of the precious metal processing service.

[0015] As a further aspect of the present invention, the step of defining the classification boundary based on the deviation magnitude refers to mapping low-risk deviation level, medium-risk deviation level, and high-risk deviation level to preset classification boundary segments respectively. The classification boundary segments are divided with a fixed numerical step size as the interval. Based on the classification boundary segment to which the mapped amount belongs, a state classification label value corresponding one-to-one with the deviation magnitude level is generated.

[0016] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By integrating acoustic frequency domain characteristics with grain microstructure parameters, an energy propagation network with spatial topological relationships is constructed. The dynamic changes in microphysics are mapped to node impedance and energy distribution gradients. Multi-dimensional heterogeneous data are converted into deeply fused vectors that directly reflect the essential characteristics of the process. Through graph structure, multi-order feature aggregation and nonlinear spatial mapping are performed to accurately extract the deep coupling correlation between multiple physical parameters. This completely breaks through the limitations and lag of traditional single hard threshold judgment, and achieves adaptive boundary division and accurate identification of physical state categories in the face of complex dynamic working conditions. It effectively eliminates false alarms and missed alarms in condition monitoring and comprehensively ensures the continuous and safe operation of the core manufacturing process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0018] Figure 1 This is a system schematic diagram of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the acoustic acquisition module in this invention; Figure 4 This is a flowchart of the grain sensing module in this invention; Figure 5 This is a flowchart of the energy modeling module in this invention; Figure 6 This is a flowchart of the impedance fusion module in this invention; Figure 7 This is a flowchart of the anomaly detection module in this invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] This invention provides an AI-based anomaly detection system for precious metal processing, such as... Figure 1-2 The diagram shown illustrates an AI-based anomaly detection system for precious metal processing. The system includes: The acoustic acquisition module acquires ultrasonic transmission and reflection signals from precious metal processing services and converts them into digital signals. It performs discrete Fourier transform to extract frequency domain amplitude components, calculates the frequency domain amplitude ratio and signal delay span, and constructs an acoustic propagation parameter matrix. The grain sensing module acquires images of the microstructure of noble metals, performs grayscale processing and segments grain boundaries, calculates the ratio of grain outline perimeter to area, measures the angle between the centroids of adjacent grains, and generates grain structure feature data. The energy modeling module, based on grain structure feature data and acoustic propagation parameter matrix, sets the grain centroid as a node, calculates the Euclidean distance between adjacent nodes as the side length, extracts the propagation drag coefficient and attenuation rate as edge weights, and constructs an energy propagation network model. The impedance fusion module extracts the energy allocation values ​​of nodes in the energy propagation network model, uses the difference in energy allocation values ​​between adjacent nodes as the impedance gradient, and performs a weighted summation with the weight parameters to generate a multi-dimensional feature vector. The anomaly detection module extracts multi-dimensional feature vectors, inputs them into a graph convolutional neural network to perform feature aggregation and nonlinear spatial mapping to obtain classification labels, and outputs anomaly detection results by comparing them with a preset physical state benchmark.

[0022] The acoustic propagation parameter matrix includes acoustic attenuation coefficient, acoustic phase velocity, and ultrasonic group velocity; the grain structure characteristic data includes average grain size, grain boundary density, and shape factor; the energy propagation network model includes network adjacency matrix, node degree matrix, and graph Laplacian matrix; the multidimensional feature vectors include acoustic impedance, energy spectral density, and feature root mean square value; and the anomaly detection results include defect spatial location, damage severity, and remaining workpiece life.

[0023] Specifically, such as Figure 2, 3 As shown, the acoustic acquisition module includes: The signal conversion submodule acquires the ultrasonic transmission signal and reflected echo signal of the precious metal processing service, converts the ultrasonic transmission signal and reflected echo signal of the precious metal processing service into a discrete digital sequence, performs a Fourier transform on the discrete digital sequence to extract the amplitude value, and generates the frequency domain amplitude component. The signal conversion submodule retrieves the ultrasonic transmission signal and reflected echo signal transmitted from the precious metal processing and testing equipment via the data acquisition interface. The ultrasonic transmission signal originates from a piezoelectric ceramic transducer with a main frequency of 5 MHz, and the reflected echo signal is the acoustic waveform data after reflection from the grain interfaces within the precious metal. After receiving the continuous waveform analog data, the analog-to-digital conversion unit of the signal conversion submodule performs periodic discretization processing at a fixed sampling rate of 100 MHz. Each sampling duration is 2 milliseconds, converting the continuous signal into a discrete digital sequence containing 200,000 data nodes. The signal conversion submodule then processes this discrete digital sequence... The sequence performs discrete Fourier transform logic operations. The operation logic is to calculate the sum of the data of each sampling point in the discrete digital sequence multiplied by the rotation factor, to obtain the complex number result corresponding to each frequency interval, and then calculate the positive square root of the sum of the squares of the real part and the imaginary part of the complex number result, which is used as the amplitude value of the corresponding frequency. In this operation, the frequency domain resolution parameter is set to 0.01 MHz. This resolution value is obtained by comparing the spectral aliasing of 100 precious metal samples with different grain sizes. When the resolution is greater than 0.02 MHz, it is impossible to distinguish the frequency components of adjacent diffraction waves. Therefore, it is set to 0.01 MHz to ensure analytical accuracy. For example, for a certain time-domain sampled sequence, the real part value at 5 MHz is calculated to be 3 volts and the imaginary part value is 4 volts. After square root operation, the amplitude value is obtained as 5 volts. All amplitude values ​​in the specified frequency band are combined and arranged in ascending order of frequency to generate frequency domain amplitude components. The advantage of this operation logic is that it accurately separates the energy distribution state of each frequency in the signal by analyzing the complex modulus, laying a high-fidelity data foundation for the dynamic evaluation of material uniformity in key processes in processing services.

