Chip inductor defect detection method and system based on machine vision

Abstract

The invention provides a chip inductor defect detection method and system based on machine vision, relates to the technical field of industrial vision intelligence, and constructs a set of brand new defect detection and diagnosis framework by synchronously characterizing the internal heat conduction characteristic and the surface stress state of a chip inductor and deeply mining the physical causal association between the internal heat conduction characteristic and the surface stress state. The core value of the framework is that through accurate risk quantification, the omission ratio of high-risk defects and the misjudgment rate of good products with normal functions can be effectively reduced; and on the other hand, a complete and data-driven intelligent manufacturing closed-loop control system is constructed for realizing real-time sorting of a single component and periodic feedback optimization of process parameters of the whole production line by tracing the cause of the defect, so that the production quality and reliability of high-end electronic components are fundamentally improved.

Description

Technical Field

[0001] This invention relates to the field of industrial vision and intelligence technology, specifically to a method and system for detecting defects in patch inductors based on machine vision. Background Technology

[0002] In the high-end electronics manufacturing sector, especially in industries such as automotive electronics, aerospace, and medical equipment where component reliability requirements reach a "zero-defect" standard, automated and intelligent quality inspection technologies are crucial for ensuring product safety and performance. As basic electronic components such as surface mount inductors evolve towards higher frequencies, higher integration, and miniaturization, their internal structures become increasingly complex. The ability to detect and diagnose various microscopic defects that may be introduced during the production process has become one of the core indicators for measuring advanced manufacturing levels. Therefore, developing intelligent quality inspection technologies capable of deeply, accurately, and efficiently identifying potential risks is the current technological development trend across the entire industry.

[0003] Currently, while technological solutions for detecting defects in surface mount inductors have made progress in automation, they still face the following core challenges in terms of diagnostic depth and accuracy: Limitations of Diagnostic Dimensions: Most mainstream machine vision inspection methods rely on optical imaging of component surfaces for information. While these methods can effectively identify surface defects that are large or have significant morphological contrast, they have inherent limitations in determining the physical causes of these defects. The system struggles to effectively distinguish between two anomalies that are morphologically similar but have vastly different risk levels: one is harmless surface processing marks, and the other is high-risk microcracks transmitted to the surface from internal structural damage (such as delamination or voids). This detection mode leads to uncertainty in diagnostic results.

[0004] The superficiality of information processing: Even with advanced algorithms, such as those disclosed in patent application CN117152139A, which utilizes an instance segmentation network model to process patch inductor images, the essence remains a deep dive into information from a single physical dimension (optical image). While such methods improve the segmentation accuracy and classification efficiency of "visible" defects, their diagnostic capabilities are limited by the constraints of the information source. When the root cause of a defect lies deep within the component and does not exhibit a typical morphology on the surface, these techniques risk missing detections.

[0005] Singleness of decision output: the output result of the prior art solution is usually a simple "pass" or "fail" binary judgment, or a rough classification based on morphology. The information entropy of such output signal is low, which is difficult to support more refined downstream quality control strategies. For example, it cannot provide quantitative basis for multi-level risk sorting (including "good product", "to be observed product", "waste product"), nor can it causally trace the defect causes, so as to provide effective, data-driven guidance for the optimization and adjustment of upstream production processes. SUMMARY

[0006] The purpose of the present application is to provide a patch inductance defect detection method and system based on machine vision, to solve the problems raised in the background art.

[0007] To achieve the above purpose, the present application provides the following technical solutions: A patch inductance defect detection method based on machine vision, the specific steps comprising: S1: obtaining first data representing the internal heat conduction characteristics of the patch inductance; S2: obtaining second data representing the surface microstructure and stress state of the patch inductance; S3: performing correlation processing on the spatially registered first data and second data to identify the spatial causal correspondence between the internal thermal anomaly area indicated by the first data and the surface stress anomaly area indicated by the second data, and generating a defect physical cause diagnosis signal accordingly; S4: generating a structured indication signal based on the defect physical cause diagnosis signal, the structured indication signal including at least a defect classification label and a quantitative risk index; S5: based on the structured indication signal, performing at least one of the following control operations: For a single patch inductance, based on its corresponding defect classification label and quantitative risk index, performing multi-level sorting operation; Statistical analysis is performed on the structured indication signal, and when it is determined that the frequency of occurrence of any category in the defect classification label meets the preset condition, a process parameter adjustment instruction is generated and sent to the upstream production equipment of the patch inductance.

[0008] A patch inductance defect detection system based on machine vision, the system is used to execute the patch inductance defect detection method based on machine vision, comprising: A multi-physical field data acquisition unit is used to obtain first data representing the internal heat conduction characteristics of the patch inductance; Obtaining second data representing the surface microstructure and stress state of the patch inductance; A spatial causality correlation diagnosis unit is configured to perform correlation processing on the first data and the second data that are spatially registered, to identify a spatial causality correspondence between an internal thermal anomaly region indicated by the first data and a surface stress anomaly region indicated by the second data, and to generate a defect physical cause diagnosis signal based on the spatial causality correspondence; A structured risk signal generation unit is configured to generate a structured indication signal based on the defect physical cause diagnosis signal, the structured indication signal including at least a defect classification label and a quantified risk index; A closed-loop linkage control execution unit is configured to perform at least one of the following control operations based on the structured indication signal: Performing a multi-level sorting operation on a single patch inductor based on the defect classification label and the quantified risk index corresponding to the patch inductor; Performing statistical analysis on the structured indication signal, and generating a process parameter adjustment instruction and sending the process parameter adjustment instruction to an upstream production device of the patch inductor when it is determined that the frequency of occurrence of any one of the defect classification labels meets a preset condition.

[0009] Compared with the prior art, the present application has the beneficial effects that: two key and mutually orthogonal physical characterization dimensions are introduced: one is the internal thermal conduction characteristic (first data) that can reveal the internal structural integrity and material uniformity of the component; the other is the surface microstructure and stress state (second data) that can finely reflect the surface microtopography and stress distribution. By synchronously acquiring these two kinds of "symptom" data of different sources but related in physics, a solid data foundation is laid for causal inference.

[0010] Through a specific correlation processing model, the spatial causality correspondence between the internal thermal anomaly region and the surface stress anomaly region is actively identified and quantified. This analysis leap from "correlation" to "causality" enables the present application to change the nature of detection from morphological recognition to physical cause diagnosis.