[0024] The delay extraction submodule, based on the frequency domain amplitude component, divides the frequency domain amplitude component by the frequency domain reference value to obtain the amplitude division ratio, reads the difference between the transmission and return times and assigns it to the signal delay span, and generates a span parameter vector by combining the amplitude division ratio and the signal delay span; The delay extraction submodule retrieves the frequency domain amplitude component output by the front-end component and reads the frequency domain reference value from the local storage unit. This frequency domain reference value is the average amplitude data obtained from 50 ultrasonic transmission tests on a standard pure gold control sample without internal structural defects under a standard constant temperature environment of 20 degrees Celsius. The frequency domain amplitude component of the current test sample is divided by this frequency domain reference value, and the quotient calculation logic is executed to obtain the amplitude division ratio. For example, when the frequency domain amplitude component is 5 volts and the frequency domain reference value is 10 volts, the calculated amplitude division ratio is 0.5. Subsequently, the timestamp of the ultrasonic wave emission recorded by the transmission synchronization trigger is read, and the timestamp of the return recorded by the echo peak detector is read. The difference between the return timestamp and the emission timestamp is obtained, and this difference is directly assigned to the signal delay span. For example, if the transmission time is 10 microseconds and the return time is 30 microseconds, the calculated signal delay span is 20 microseconds. Combining the aforementioned amplitude division ratio and the signal delay span, a weighted summation calculation logic is performed to generate a span parameter vector. The weighting coefficient for the amplitude division ratio is set to 0.6, and the weighting coefficient for the benchmarked signal delay span is set to 0.4. This setting is based on the fact that parameter sensitivity tests on 200 samples of different thicknesses revealed that amplitude changes have a 50% higher sensitivity score to internal material damage than delay changes. Therefore, the parameter vector is assigned... With a weighting coefficient of 0.6, the amplitude division ratio of 0.5 is multiplied by the weighting coefficient of 0.6 to obtain 0.3. The signal delay span of 20 microseconds is benchmarked to 2 and then multiplied by the weighting coefficient of 0.4 to obtain 0.8. The sum of the two results in a span parameter vector of 1.1. The advantage of this operation logic is that it comprehensively characterizes the beam propagation hindrance state by integrating the dual acoustic indicators of amplitude attenuation and time delay. The establishment of this acoustic hindrance index system directly provides a highly sensitive reference constraint for the closed-loop calibration of process parameters in downstream automated processing services.

[0025] The parameter construction submodule calls the span parameter vector, calculates the product of the span parameter vector and the preset attenuation compensation coefficient value, maps the product to the preset blank multidimensional array according to the row and column dimensions, and establishes the acoustic propagation parameter matrix. The parameter construction submodule extracts the generated span parameter vector 1.1 and retrieves the preset attenuation compensation coefficient value from the system configuration database. This preset attenuation compensation coefficient value depends on the specific material density of the precious metal. For the gold material currently being processed, with a density of 19.3 grams per cubic centimeter, the preset attenuation compensation coefficient value was determined to be 1.05 through previous attenuation calibration experiments on 10 gold plates of different thicknesses. The multiplication logic is executed, multiplying the span parameter vector 1.1 by the preset attenuation compensation coefficient value 1.05, resulting in an attenuation correction value of 1.155. Subsequently, the parameter construction submodule initializes a preset blank multidimensional array with a dimension of 10 rows and 10 columns in memory. The initial internal elements of this multidimensional array are all assigned a value of 0. The parameter construction submodule obtains the two-dimensional physical coordinates of the current signal acquisition point on the metal surface. The horizontal position of this physical coordinate is divided by the spatial step size of 2 millimeters to obtain the row index number, and the vertical position is divided by the spatial step size of 2... The column index is obtained by measuring millimeters. For example, when the physical coordinates are 6 millimeters horizontally and 8 millimeters vertically, the row index is calculated to be 3 and the column index to be 4. The calculated attenuation correction result value of 1.155 is precisely mapped to the storage unit in the 3rd row and 4th column of the preset blank multidimensional array, overwriting the original 0 value. This logic is repeated to traverse the correction result values ​​corresponding to the detection sampling points and complete the row and column dimension mapping. After the mapping operation of all sampling points is completed, the original blank multidimensional array is updated to an acoustic propagation parameter matrix filled with specific acoustic parameters. Based on this calculation, the overall processing service platform can accurately locate abnormal micro-sites inside the material according to the spatial coordinate mapping distribution of this matrix, and adaptively adjust the processing service feed rate and cutting threshold in subsequent construction operations according to the distribution state.

[0026] Specifically, such as Figure 2 , 4 As shown, the grain sensing module includes: The boundary segmentation submodule acquires images of the microstructure of precious metals, converts the images into a single-channel grayscale matrix, extracts the pixel step position separation region within the single-channel grayscale matrix, and generates a set of grain boundary pixels. The boundary segmentation submodule connects to a high-magnification optical microscopy imaging device to acquire microscopic images of the precious metal surface after polishing and etching. The microscopic image is a high-definition color digital image containing three channels: red, green, and blue. Following the International Commission on Illumination (ICI) standard formula, the boundary segmentation submodule multiplies the red channel component by 0.299, the green channel component by 0.587, and the blue channel component by 0.114. These products are then added together to obtain a single grayscale value. This weighted summation transformation logic is executed pixel-by-pixel to convert the microscopic image into a single-channel grayscale matrix. For each central pixel in the single-channel grayscale matrix, the grayscale value of its right-hand neighboring pixel is extracted. The grayscale value of the central pixel is obtained by subtracting the grayscale value of the right-hand neighboring pixel from the grayscale value of the central pixel. Similarly, the grayscale value of the central pixel is obtained by subtracting the grayscale value of the grayscale value of the lower neighboring pixel from the grayscale value of the central pixel. The sum of the absolute values ​​of the horizontal and vertical grayscale value differences is calculated as the final grayscale value difference. This process extracts the pixel step position separation region within the single-channel grayscale matrix. The maximum and minimum gray values ​​in the single-channel grayscale matrix are read. The maximum gray value is subtracted from the minimum gray value to obtain the grayscale dynamic range. Pixels with grayscale value differences exceeding a preset threshold are marked as boundary candidate points. The preset threshold ranges from 10% to 20% of the grayscale dynamic range of the single-channel grayscale matrix. For the image with a dynamic range of 200, 15% is extracted as the preset threshold, i.e., 30. This threshold ratio is obtained by comparing and verifying 50 microscopic images under different lighting conditions. Below 10% will introduce too many artifact noise, while above 20% will lead to the loss of fine grain boundaries. When the grayscale value difference of a pixel is 45, it is determined to be greater than 30 and its coordinates are recorded as a boundary candidate point. The accumulated boundary candidate points generate a grain boundary pixel set.

[0027] Table 1: Local Validation Data Table for Boundary Detection of Precious Metal Microscopic Images; As shown in Table 1, the boundary segmentation submodule completes the classification and determination of the attributes of adjacent pixels by comparing the difference in gray values ​​obtained by comparison with the preset threshold calculated based on the dynamic range.

[0028] The morphological parameter extraction submodule calls the grain boundary pixel set, accumulates the edge pixel count to obtain the grain outline perimeter, counts the internal pixels to obtain the cross-sectional area, divides the grain outline perimeter by the cross-sectional area to obtain the outline ratio value, measures the intersection angle of the centroid connection line of adjacent grains to obtain the centroid connection line angle, and generates morphological space parameter values. The morphological parameter extraction submodule calls the grain boundary pixel set generated by the preprocess, and uses the connected component labeling algorithm to identify each closed region formed by the boundary pixels as an independent grain. For each independent grain, the total number of pixels in the grain boundary pixel set that constitutes its outer contour is counted, and the total number is directly accumulated to obtain the grain contour perimeter of the independent grain. At the same time, the morphological parameter extraction submodule traverses the non-boundary pixels inside the closed region, and directly obtains the cumulative number of internal pixels as the cross-sectional area. The division operation logic is executed to divide the grain contour perimeter by the cross-sectional area, and the quotient of the two is used as the contour ratio value. For example, if the perimeter of a certain independent grain is 200 pixels and the cross-sectional area is 2000 pixels, dividing 200 by 2000 yields a contour ratio of 0.1. Then, the average values ​​of the horizontal and vertical coordinates of the pixels inside the independent grain are calculated. This set of average values ​​is combined as the centroid coordinates of the grain. In the spatial coordinate system, the centroids of this grain and another adjacent grain are selected. These two centroid coordinates are extracted and connected by a straight line to form a centroid connection line. The morphological parameter extraction submodule obtains the intersection angle between the connection line and the horizontal reference axis by calculating the arctangent of the slope of the centroid connection line. This angle is taken as the centroid connection line angle. Combining the contour ratio of 0.1 and the centroid connection line angle of 45 degrees, the two are organized into a two-dimensional data set to generate the morphological spatial parameter values ​​for this grain.