[0011] Based on the deep understanding of the causes of defects, the present application can generate a structured indication signal with highly concentrated information. The signal contains two core elements: a qualitative "defect classification label" (including "internal structural defect" and "surface processing defect") and a quantitative "quantified risk index". This structured output provides decision accuracy and flexibility for subsequent intelligent control, realizing a complete information link from detection to diagnosis and then to decision. BRIEF DESCRIPTION OF DRAWINGS

[0012] Fig. 1 The figure is a schematic diagram of the overall system module flow of the present application; Fig. 2 The figure is a schematic diagram of the execution logic of steps S1 to S3 of the present application; Fig. 3 The execution logic diagram for steps S4 and S5 of the present application. DETAILED DESCRIPTION

[0013] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0014] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the concept of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0015] Embodiment one: Please refer to Figs. 1-2 The present application provides a technical solution: A patch inductance defect detection method based on machine vision, comprising the following steps: S1: obtaining first data representing the internal heat conduction characteristics of the patch inductance; S2: obtaining second data representing the surface microstructure and stress state of the patch inductance; S3: performing correlation processing on the spatially registered first data and second data to identify the spatial causal correspondence between the internal thermal anomaly area indicated by the first data and the surface stress anomaly area indicated by the second data, and generate a defect physical cause diagnosis signal accordingly; S4: generating a structured indication signal based on the defect physical cause diagnosis signal, the structured indication signal at least including a defect classification label and a quantitative risk index; S5: based on the structured indication signal, performing at least one of the following control operations: For a single patch inductance, based on its corresponding defect classification label and quantitative risk index, performing a multi-level sorting operation; Statistical analysis is performed on the structured indication signal, and when it is determined that the frequency of occurrence of any category in the defect classification label meets a preset condition, a process parameter adjustment instruction is generated and sent to the upstream production equipment of the patch inductance.

[0016] Further explanation: in this embodiment, step S3, that is, through a specific correlation processing model, establishes the physical causal correspondence between internal heat conduction anomaly (first data) and surface stress / microstructure anomaly (second data), thereby generating a diagnosis signal that can reveal the "cause" of the defect; the essence of step S3 is to "identify the spatial causal correspondence".

[0017] The embodiment introduces a hierarchical and dynamically weighted causal inference paradigm. Through the candidate association pair generation stage, the data is preliminarily screened and paired; secondly, in the core causal confidence evaluation stage, a key internal defect severity degree parameter is introduced, and a nonlinear causal association weight is dynamically generated based on the parameter, which is used to modulate the information propagation process in the graph neural network; so that the inference engine can focus the computing resources and model attention on the "high-risk" associations caused by "serious" internal defects, thereby improving the accuracy and robustness of diagnosis.

[0018] Further, S1 comprises: applying a transient thermal excitation to the patch inductor; and, Collecting and processing the temperature decay process of the surface of the patch inductor after the transient thermal excitation to generate the first data containing at least a thermal relaxation time constant map.

[0019] Further, S2 comprises: Irradiating the patch inductor with light of a preset polarization state; and, Collecting and processing the scattered light on the surface of the patch inductor to generate the second data containing at least a linear polarization degree map and a linear polarization angle map.

[0020] Further, in S3, the association processing is realized by a preset defect physical cause inference engine, which is used to model the abnormal features extracted from the first data and the second data to infer the physical causal relationship between the internal thermal abnormal area and the surface stress abnormal area.

[0021] Further, the defect physical cause inference engine is a graph neural network model; the graph neural network model is configured to: Take the thermal abnormal points in the first data and the stress abnormal points in the second data as nodes of a graph; Based on the spatial proximity relationship between nodes and the preset material mechanics constraints, the edges of the graph are constructed; and, By propagating node information on the graph, a quantitative evaluation result of the spatial causal correspondence relationship is output.

[0022] Further, the association processing of the defect physical cause inference engine comprises: A candidate association pair generation stage for screening candidate abnormal pairs that exist in a causal association in space from all thermal abnormal points and stress abnormal points based on spatial proximity; and, A causal confidence evaluation stage, wherein the graph neural network model is only constructed and information is propagated for the candidate abnormal pairs, and finally a causal confidence score is formed.

[0023] Further, the cause-effect confidence evaluation stage further comprises: Based on the first data, an internal defect severity metric is determined for each hot anomaly point; In the information propagation process of the graph neural network model, the information propagated from the hot anomaly point node to the stress anomaly point node is modulated by a dynamic cause-effect correlation weight; The dynamic cause-effect correlation weight is dynamically generated based on the internal defect severity metric of the hot anomaly point node through a preset nonlinear mapping function.

[0024] The following specific implementation is described: In the embodiment, all configurable operating parameters, including but not limited to various thresholds, coefficients, mode selection identifiers mentioned below, are pre-defined and stored in a structured external data carrier. In a preferred embodiment, the data carrier is a file existing in an Excel format. When initializing, the processing device executing the method of the present application will read the file through a data loading module, load its content into the working memory, and all subsequent algorithm steps will directly query and obtain the required parameters from the memory data structure, so as to realize the technical decoupling of the core algorithm logic and the specific application strategy.

[0025] The thermal relaxation time constant is denoted as τ relax This parameter represents the rate at which the temperature of any point on the surface of the patch inductor decays from the peak value to the initial state after being instantaneously excited. The value is closely related to the local heat capacity and thermal conductivity of the point, and internal defects will increase the value. The thermal relaxation time constant of each pixel point is determined by the following steps executed by the data processing module: Obtain the time-temperature sequence data for the pixel point collected by the high-speed infrared thermal imager, which includes multiple time sampling points and their corresponding temperature values. Read the preset reference ambient temperature value from the external configuration file. Preprocess the temperature values in the time-temperature sequence data, i.e., subtract the reference ambient temperature value from the temperature value of each sampling point to obtain the temperature rise sequence. Select the maximum value in the temperature rise sequence as the peak temperature rise. Apply a preset exponential decay model to the temperature rise sequence for nonlinear least squares fitting. The idea of the exponential decay model is derived from Newton's cooling law in thermodynamics, and its calculation logic is described as follows: the temperature rise value at any time is equal to the peak temperature rise multiplied by the power of the natural constant with the negative ratio of the current time to the to-be-solved time constant as the exponent. In the fitting process, the to-be-solved time constant that minimizes the sum of squares of residuals between the model prediction value and the actual observation value is determined as the thermal relaxation time constant τ relax of the pixel point.

[0026] The thermal relaxation time constant τ relax is determined as follows: The original signal is derived from a high frame rate infrared thermal imager, whose key performance indicators are: the detector type is a cooled mid-wave infrared (MWIR) detector to ensure high thermal sensitivity (NETD < 20 mK); its highest sampling frame rate is not less than 1000 Hertz (Hz) to accurately capture the transient thermal response process. The original signal collected is a three-dimensional data cube represented by temperature values, denoted as T(x, y, t), where x and y are pixel spatial coordinates, and t is the time sequence. The data cube completely records the complete temperature decay curve of each pixel point on the patch inductor surface after being subjected to transient thermal excitation.