[0029] The structural feature statistics submodule, based on the morphological space parameter values, accumulates the number of independent grains within the field of view to obtain the number of grains in the region, and writes the number of grains in the region and the morphological space parameter values ​​into the feature record matrix according to the grain index number to generate grain structure feature data. The structural feature statistics submodule receives the morphological space parameter values ​​of independent grains and starts a counter within the global image field of view. It traverses the marked closed connected regions one by one. Each time a valid and complete independent grain is encountered, the counter value is incremented by 1. After the complete field of view is traversed, the counting stops, and the final value accumulated by the counter is directly obtained as the number of grains in the region. For example, if the counter finally stops at 150, then the number of grains in the region is 150. Based on the traversal order, each of the 150 independent grains is assigned a unique consecutive integer from 1 to 150 as a grain index number. A multi-dimensional feature recording matrix is ​​allocated in memory. The number of rows in this feature recording matrix is ​​equal to the number of 150 grains in the region, and the number of columns corresponds to the various parameter dimensions. The number of 150 grains in the region and the morphological space parameter values ​​of each grain are written item by item according to the grain index number. Specifically, for the grain with index number 1, its contour ratio value of 0.1 and the angle between the centroid line and the grain are filled into a specific column of the first row of the feature recording matrix. The total number of regions 150 is recorded in the basic information column of the first row. This writing action is repeated until the information of the 150th grain is completely recorded. After all the grain information is structured and encapsulated, grain structure feature data is generated.

[0030] Specifically, such as Figure 2 , 5 As shown, the energy modeling module includes: The node distance calculation submodule extracts the centroid coordinates of the grains in the grain structure feature data as spatially distributed nodes, and performs linear span association mapping based on the spatial absolute position features of adjacent spatially distributed nodes in the preset coordinate system to generate node Euclidean distance vectors. The node distance calculation submodule reads the grain structure feature data output from the previous level, and extracts the centroid coordinates of each grain stored in the feature record matrix. Since the centroid coordinates represent the real center of gravity of the grain on the physical slice of the material cross section, each extracted grain centroid coordinate is instantiated as a spatially distributed node, and its absolute spatial position features are recorded in a preset two-dimensional coordinate system composed of the horizontal and vertical axes. A double-loop pairing mechanism is initiated to select any two spatially distributed nodes. Based on the absolute spatial position characteristics of adjacent spatially distributed nodes in a preset coordinate system, a straight-line span association mapping is performed. The mapping calculation logic is as follows: extract the x-coordinate value of the first spatially distributed node and subtract the x-coordinate value of the second spatially distributed node, calculate the square of the difference, extract the y-coordinate value of the first spatially distributed node and subtract the y-coordinate value of the second spatially distributed node, calculate the square of the difference, add the two squared results and take their positive square root to calculate the Euclidean straight-line span between the two points. For example, if the x-coordinate of the first node is 10 and the y-coordinate is 20, and the x-coordinate of the second node is 13 and the y-coordinate is 24, the square of the difference in x-coordinates is 9, the square of the difference in y-coordinates is 16, and the sum is 25. The positive square root is used to calculate the straight-line span as 5. The straight-line spans calculated between the paired nodes are arranged and combined into a one-dimensional array according to the node index pairs to generate the node Euclidean distance vector.

[0031] The edge weight extraction submodule, based on the node Euclidean distance vector and the acoustic propagation parameter matrix, reads the stagnation value and the decrease ratio at the spatially distributed nodes in the acoustic propagation parameter matrix, and appends the product of the stagnation value and the decrease ratio to the corresponding position of the node Euclidean distance vector to generate the propagation edge weight matrix. The edge weight extraction submodule extracts a pair of spatially distributed node indices from the previously generated node Euclidean distance vector and the previously constructed acoustic propagation parameter matrix. It then uses the physical coordinates corresponding to these spatially distributed nodes to reverse map and query the acoustic propagation parameter matrix. The module reads the stagnation value and decay rate recorded at the corresponding physical coordinate position in the acoustic propagation parameter matrix. The stagnation value is quantized by the product of the local material density and the ultrasonic velocity, and the decay rate is simplified by the acoustic attenuation coefficient. The edge weight extraction submodule executes the multiplication operation logic, directly multiplying the stagnation value obtained from the lookup table with the decay rate to obtain the local acoustic conduction impedance. For example, if the impedance value is found to be 2.5 and the decreasing rate is 0.4, multiplying 2.5 by 0.4 yields a local acoustic conduction impedance of 1.0. Then, this product is appended to the position of the distance element between the two nodes in the node Euclidean distance vector. Using multiplication correction logic, the original node Euclidean distance value is multiplied by the local acoustic conduction impedance. The previously calculated straight-line span of 5 is multiplied by 1.0 to obtain a corrected edge weight value of 5.0. The corrected edge weight values ​​calculated between node pairs are integrated into a two-dimensional square matrix, where the row index and column index correspond to the starting node and the ending node, respectively, to generate the propagation edge weight matrix.

[0032] Table 2: Record of Calculation of Inter-Node Transmission Characteristic Parameters; As shown in Table 2, the edge weight extraction submodule uses the extracted acoustic parameters to perform multiplier correction on the basic geometric span, and obtains the edge weight values ​​that comprehensively reflect the physical conduction properties.

[0033] The network topology construction submodule calls the propagation edge weight matrix, extracts the non-zero elements in the propagation edge weight matrix as interconnected edges, and connects the spatially distributed nodes in sequence according to the interconnected edges to establish an energy propagation network model. The network topology construction submodule calls the previously established propagation edge weight matrix and initiates a matrix element scanning mechanism. It reads each value in the propagation edge weight matrix row by row and column by column, extracting non-zero elements as interconnecting edges. The logic for extracting these interconnecting edges iterates through all elements of the propagation edge weight matrix, checking if the element value is greater than 0. The row and column index pairs with non-zero values ​​are recorded as endpoint index pairs for the interconnecting edges. For example, when the value in row 2, column 3 is 12.0, since it is greater than 0, the index pairs 2 and 3 are recorded as endpoint index pairs. The endpoint indices are arranged sequentially according to row priority in the propagation edge weight matrix, forming a rule-ordered sequence of interconnected edge indices. A directed connection is established between the two spatially distributed nodes corresponding to each endpoint index pair in this sequence, i.e., from the starting node represented by the row index to the ending node represented by the column index. The set of directed connections generated by the endpoint index pairs constitutes the edge set of the energy propagation network model. Based on these interconnected edges, spatially distributed nodes are connected sequentially. Combining the Euclidean position information of the nodes, a complete energy propagation network model is finally established. This model is essentially a weighted directed graph data structure. The advantage of this operational logic is that it significantly compresses the storage complexity of the graph structure by eliminating invalid connected paths with zero weight and filters out absolute sound barrier regions that sound waves cannot penetrate.

[0034] Specifically, such as Figure 2 , 6 As shown, the impedance fusion module includes: The node extraction submodule extracts the energy allocation values ​​of the nodes in the energy propagation network model, and performs mapping on the energy allocation values ​​of the nodes in the energy propagation network model according to the topological space sequence, integrating the node energy values ​​into an array range to generate an energy sequence vector; The node extraction submodule reads the initial ultrasonic energy intensity of each spatially distributed node in the established energy propagation network model as the node energy allocation value. This energy allocation value is calculated by the external ultrasonic field injection power after conduction attenuation through the first layer node. The topology space sequence is retrieved from the local storage unit. This topology space sequence is the node traversal order number generated by the breadth-first search algorithm of the network model graph. The energy allocation values ​​of the nodes in the energy propagation network model are sorted and mapped according to the topology space sequence. The specific logic involves extracting the energy values ​​scattered across each node in a breadth-first search order from 1 to N. For example, if the breadth-first search sequence prioritizes node 3 over node 1, with node 3 having an energy allocation of 50 microjoules and node 1 having an energy allocation of 45 microjoules, then the 50 microjoules are extracted first, followed by the 45 microjoules. The extracted energy values ​​from each node are then sequentially integrated into a continuous one-dimensional array interval, forming a set of values ​​arranged according to the decreasing depth of the network hierarchy. This generates an energy sequence vector, which directly reflects the layer-by-layer dissipation characteristics of ultrasonic energy during transmission between layers of a dense noble metal lattice network. This provides a basic quantitative basis for evaluating the internal acoustic uniformity of materials at different processing thicknesses and densities in noble metal processing services.

[0035] The gradient calculation submodule calls the energy sequence vector, extracts the parameter element pairs of adjacent topological connection states in the energy sequence vector, performs subtraction operation on the parameter element pairs to obtain the node energy range index, and integrates the node energy range index to perform alignment and arrangement to generate the node impedance gradient matrix. The gradient calculation submodule calls the previously processed energy sequence vector, extracts parameter element pairs from adjacent topological connections in the energy sequence vector, that is, the energy values ​​corresponding to parent and child nodes with direct head-to-tail connections in the topological graph. It performs a subtraction operation on these parameter element pairs, subtracting the energy value of the child node from the energy value of the parent node to obtain the node energy range index. This range index reflects the energy loss gradient within a single propagation span. For example, if the parent node's energy value is 50 microjoules and the corresponding directly connected child node's energy value is 45 microjoules, the gradient calculation submodule subtracts 45 from 50 to calculate the node energy range index as 5 microjoules. It then traverses the directly connected edges in the network model and calculates the energy range index of each node at both ends. The range calculation results are collected, and the node energy range indexes are integrated and aligned using alignment logic. This alignment logic sorts the calculated range indices according to the starting node index recorded in the edge set. The sorted range indices are then filled into a single-column matrix structure with the same number of rows as the total number of connected edges, generating the node impedance gradient matrix. When precious metals such as gold, silver, and platinum undergo high-intensity processing such as forging and wire drawing, microscopic tearing or grain boundary dislocation accumulation is easily generated inside them. Structural defects will significantly increase the local acoustic wave propagation resistance. Therefore, this nodal impedance gradient matrix can accurately capture and locate the abrupt changes in microscopic mechanical properties caused by mechanical stress concentration or fatigue damage inside precious metal materials.