[0027] Before model fitting, the original signal T(x, y, t) is preprocessed. The pipeline includes: Extract several frames of images before the occurrence of thermal excitation, calculate the average temperature of each pixel point as its initial reference temperature T base (x, y). Then, subtract the initial reference temperature T base (x, y) of all time points in the data cube from the temperature value T(x, y, t) to obtain the temperature rise data cube ΔT(x, y, t).

[0028] To reduce the influence of detector noise on subsequent fitting accuracy, a low-pass filter is applied to ΔT(x, y, t) in the time dimension. In the preferred embodiment, a third-order Savitzky-Golay filter is used, with a window length of 5 time sampling points.

[0029] Extract the calculation model of the thermal relaxation time constant τ relax , whose physical basis is derived from Newton's cooling law in thermodynamics, which describes the temperature decay process of an object over time under certain conditions following an exponential law.

[0030] For each spatial coordinate (x, y) in the preprocessed temperature rise data cube ΔT(x, y, t), the parameter extraction module performs the following calculation logic to determine its thermal relaxation time constant τ relax value: Obtain the temperature rise time series ΔT series of the point, and the peak temperature rise ΔT peak in the series, i.e., the maximum value in the series.

[0031] Using the nonlinear least squares method, the temperature rise time series ΔT series is fitted to a single exponential decay model. The calculation logic of the model is described as follows: the predicted temperature rise value at any time t is calculated by calculating the peak temperature rise ΔT peakThe product of the exponential term is obtained; the exponential term is calculated in the following way: taking the natural constant e as the base, multiplying the time t by negative one, and dividing the result by the parameter to be solved, which is the thermal relaxation time constant τ relax .

[0032] The goal of the fitting algorithm is to find the thermal relaxation time constant τ relax value that minimizes the sum of the squares of the differences between all the predicted temperature rise values calculated by the above model and the actual observed temperature rise values over the entire time series. The thermal relaxation time constant τ relax value obtained after the Levenberg-Marquardt algorithm converges is finally determined as the thermal relaxation time constant of the pixel point.

[0033] The linear polarization degree is denoted as DoLP, and the linear polarization angle is denoted as AoLP: DoLP represents the proportion of the linearly polarized component in the scattered light and is sensitive to surface roughness and material changes. AoLP represents the vibration direction of the linearly polarized component and is extremely sensitive to the directionality of the surface microstructure and the birefringence effect caused by mechanical stress; the DoLP and AoLP values of each pixel point are determined by the following steps: Obtain four light intensity values corresponding to four detection directions of 0°, 45°, 90°, and 135° collected by the polarization camera in a single exposure, denoted as I0, I 45 , I 90 , and I 135 .

[0034] According to the Stokes vector theory used to describe the state of polarized light in optics, three Stokes parameters S00, S11, and S22 are calculated. S00 is calculated by adding I0 and I 90 . S11 is calculated by subtracting I 90 from I0. S22 is calculated by subtracting I 45 from I 135 .

[0035] Determination of the linear polarization degree DoLP: calculate the square of S11 and the square of S22, and add them to obtain an intermediate sum. Calculate the arithmetic square root of the intermediate sum. Divide the arithmetic square root by the parameter S00, and the quotient obtained is the DoLP value.

[0036] Determination of the linear polarization angle AoLP: calculate the ratio of S22 to S11. Calculate the inverse tangent value of the ratio. Multiply the inverse tangent value by one-half, and the result obtained is the AoLP value.

[0037] The specific determination method of DoLP and AoLP is as follows: The original signals are captured from an industrial CMOS camera integrated with an array of on-chip micro-polarizers. Each 2x2 pixel unit (super-pixel) of the camera is covered with four linear polarized filters of different orientations, 0°, 45°, 90° and 135° respectively.

[0038] After a single exposure of the camera, four registered gray-scale images of the same full resolution of the camera are outputted directly by the internal demosaicing algorithm, representing the light intensity in the four polarization directions, denoted as I0(x,y), I 45 (x,y), I 90 (x,y), I 135 (x,y).

[0039] Parameter calculation based on Stokes vector model: The calculation of parameters follows the Stokes vector representation method used in the field of optics to describe the state of polarized light. Specifically, for each pixel coordinate (x,y), the polarization calculation module performs the following logic: Obtain the four light intensity values I0, I 45 , I 90 , I 135 . Calculate the first component S00: add I0 and I 90 . Calculate the second component S11: subtract I 90 from I0. Calculate the third component S22: subtract I 45 from I 135 .

[0040] Calculate the square of S11 and the square of S22 respectively; add the two square values to obtain an intermediate sum.

[0041] Calculate the arithmetic square root of the intermediate sum; divide the arithmetic square root by S00, and the quotient obtained is determined as DoLP. The value is normalized in the interval [0,1].

[0042] Calculate the ratio of S22 to S11. Call the standard four-quadrant arctangent function with S22 as the first input (corresponding to the y coordinate) and S11 as the second input (corresponding to the x coordinate) to calculate the radian value. Multiply the calculated radian value by one-half, and the result obtained is determined as AoLP. The value ranges in the interval [-π / 2, +π / 2] or [0,π].

[0043] Internal defect severity metric, denoted as S(internal): This is a dimensionless parameter that is normalized and quantifies the severity of internal defects corresponding to a thermal anomaly point. Its numerical range is limited to the interval [0,1], and the larger the value, the more severe the internal defects. The present embodiment performs the following steps to determine the severity metric of each thermal anomaly point: read the standard thermal relaxation time constant τ standard and the maximum tolerated thermal relaxation time constant τ max . τ standard representing the average τ relax value of the defect-free good region max representing the τ relax upper limit set empirically to indicate a serious defect.

[0044] For a given thermal anomaly point, obtain its thermal relaxation time constant τ relax . Calculate the difference between τ relax and τ standard . If the difference is negative or zero, S(internal) is set to 0. If the difference is positive, further calculate the difference between τ max and τ standard as the denominator. Divide the obtained difference by the obtained denominator to obtain a quotient. If the quotient is greater than 1, S(internal) is set to 1; otherwise, the value of S(internal) is the quotient. This process ensures that S(internal) is strictly normalized to the interval [0, 1].

[0045] The embodiment specifically determines the internal defect severity metric S(internal) as follows: through controllable experiments, determine two key internal constants used to calculate S(internal): the standard thermal relaxation time constant τ standard and the maximum tolerated thermal relaxation time constant τ max .

[0046] Prepare two groups of patch inductor samples. The first group is the "at least 30 good sample set" confirmed by X-ray non-destructive testing (NDT) or final line testing (FVT) to be free of any internal defects. The second group is the "at least 30 critical defect sample set" by intentionally introducing process deviations during production to reduce sintering pressure, or by means such as focused ion beam (FIB) to manufacture internal cavities or micro-cracks with sizes just above the design specification upper limit.

[0047] Set up the following thermal excitation and infrared acquisition system exactly the same as the actual production line: Perform the experiment under the same temperature and humidity (25°C±1°C, 50%±5%RH) as the production line.

[0048] For each sample in the "good sample set", repeat the aforementioned τ relax measurement process at least 10 times. Record the τ relax value of all samples, all measurement times, and all pixel points.

[0049] For each sample in the "critical defect sample set", repeat the measurement 10 times as well. Record the τ relax value of the pixel located right above the known defect region.

[0050] For all the collected good product τ relax data, calculate the statistical average value. This average value is determined as τ standard .

[0051] For all the collected critical defect τ relax data, calculate the 99th percentile of the statistical distribution. The percentile is used instead of the maximum value in this embodiment to exclude the influence of extreme outliers and enhance the robustness of the parameter. The value of the 99th percentile is determined as the maximum tolerable thermal relaxation time constant τ max . Substitute the calibrated τ standard and τ max into the calculation logic of S(internal) and apply them to the two groups of sample data. Verify that the S(internal) values of all good product samples tend to 0, while the S(internal) values of all critical defect samples tend to 1, to confirm the effectiveness of the calibration results.

[0052] The dynamic causal correlation weight, denoted as W causal , is a dynamic weight factor normalized to the interval [0, 1]. It is used to adjust the influence strength of the internal thermal anomaly node on its associated surface stress anomaly node during the information propagation process of the graph neural network. The core idea is that the more serious the internal defect, the higher the causal correlation credibility of the internal defect and the surface anomaly, and the larger the weight. The weight generation module dynamically generates this weight for each thermal anomaly point in the candidate correlation pair through the following steps: Obtain the internal defect severity metric S(internal) of the thermal anomaly point. Read two parameters that control the shape of the non-linear mapping from the external configuration file: the gain coefficient k and the offset x0.

[0053] Apply a preset S-shaped activation function for non-linear mapping. The calculation model is: subtract the offset x0 from S(internal) to get a difference; multiply the difference by the negative gain coefficient k; calculate the power with the natural constant as the base and the product obtained in the previous step as the exponent; add 1 to the power value; calculate the quotient of 1 and the sum obtained in the previous step. The final calculation result is the dynamic causal correlation weight W causal . This embodiment specifically adopts the mathematical form of the Sigmoid function; it can map inputs of any range to the interval [0, 1]; The gain coefficient k controls the steepness of the S-shaped curve near the inflection point. The greater the value of k, the steeper the curve, and the more dramatic the transition from 0 to 1 in the weight, representing a more "sensitive" switching logic.

[0054] The offset x0 determines the center position of the S-shaped curve, i.e. when the input internal defect severity metric S (internal) is equal to x0, the output dynamic causal correlation weight W causal is exactly 0.5. It represents the "inflection point" or "threshold" of the decision. The values of k and x0 are determined as hyperparameters during the system integration and debugging phase. The goal of optimization is to maximize the root cause diagnosis accuracy (RCDA) indicator mentioned above. In a specific embodiment, the Grid-Search method is used to perform combination testing within the preset k value range [1, 20] and x0 value range [0.1, 0.9], and finally the (k, x0) value that makes RCDA reach the highest on the validation set is selected and fixed in the external configuration file.

[0055] Further, the configuration file is set to contain at least the following fields: parameter name, parameter value, and parameter description. A specific example is shown in Table 1 as follows: Table 1: Configuration file contains field table ; The graph neural network fusion method assigns static or uniform weights to all edges established based on spatial proximity, which has limitations when dealing with scenarios with clear physical causal strength relationships. It cannot reflect the expert diagnosis logic that "a serious internal defect, even if it produces weak stress signals on the surface, its relevance should be highly valued". Therefore, the hierarchical and dynamically weighted defect physical cause inference engine designed in this embodiment has the following internal working mechanism: In the first stage, candidate correlation pairs are generated: abnormality detection is performed on the thermal relaxation time constant map and the linear polarization angle map to obtain the thermal abnormal point set and the stress abnormal point set, respectively. A k2-d2 tree (k-dimensional-tree) spatial index structure from the field of computational geometry is used. A two-dimensional k2-d2 tree is constructed for the stress abnormal point set.

[0056] For each point in the thermal abnormal point set, a radius search is performed in the k2-d2 tree. A preset radius around its spatial position is read from the external configuration file, and all stress abnormal points within the preset radius are found. Each thermal abnormal point and all the stress abnormal points it searches form a candidate correlation pair.

[0057] The second stage, the causal confidence evaluation: only for the candidate association pair set generated in the last stage, the graph neural network is constructed. The nodes of the graph are all the abnormal points in the candidate association pairs. In the iteration process of information propagation performed by the graph neural network, when information is propagated from the hot anomaly point node i to the stress anomaly point node j, the propagation strength is dynamically adjusted.

[0058] The standard graph convolution network information propagation logic is described as follows: the new feature of node j is obtained by performing linear transformation on the feature of node i, multiplying the static normalization coefficient, and finally performing aggregation.

[0059] In the present application, the process of the standard graph convolution network information propagation logic is modified as follows: after the linear transformation and before the aggregation, the information flow must be additionally multiplied by the dynamically calculated dynamic causal association weight W causal .

[0060] Since the dynamic causal association weight W causal is a nonlinear increasing function of the internal defect severity metric S (internal): when the internal defect severity metric S (internal) tends to 0, the dynamic causal association weight W causal tends to a smaller value, and the information propagation is inhibited; when the internal defect severity metric S (internal) tends to 1, the dynamic causal association weight W causal tends to 1, and the information propagation is enhanced.

[0061] After multiple rounds of iteration and propagation, a causal confidence score between 0 and 1 is finally output for each pair of candidate association pairs, which is the quantitative evaluation result of the spatial causal correspondence relationship between them.

[0062] The overall calculation process of the above steps S1 to S3 is as follows: The overall calculation process of the embodiment of the present application is executed by the central control program, and the specific steps are as follows: The control program starts and executes the data loading module to read all algorithm parameters and system configurations from a specified external spreadsheet file and stores them in the memory.