[0036] The feature fusion submodule, based on the node impedance gradient matrix, calls the corresponding weight parameters of the energy propagation network model, reads the numerical relationship between the weight parameters and the elements of the node impedance gradient matrix, multiplies and accumulates the values ​​of the node impedance gradient matrix elements with the corresponding weight parameters, and generates a multidimensional feature vector. The feature fusion submodule, based on the previously generated node impedance gradient matrix, calls the corresponding weight parameters of the pre-trained deep neural network that matches the energy propagation network model structure. This deep neural network contains one input layer, three hidden layers using linear rectified activation functions, and one output layer. Currently, it calls the connection weight parameter matrix between the first hidden layer and the input layer. It reads the corresponding numerical relationship between the elements of the weight parameter matrix and the node impedance gradient matrix, and performs dot multiplication and addition logic, multiplying and accumulating the values ​​of the node impedance gradient matrix elements with the corresponding network connection weight parameters. Specifically, the first element of the impedance gradient matrix is ​​extracted as 5, corresponding to a network weight parameter of 0.8; the second element is extracted as 3, corresponding to a network weight parameter of 0.5; multiplying 5 by 0.8 yields 4.0; multiplying 3 by 0.5 yields 1.5; then accumulating 4.0 and 1.5 gives the pre-activation input value of a single neuron as 5.5. After performing a complete matrix multiplication operation by traversing the entire gradient matrix and weight parameters, the accumulated output result is input to the deep neural network structure processing component to generate a multidimensional feature vector in high-dimensional space. This high-dimensional spatial feature fully maps the complex perturbation law of the nonlinear attenuation of sound waves caused by the high density crystal and high ductility of precious metals. This enables the intelligent control center to keenly retrieve the historical data patterns of processing services and make forward-looking process intervention and adjustment based on the structural fusion features of precious metals extracted by deep neural networks in precious metal processing operations facing random mutations and extremely challenging working conditions.

[0037] Specifically, such as Figure 2 , 7 As shown, the anomaly detection module includes: The feature aggregation submodule extracts multidimensional feature vectors, collects a preset topological space connection matrix, performs multiplication calculations between the elements in the multidimensional feature vectors and the elements in the topological space connection matrix, extracts the calculated parameters of adjacent nodes, performs accumulation, and generates a node feature aggregation matrix. The feature aggregation submodule extracts the multidimensional feature vectors generated in the preceding process, which contain output information from multiple hidden layers. It then acquires a pre-set topology space connection matrix via the memory bus. This topology space connection matrix is ​​the normalized graph adjacency matrix, where a value of 1 indicates a connection between nodes, and 0 indicates no connection. Utilizing the core computational mechanism of the graph convolutional neural network, it performs standard matrix multiplication with the elements in the topology space connection matrix, multiplying the high-dimensional element matrix within the multidimensional feature vectors. Through this matrix multiplication, each node automatically obtains the high-dimensional feature information of its direct neighbors. Subsequently, it extracts the calculated neighboring node feature parameters and performs aggregation on the node's own dimension. For example, if node 1's feature vector value is 2.0, its neighbor node 2's feature vector value is 3.0, and its neighbor node 3's feature vector value is 4.0, after multiplication and aggregation using the connection matrix, node 1's aggregated feature value is updated to the sum of 2.0, 3.0, and 4.0, resulting in 9.0. This neighborhood information aggregation and update operation is performed on all network nodes, generating a node feature aggregation matrix that fully integrates local topology and neighbor context information.

[0038] The spatial mapping submodule detects the activation function mapping parameters based on the node feature aggregation matrix, performs mapping on the node feature aggregation matrix using the activation function mapping parameters, calculates the deviation of the mapping amount from the preset state baseline value, delineates the classification boundary based on the deviation amount, and generates state classification label values. The spatial mapping submodule, based on the updated node feature aggregation matrix, detects and retrieves the S-type activation function mapping parameters from the configuration parameter library. It then uses this activation function calculation mechanism to perform a non-linear mapping on each element value in the node feature aggregation matrix, compressing the original aggregated feature values ​​to an output range of 0 to 1 to obtain the mapping value. For example, if the input feature aggregation value is 9.0, after processing by the S-type activation function calculation logic, the output mapping value is 0.999. Subsequently, the system's preset state baseline value is read as 0.5. The obtained mapping value is subtracted from the preset state baseline value, and the absolute value is taken to calculate the deviation of the mapping value relative to the preset state baseline value. For example, 0.999 minus 0.5 yields an absolute deviation of 0.499. The spatial mapping submodule uses this deviation to define classification boundaries. This definition logic refers to classifying low-risk deviation levels, medium-risk deviation levels, and high-risk deviation levels as... The deviation level is mapped to three preset classification boundary segments. The classification boundary segments are divided with a fixed numerical step size of 0.2. Specifically, a deviation of 0 to 0.2 belongs to the low-risk segment, 0.2 to 0.4 belongs to the medium-risk segment, and greater than 0.4 belongs to the high-risk segment. Based on the classification boundary segment to which the current deviation of 0.499 belongs, it is determined that it falls into the interval greater than 0.4. Thus, a state classification label value of 3 is generated, which corresponds one-to-one with the high-risk deviation level, and the final quantitative state classification label value is generated.

[0039] Table 3: Comparison and Judgment Table of Node Deviation Amount and Risk Classification; As shown in Table 3, the spatial mapping submodule outputs discretized classification labels clearly using a fixed step size rule based on the difference between the mapping amount after the node features are processed by the activation function and the baseline value.