[0063] The control program sends instructions to the synchronization controller to trigger the heat excitation source, the high-speed infrared thermal imager and the polarization camera in turn to obtain the original thermal image sequence and the four-channel polarized light intensity image.

[0064] The thermal image sequence is processed to generate a thermal relaxation time constant map. The four-channel polarized light intensity image is processed to generate a linear polarization degree map and a linear polarization angle map; the pre-calibrated affine transformation matrix is applied to align the above three feature maps to a unified pixel coordinate system.

[0065] According to the thermal relaxation time constant map and the reference parameters in the profile, an internal defect severity metric map is calculated.

[0066] The defect physical cause inference engine (DPIE) performs a first stage: calling a defined candidate association pair generation algorithm, inputting the registered feature map, and outputting a candidate association pair list.

[0067] The defect physical cause inference engine (DPIE) performs a second stage: constructing a graph neural network for the candidate association pair list. In the information propagation process, for each thermal anomaly point node, the severity metric is queried from the internal defect severity metric map, and the dynamic causal association weight W is dynamically calculated causal , and the weight is applied to modulate the information propagation. After the graph neural network is executed, the causal confidence score of each candidate association pair is output, forming the final defect physical cause diagnosis signal. The signal is then sent to the subsequent decision and control module.

[0068] It should be noted that: in order to ensure that the "first data" obtained from the infrared thermal imager and the "second data" obtained from the polarization camera are accurately aligned at the pixel level, the present embodiment adopts an offline calibration method to determine an affine transformation matrix. The affine transformation matrix is used to convert the pixel coordinate system of the infrared image to the pixel coordinate system of the polarization camera; The calibration platform includes a rigid optical support for co-axially or off-axially fixing the infrared thermal imager and the polarization camera at a preset relative pose. The calibration board used is an alumina ceramic board with a high-emissivity coating (emissivity > 0.95) sprayed on it. The board is processed with a 10-row and 10-column circular marker array, and the marker points are bare metal (low emissivity) circles with a diameter of 3 mm. The spacing between the centers of all the marker points is 15 mm in the horizontal and vertical directions, and the processing accuracy is better than 0.01 mm. This design ensures that the marker points have high contrast in the infrared image (based on temperature difference or emissivity difference) and the visible light image.

[0069] The calibration board is placed in the common field of view of the two cameras. By moving the calibration board, 15 pairs of infrared images and polarization images are synchronously collected from at least 15 different spatial positions and poses.

[0070] For each visible light polarization image, an image moment-based algorithm is called to accurately calculate the sub-pixel center coordinates of each circular marker point.

[0071] For each infrared image, a threshold segmentation algorithm is applied to highlight the low-emissivity marker points, and then an image moment-based algorithm is also called to calculate the sub-pixel center coordinates of the marker points.

[0072] For each pair of images collected, the N feature points coordinates (u ir ,v ir ) extracted from the infrared image are one-to-one corresponding to the N feature points coordinates (up,vp) extracted from the polarized image according to the row and column positions of the markers in the array, forming N matching coordinate pairs.

[0073] The calibration computing module collects all the matching coordinate pairs collected and performs the following calculation logic to solve the affine transformation matrix containing six unknown parameters (a, b, c, d, e, f): Input coordinate point set: {(u ir ,v ir )}; u ir represents the horizontal (column) pixel coordinate of the i-th marker point in the infrared image. v ir represents the vertical (row) pixel coordinate of the i-th marker point in the infrared image. This set {(u ir ,v ir )} constitutes the source coordinate point set to be transformed.

[0074] Target coordinate point set (target coordinate system): {(up,vp)}; up represents the horizontal (column) pixel coordinate of the same physical point corresponding to the i-th infrared marker point in the polarized image. vp represents the vertical (row) pixel coordinate of the same physical point corresponding to the i-th infrared marker point in the polarized image. This set {(up,vp)} constitutes the true value target point set for calibration.

[0075] Core parameters to be solved for affine transformation matrix: a, b, c, d, e, f; these parameters collectively define a linear transformation in a two-dimensional plane. Each of them has a clear geometric meaning, collectively describing all the operations required to "twist" the source coordinate system into the target coordinate system: c describes the overall shift in the horizontal direction. Its physical meaning is to translate the entire infrared image as a whole in the horizontal direction by c pixels. f describes the overall shift in the vertical direction. Its physical meaning is to translate the entire infrared image as a whole in the vertical direction by f pixels. a describes the scaling ratio in the horizontal direction. If the value of a is 1.05, it means that the infrared image is enlarged by 5% in the horizontal direction. e describes the scaling ratio in the vertical direction. If the value of e is 0.98, it means that the infrared image is reduced by 2% in the vertical direction. b describes the degree of skew or tilt of the image. Specifically, it quantifies the degree of influence of the vertical coordinate (v ir ) on the final horizontal coordinate (up). A non-zero b value means that the vertical lines in the image will be tilted after transformation. d describes the degree of skew or tilt of the image. It quantifies the degree of influence of the horizontal coordinate (u irthe degree of influence on the final vertical coordinate (vp). A non-zero value of d means that horizontal lines in the image will be tilted after the transformation. The four parameters a, b, d, e together determine the scaling, rotation and shearing of the source image. c and f are responsible for the final translation alignment after the above operations.

[0076] Obtain all the matched coordinate pairs set; for each matched pair, according to the mathematical principle of affine transformation (i.e., up = a x u ir + b x v ir + c and vp = d x u ir + e x v ir + f), construct an overdetermined linear equation set.

[0077] Solve the overdetermined equation set by using the linear least square method based on singular value decomposition (SVD). This method can find a set of optimal parameters (a, b, c, d, e, f) so that the sum of the squared Euclidean distances between the new coordinate points obtained by the transformation defined by the set of parameters and the corresponding actual polarized camera coordinate points is minimized. Assemble the six solved parameters into an affine transformation matrix of three rows and three columns. The first row of the matrix is [a, b, c], the second row is [d, e, f], and the third row is [0, 0, 1]. The calculated affine transformation matrix is serialized and saved as a configuration file. When the system is running normally, the matrix is first loaded into memory.

[0078] Further, the core result finally output by the defect physical cause inference engine (DPIE) is the causal confidence score for each candidate abnormal pair. The causal confidence score is a dimensionless value normalized to the interval [0, 1]; When the causal confidence score tends to 1, it means that the defect physical cause inference engine judges that the internal thermal anomaly indicated by the first data is the root physical cause of the surface stress anomaly corresponding to it in space, with a higher confidence. This state clearly points to the "endogenous and exogenous" high-risk complex defect mode.

[0079] When the causal confidence score tends to 0, it means that the defect physical cause inference engine judges that the physical causal relationship between the internal thermal anomaly and the surface stress anomaly is weaker. This state indicates that even if the two are adjacent in space, they belong to two independent and unrelated events; an example is a harmless internal material benign heterogeneous point that happens to coincide with an independent surface processing scratch in projection.

[0080] The causal confidence score represented by the final output of the defect physical cause inference engine in this embodiment is affected by a series of key input parameters in a hierarchical and nonlinear manner. Its internal logical relationship conforms to the real physical law and expert diagnosis experience: thermal relaxation time constant τ relax The increase in the value will lead to a monotonic increase in the "internal defect severity metric". It is a positively correlated piecewise linear relationship; thermal relaxation time constant τ relax is a direct physical quantity representing the degree of difficulty in local heat dissipation. According to the theory of heat conduction, internal defects (including cavities, delamination) as obstacles to heat flow will prolong the time of heat dissipation, thereby increasing the value of thermal relaxation time constant τ relax . Therefore, the larger the value of thermal relaxation time constant τ relax , the more serious the internal structural defects it corresponds to.

[0081] The increase in the internal defect severity metric will lead to a monotonic increase in the "dynamic causal correlation weight"; it is a positively correlated, nonlinear relationship; the dynamic causal correlation weight aims to simulate the "risk attention" mechanism in expert diagnosis. In physical logic, the more serious the internal structural defects, the greater the possibility and intensity of observable anomalies on the surface through the stress conduction mechanism. When the internal defect severity metric is low, its causal correlation with any surface anomaly should be given a very low weight; and when the internal defect severity metric exceeds the critical value, its correlation weight should also be quickly increased and saturated. A S-shaped nonlinear mapping function is used to generate the dynamic causal correlation weight, which reproduces this nonlinear decision logic from "basic neglect" to "high attention".

[0082] The increase in the value of the dynamic causal correlation weight will lead to a monotonic increase in the information intensity propagated from the corresponding thermal anomaly point node to the stress anomaly point node in the graph neural network model, thereby ultimately leading to a monotonic increase in the "causal confidence score" of the anomaly pair; in the information propagation mechanism of the graph neural network, when the dynamic causal correlation weight approaches 0, even if the two nodes are spatially adjacent, the information flow from the thermal anomaly node will be severely attenuated and cannot have a substantial impact on the final state of the neighbor node (stress anomaly point). On the contrary, when the dynamic causal correlation weight approaches 1, the information flow is transmitted without loss or enhanced, so that the features of the thermal anomaly node can strongly influence the judgment of its associated stress anomaly node.

[0083] The core technical feature of the embodiment is that the method for processing defect physical cause diagnosis signals and generating linkage control instructions converts the diagnosis results output by the upstream inference engine, containing multi-dimensional physical information, into a structured indication signal composed of a qualitative "defect classification label" and a quantitative "quantitative risk index" through a structured, hierarchical, and dynamically adaptive risk quantification model, which has clear physical meaning and decision guidance value. This signal is then used to simultaneously drive real-time sorting of individual units at the microscopic level and process cycle feedback at the macroscopic level, thereby building an intelligent manufacturing closed loop from detection, diagnosis to prevention.

[0084] Further explanation: in S4, the defect classification label is selected from a preset label set, and the label set at least includes: "internal structural defect", "surface processing defect", and "endogenous and exogenous composite defect"; The quantitative risk index is a normalized value representing the failure probability of the patch inductor under a preset working condition.

[0085] Further explanation: the generation of the quantitative risk index is realized through a two-stage hierarchical risk fusion model; specifically configured as: In the first stage, based on the defect classification label, a reference risk value representing the prior hazard level of the defect category is determined from a preset defect risk reference library; In the second stage, at least one dynamic severity metric representing the severity of the current physical state of the defect is determined based on the first data and the second data corresponding to the patch inductor, and the reference risk value is nonlinearly modulated using the dynamic severity metric to generate the final quantitative risk index.

[0086] Further explanation: the dynamic severity metric further includes a time dimension enhancement factor; The dynamic risk escalation module is configured to continuously monitor the change trend of the internal thermal abnormality feature over time indicated by the first data; When it is determined that the internal thermal abnormality feature presents a deterioration trend, the value of the time dimension enhancement factor is dynamically increased; The time dimension enhancement factor is used to amplify the modulation amplitude of the dynamic severity metric to the reference risk value in the nonlinear modulation process in the second stage.

[0087] Further explanation: the dynamic severity metric further includes a spatial dimension correlation factor; The spatial consistency evaluation module is configured to: a spatial distance between a geometric center of the internal thermal anomaly region indicated by the first data and a geometric center of the surface stress anomaly region indicated by the second data; based on the spatial distance, determining a numerical value of the spatial-dimension correlation factor, wherein the smaller the spatial distance is, the larger the numerical value of the spatial-dimension correlation factor is; the spatial-dimension correlation factor is used to further amplify the modulation amplitude of the dynamic severity metric on the baseline risk value in the nonlinear modulation process of the second stage.

[0088] The following specific implementation is made to the above: In the embodiment, the calculation model for generating the structured indication signal involves the following key parameters. All these parameters, if they are configurable items, are pre-defined and stored in an external, local storage spreadsheet file, which is read into memory by the parameter loading module at system initialization.

[0089] a defect classification label, denoted as Label defect : which represents a categorical variable of string type; the physical meaning is a qualitative judgment of the fundamental physical cause of the currently detected defect. Its determination method is: obtain the classification probability vector output by the defect physical cause inference engine (DPIE) in the aforementioned step S3, each dimension of the vector corresponds to a pre-set defect category; the label determination module determines the category name corresponding to the dimension with the highest probability value as the defect classification label Label defect of the defect by performing the maximum probability value operation. The value set of the label is pre-set to at least include: "internal structural defect", "surface processing defect", and "endogenous and exogenous composite defect"; this parameter is the direct output of the defect physical cause inference engine (DPIE) as the core algorithm.

[0090] The application target of DPIE is to classify the fundamental physical cause of the detected defect. DPIE preferably adopts a graph attention network (Graph-Attention-Network, GAT) architecture. Its structure includes: an input layer for receiving graph node features composed of thermal anomaly points and stress anomaly points; three GAT convolution layers for aggregating neighbor node information through attention mechanism; an output layer, which is a fully connected layer with a Softmax activation function, the number of output nodes of which is equal to the total number of pre-set defect categories.

[0091] The training process includes failure samples with root causes precisely labeled using X-rays and SEM. It contains at least 3000 samples, covering all predefined defect categories with a relatively balanced category distribution. Each sample's label serves as the training label. The dataset is divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. The cross-entropy loss function is used. The Adam optimizer is employed; the initial learning rate is set to 0.001, the batch size to 32, and the training epochs to 200, with early stopping to prevent overfitting. After training, the optimal model weights are fixed and deployed to the detection system. During inference, the category corresponding to the dimension index with the largest value in the DPIE output Softmax layer vector is determined as the label. defect .