[0040] The status comparison submodule calls the status classification label value, obtains the physical operating status quantity of the equipment, compares the numerical difference between the status classification label value and the physical operating status quantity of the equipment, determines the risk category when the numerical difference exceeds the safety tolerance limit, and generates the abnormal detection result of the precious metal processing service. The state comparison submodule calls the state classification label value issued by the space mapping submodule through an internal communication protocol. For example, the previously generated state classification label value is 3. Simultaneously, the state comparison submodule reads the physical operating status quantity of the equipment from the programmable logic controller of the precious metal processing production line. This physical operating status quantity is an integer value collected by the vibration sensor of the processing table and normalized and quantized. For example, if the currently read physical operating status quantity is 1, the state comparison submodule uses a difference operation logic to compare the numerical difference between the state classification label value and the physical operating status quantity. Subtracting the physical operating status quantity 1 from the state classification label value 3 yields a numerical difference of 2. The submodule then reads the safety tolerance extreme limit set to 1 from the safety threshold setter and determines that the currently calculated numerical difference 2 is greater than... The safety tolerance limit is 1. Therefore, the system determines that the difference in value has exceeded the safety tolerance limit. The system triggers a conditional judgment command to determine that the detected target falls into the category of process instability risk caused by precious metal material deterioration. Based on the above logical comparison results, a diagnostic report containing the risk category code and specific difference value is output, generating an anomaly detection result for the current precious metal processing service. The advantage of this operational logic is that by cross-validating the logical labels derived from microscopic features with the physical labels fed back by macroscopic equipment, it greatly reduces the downtime accident rate caused by unilateral data false alarms. This significantly reduces the processing schedule delay costs caused by blind shutdowns while comprehensively ensuring the safety of the entire precious metal processing service system and the delivery quality of the final product under high-frequency orders.

[0041] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. An AI-based anomaly detection system for precious metal processing, characterized in that, The system includes: The acoustic acquisition module acquires ultrasonic transmission and reflection signals from precious metal processing services and converts them into digital signals. It performs discrete Fourier transform to extract frequency domain amplitude components, calculates the frequency domain amplitude ratio and signal delay span, and constructs an acoustic propagation parameter matrix. The grain sensing module acquires images of the microstructure of noble metals, performs grayscale processing and segments grain boundaries, calculates the ratio of grain outline perimeter to area, measures the angle between the centroids of adjacent grains, and generates grain structure feature data. The energy modeling module, based on the grain structure feature data and the acoustic propagation parameter matrix, sets the grain centroid as a node, calculates the Euclidean distance between adjacent nodes as the side length, extracts the propagation drag coefficient and attenuation rate as the side weights, and constructs an energy propagation network model. The impedance fusion module extracts the energy allocation values ​​of the nodes in the energy propagation network model, uses the difference in energy allocation values ​​of adjacent nodes as the impedance gradient, and performs a weighted summation with the weight parameters to generate a multi-dimensional feature vector. The anomaly detection module extracts the multidimensional feature vector, inputs it into a graph convolutional neural network to perform feature aggregation and nonlinear spatial mapping to obtain classification labels, and outputs anomaly detection results by comparing them with a preset physical state benchmark.

2. The AI-based anomaly detection system for precious metal processing according to claim 1, characterized in that, The acoustic propagation parameter matrix includes acoustic attenuation coefficient, acoustic phase velocity, and ultrasonic group velocity; the grain structure feature data includes average grain size, grain boundary density, and shape factor; the energy propagation network model includes network adjacency matrix, node degree matrix, and graph Laplacian matrix; the multidimensional feature vector includes acoustic impedance, energy spectral density, and feature root mean square value; and the anomaly detection results include defect spatial location, damage severity, and remaining workpiece life.

3. The AI-based anomaly detection system for precious metal processing according to claim 1, characterized in that, The acoustic acquisition module includes: The signal conversion submodule acquires the ultrasonic transmission signal and reflected echo signal of the precious metal processing service, converts the ultrasonic transmission signal and reflected echo signal of the precious metal processing service into a discrete digital sequence, performs a Fourier transform on the discrete digital sequence to extract the amplitude value, and generates the frequency domain amplitude component. The delay extraction submodule, based on the frequency domain amplitude component, divides the frequency domain amplitude component by the frequency domain reference value to obtain the amplitude division ratio, reads the difference between the transmission and return times and assigns it to the signal delay span, and generates a span parameter vector by combining the amplitude division ratio and the signal delay span; The parameter construction submodule calls the span parameter vector, calculates the product of the span parameter vector and the preset attenuation compensation coefficient value, maps the product to a preset blank multidimensional array according to the row and column dimensions, and establishes the acoustic propagation parameter matrix.