[0092] The benchmark risk value is denoted as Risk. base : A floating-point number ranging from [0,1]. Its physical meaning is the prior risk level for different defect categories defined based on historical data and an expert knowledge base. It is determined as follows: the system maintains a "defect category - baseline risk" mapping table (stored in an external configuration file), and this table is used when determining the Label. defect Then, by looking up a table, the corresponding benchmark risk value (Risk) can be directly obtained. base The baseline risk value (Risk) corresponding to "surface processability defects". base The value is set to 0.3 for "internal structural defects," 0.6 for "endogenous-exhibitive composite defects," and 0.9 for "endogenous-exhibitive composite defects." This parameter maps the qualitative classification of different defects to a quantitative value representing their prior hazard level. Its determination process is based on statistical offline calibration experiments.

[0093] This embodiment establishes a quantitative mapping relationship between "defect classification labels" and "long-term failure probability".

[0094] For each of the 1000 failure samples, multimodal image data was acquired. The sample was then meticulously dissected and analyzed using X-ray microscopy and SEM. At least two senior failure analysis engineers jointly determined the root cause of the failure and categorized it into a pre-defined tag set (including "internal voids leading to delamination," "surface microcrack propagation," etc.). The engineers' "gold standard" tags were then compared with the labels automatically generated by the system of this invention. defectThe classification accuracy of DPIE is verified by comparison. The number of samples finally determined to belong to the three categories of "internal structural defects", "surface processing defects" and "endogenous and exogenous compound defects" among all failure samples is counted, denoted as N(internal), N(surface), and N(compound) respectively. The failure contribution rate of each category is calculated. The contribution rate P(internal) of "internal structural defects" is calculated as follows: N(internal) is divided by N(internal), N(surface), and N(compound) respectively, then the sum of the three quotients is obtained, and finally N(internal) is divided by the sum. The calculated failure contribution rate of each category is used as the initial value of the corresponding reference risk value Risk internal . base

[0095] The normalized internal defect severity S(internal, norm) is a floating-point number with a value range in the interval [0, 1]. Its physical meaning is the quantitative evaluation of the severity of internal thermal anomalies. Its determination method is described in the aforementioned embodiment of the internal defect severity metric, and S(internal, norm) is the internal defect severity metric value.

[0096] The normalized surface stress severity is denoted as S(stress, norm): a floating-point number with a value range in the interval [0, 1]. Its physical meaning is the quantitative evaluation of the severity of surface stress anomalies. Its determination method is as follows: the stress analysis module calculates the average value of the absolute value of the linear polarization angle (AoLP) of all pixel points in the stress anomaly area, obtaining AoLP mean ; then the module performs the following calculation logic for normalization: two parameters in the configuration file, "normal stress upper limit (AoLP norm )" and "limit stress threshold (AoLP max )", are obtained; calculate "AoLP mean " minus AoLP norm " divided by the difference between "calculated AoLP max " and AoLP norm "; the quotient of the calculation result is truncated in the interval [0, 1] to obtain S(stress, norm).

[0097] It should be noted that AoLP norm , AoLP max ​is determined by offline calibration experiments. The detailed steps of this experiment are as follows: prepare two sample sets, i.e. a healthy sample set containing at least 200 samples confirmed by X-ray detection to be free of defects, and a limit defect sample set containing at least 200 samples confirmed by SEM to have serious defects. Secondly, perform complete detection on each sample in the two sample sets, and record the average thermal relaxation time constant and the average absolute linear polarization angle of each sample, respectively. Then, the parameter calibration module performs the following calculation logic: For all the AoLP_mean values recorded for the healthy sample set, calculate the 99% percentile of the statistical distribution thereof, and determine this percentile value as the "normal stress upper limit (AoLP norm )". For all the AoLP mean values recorded for the limit defect sample set, calculate the 99% percentile of the statistical distribution thereof, and determine this percentile value as the "limit stress threshold (AoLP max )".

[0098] Time dimension enhancement factor, denoted as γt: a floating-point number with a value range greater than or equal to 1. Its physical meaning is to characterize whether the internal defect has a dynamic trend of deterioration over time. Its determination method, the dynamic risk upgrading module performs the following calculation logic: maintain a queue of S(internal, norm) historical values for the same patch inductance in consecutive production batches; apply first-order difference calculation to the data in this queue to obtain the change rate ΔS of S(internal, norm); obtain a preset "change rate gain coefficient kt"; calculate the product of ΔS and kt; add the product to the reference value 1 to obtain γt. If ΔS is negative or zero, then γt is 1, with no enhancement effect.

[0099] Spatial dimension correlation factor, denoted as γs: the spatial consistency evaluation module performs the following calculation logic: calculate the Euclidean distance d1 between the two center points; obtain a preset "spatial attenuation coefficient ks" and a "maximum correlation factor γs max ", and the preferred value in this embodiment is 2.0. Then, perform the following steps: First step: multiply the Euclidean distance d1 by the negative ks to obtain an exponential term; second step: calculate the power of the exponential term with the natural constant e as the base to obtain an attenuation factor; third step: subtract the reference value 1 from the maximum correlation factor γs max to obtain the maximum enhancement amplitude; fourth step: multiply the maximum enhancement amplitude by the attenuation factor obtained in the second step; fifth step: add the result of the fourth step to the reference value 1 to obtain the final γs. When the two center points completely coincide (d1 = 0), the attenuation factor is 1, and γs takes the maximum value γs max , i.e. 2.0; when the distance d1 tends to infinity, the attenuation factor tends to 0, and γs smoothly attenuates to its reference value 1, i.e. loses the enhancement effect.

[0100] Note that the rate of change gain coefficient kt is determined by sensitivity tuning experiments. In the experiments, a series of defects with known rates of deterioration (5% increase per hour) are artificially created, and the value of kt is adjusted so that the risk index output is stable when the deterioration rate reaches the preset warning line, triggering the threshold of “to be reviewed” or “waste”.

[0101] The spatial attenuation coefficient ks is determined by spatial correlation calibration experiments. In the experiments, a large number of “endogenous and exogenous complex defects” samples are analyzed, and the spatial distance distribution of internal thermal anomaly centers and surface stress anomaly centers is counted. The value of ks is set so that when the distance exceeds the 95% percentile of the distribution, the value of γs is attenuated to close to 1.05, i.e. basically losing the enhancement effect.