4. The AI-based anomaly detection system for precious metal processing according to claim 1, characterized in that, The grain sensing module includes: The boundary segmentation submodule acquires images of the microstructure of precious metals, converts the images into a single-channel grayscale matrix, extracts the pixel step position separation region within the single-channel grayscale matrix, and generates a set of grain boundary pixels. The morphological parameter extraction submodule calls the grain boundary pixel set, accumulates the edge pixel count to obtain the grain outline perimeter, counts the internal pixel points to obtain the cross-sectional area, divides the grain outline perimeter by the cross-sectional area to obtain the outline ratio value, measures the intersection angle of the centroid connection line of adjacent grains to obtain the centroid connection line angle, and generates morphological space parameter values. The structural feature statistics submodule, based on the morphological space parameter value, accumulates the number of independent grains within the field of view to obtain the number of grains in the region, and writes the number of grains in the region and the morphological space parameter value into the feature record matrix according to the grain index number to generate grain structure feature data.

5. The AI-based anomaly detection system for precious metal processing according to claim 4, characterized in that, The extraction of pixel step position separation region in single-channel gray-scale matrix refers to calculating the difference in gray-scale values ​​of adjacent pixels in single-channel gray-scale matrix, and marking the position of pixel where the difference in gray-scale value exceeds a preset threshold as a boundary candidate point. The preset threshold ranges from 10% to 20% of the gray-scale dynamic range of single-channel gray-scale matrix.

6. The AI-based anomaly detection system for precious metal processing according to claim 1, characterized in that, The energy modeling module includes: The node distance calculation submodule extracts the grain centroid coordinates in the grain structure feature data as spatial distribution nodes, performs linear span association mapping based on the spatial absolute position features of adjacent spatial distribution nodes in a preset coordinate system, and generates node Euclidean distance vectors. The edge weight extraction submodule, based on the node Euclidean distance vector and the acoustic propagation parameter matrix, reads the stagnation value and the decrease ratio at the spatially distributed nodes in the acoustic propagation parameter matrix, and appends the product of the stagnation value and the decrease ratio to the corresponding position of the node Euclidean distance vector to generate the propagation edge weight matrix. The network topology construction submodule calls the propagation edge weight matrix, extracts the non-zero elements in the propagation edge weight matrix as interconnected edges, and connects the spatially distributed nodes sequentially according to the interconnected edges to establish an energy propagation network model.

7. The AI-based anomaly detection system for precious metal processing according to claim 6, characterized in that, The step of extracting the non-zero element positions in the propagation edge weight matrix as interconnected edges involves traversing all elements of the propagation edge weight matrix, recording the row and column index pairs with non-zero element values ​​as endpoint index pairs of interconnected edges, and arranging the endpoint index pairs in the row priority order in the propagation edge weight matrix to form an interconnected edge index sequence. A directed connection relationship is established between the two spatially distributed nodes corresponding to each endpoint index pair in the interconnected edge index sequence. The set of all directed connections constitutes the edge set of the energy propagation network model.

8. The AI-based anomaly detection system for precious metal processing according to claim 1, characterized in that, The impedance fusion module includes: The node extraction submodule extracts the energy allocation values ​​of the nodes in the energy propagation network model, and performs mapping on the energy allocation values ​​of the nodes in the energy propagation network model according to the topological space sequence, integrating the node energy values ​​into an array range to generate an energy sequence vector; The gradient calculation submodule calls the energy sequence vector, extracts the parameter element pairs of adjacent topological connection states in the energy sequence vector, performs subtraction operation on the parameter element pairs to obtain the node energy range index, and integrates the node energy range index to perform alignment and arrangement to generate the node impedance gradient matrix. The feature fusion submodule, based on the node impedance gradient matrix, calls the corresponding weight parameters of the energy propagation network model, reads the numerical relationship between the weight parameters and the elements of the node impedance gradient matrix, multiplies and accumulates the values ​​of the node impedance gradient matrix elements with the corresponding weight parameters, and generates a multidimensional feature vector.

9. The AI-based anomaly detection system for precious metal processing according to claim 1, characterized in that, The anomaly detection module includes: The feature aggregation submodule extracts the multidimensional feature vector, collects a preset topological space connection matrix, performs multiplication calculations between the elements in the multidimensional feature vector and the elements in the topological space connection matrix, extracts the calculated adjacent node parameters, performs accumulation, and generates a node feature aggregation matrix. The spatial mapping submodule detects the activation function mapping parameters based on the node feature aggregation matrix, performs mapping on the node feature aggregation matrix using the activation function mapping parameters, calculates the deviation of the mapping amount from the preset state baseline value, delineates the classification boundary based on the deviation amount, and generates a state classification label value. The status comparison submodule calls the status classification label value to obtain the physical operating status quantity of the equipment, compares the numerical difference between the status classification label value and the physical operating status quantity of the equipment, determines the risk category when the numerical difference exceeds the safety tolerance limit, and generates the abnormal detection result of the precious metal processing service.

10. The AI-based anomaly detection system for precious metal processing according to claim 9, characterized in that, The method of defining classification boundaries based on deviation magnitude refers to mapping low-risk, medium-risk, and high-risk deviation levels to preset classification boundary segments. The classification boundary segments are divided with a fixed numerical step size. Based on the classification boundary segment to which the mapped value belongs, a status classification label value corresponding one-to-one with the deviation magnitude level is generated.

Images