[0102] The quantified risk index is denoted as Risk final : a floating point number with a value range in the interval [0, 1], which is one of the final outputs of this step. Its physical meaning is the final quantitative evaluation of the failure probability of the patch inductor under the preset working condition after considering the defect category, current severity, dynamic deterioration trend, and spatial correlation.

[0103] This embodiment is designed to overcome the limitations of conventional linear weighting methods, which cannot reflect the complex nonlinear relationship between different dimensions of information, and are prone to misjudgment due to extreme values of a single indicator. This embodiment designs the following hierarchical, time-space dynamic enhancement risk quantization engine: This embodiment discloses that the final failure risk of a patch inductor is not simply the sum of its multiple physical characteristics. It is determined by the fundamental nature of the defect (Label defect → Risk base ) that determines the baseline level of its risk. Secondly, the risk will fluctuate according to its current physical severity [S (internal, norm), S (stress, norm)]. Defects that are dynamically deteriorating (γt>1) and have a high degree of physical cause (γs>1) should be nonlinearly and significantly improved in risk level.

[0104] The traditional fixed weight summation method has a fundamental flaw: it cannot distinguish between “a defect with a high-risk category but a low current value” and “a defect with a low-risk category but a very high current value”, and it cannot model “meta-features” such as dynamic changes and spatial correlations. This embodiment uses the two-stage hierarchical risk fusion model described above.

[0105] The first stage (determine the baseline): based on Label defect , look up table to get Risk base .

[0106] Second stage (non-linear modulation): The core of this stage is the dynamic severity aggregation module. This module aggregates two severity metrics, whose calculation logic is: take the maximum value in S(internal, norm) and S(stress, norm) as the base severity S base . This maximum value taking strategy, the final risk is determined by the most severe one. Subsequently, the modulation factor calculation module performs the following logic: multiply γtand γs, to get the composite enhancement factor γ compound . The risk fusion module performs the following calculation logic to generate the quantized risk index Risk final : calculate the product of S base and γ compound ; add this product to Risk base , to get an intermediate risk value; apply the sigmoid saturation function to constrain and map this intermediate risk value to the [0, 1] interval, to get the final quantized risk index Risk final .

[0107] To ensure that the final output risk index is located in the [0, 1] interval and has good nonlinear response characteristics, the risk mapping module applies the logistic function to perform saturation mapping processing on the intermediate risk value, to obtain the final risk index Risk final . The calculation logic performed by this module is as follows: obtain the input intermediate risk value Risk mid , and two preset shaping parameters in the configuration file, "mapping center point" and "mapping slope". Then, perform the following calculation steps: first, subtract the mapping center point from the intermediate risk value Risk mid ; second, multiply the result of the first step by the negative mapping slope, to get the exponential term; third, calculate the power of the exponential term with the natural constant e as the base; fourth, add 1 to the result of the third step; fifth, divide the result of the fourth step by 1, to obtain the final risk index Risk final . In this embodiment, the "mapping center point" is preferably set to 0.7, and the "mapping slope" is preferably set to 10, so that when the intermediate risk value is near 0.7, the final risk index will change most sensitively.

[0108] When S(internal, norm) and S(stress, norm) are both 0, S base is 0, and the risk index Risk final will approach Risk base infinitely (if Risk base itself is also 0, the result is 0), which is consistent with the logic of the defect-free state.

[0109] When the defect severity is high (i.e., S base=1)but no worsening trend and weak spatial correlation (γt=1, γs=1), the risk index Risk final will be mainly determined by Risk base and S base , and will not be over-amplified.

[0110] When the current severity of the defect is not high, i.e. S base =0.4, but the worsening trend is significant (γt=1.5) and the spatial consistency is high (γs=1.8), the product of γ compound and S base will be a larger value (2.7), which will boost the final risk index Risk compound , achieving a “warning” of the potential risk, even if S final is not large.

[0111] In a specific execution cycle, the complete calculation and control flow of steps S4 and S5 in the embodiment is as follows: A defect physical cause diagnosis signal for a single patch inductor is received from step S3, which contains at least the classification probability vector output by the DPIE, the S (internal) value, and the AoLP distribution map of the stress abnormal area. The label determination module is called to perform the maximum probability value operation on the classification probability vector to generate Label defect .

[0112] Parallelly, each parameter calculation module is called: Risk base is determined by table lookup. S (internal, norm) and S (stress, norm) are calculated and normalized.

[0113] The dynamic risk upgrade module is called to calculate γt in combination with historical data. The spatial consistency evaluation module is called to calculate γs.

[0114] The above parameters are fed into the hierarchical, time-space dynamically enhanced risk quantification engine to calculate the final risk index Risk final .

[0115] Label defect and the risk index Risk final are packaged into a structured indication signal.

[0116] The decision logic module reads the indication signal. The sorting thresholds preset in the configuration file are obtained, which are the high-risk sorting threshold Th1=0.9 and the medium-risk sorting threshold Th2=0.5.

[0117] If the risk index Risk final >Th1, a control instruction is generated to drive the mechanical arm to sort the inductor to the “waste” channel. If Th2<Riskfinal If the value is ≤Th1, then sort it to the "Pending Re-inspection" channel. Otherwise, sort it to the "Good Product" channel.

[0118] The statistical monitoring module stores this indication signal in a short-term database. Statistical analysis is performed at preset time intervals (every 5 minutes): this analyzes all labels within the past hour. defect The number of samples for "internal structural defects" is calculated; their frequency of occurrence is then compared with the frequency threshold of 5% set in the configuration file.

[0119] If the triggering condition is met (frequency > 5%), the process parameter adjustment suggestions associated with "internal structural defects" (including "suggesting to increase the sintering temperature by 5°C") are retrieved from the "defect-countermeasure" knowledge base, a process parameter adjustment instruction is generated, and sent to the upstream manufacturing execution system (MES) via the industrial Ethernet interface.

[0120] In a preferred embodiment of this invention, the "Defect-Countermeasure" knowledge base is a structured and scalable database table. The database table is constructed to map machine-recognizable defect labels to human-understandable and machine-executable process adjustment instructions. The table construction process is as follows: The implementer organizes multiple rounds of expert interviews with process engineers, equipment engineers, and quality engineers to systematically analyze the causal relationships between various defect phenomena and their most likely upstream process parameters. Secondly, historical production data, especially design experiment data, is retrospectively analyzed to verify and quantify these causal relationships using data mining. Finally, this verified knowledge is solidified into the database table. The table contains at least the following four fields: Defect tags: storage and label defect Completely consistent defect category names. Recommended instructions: Store structured instruction strings, including {parameter: SinteringTemp, operation: increase, value: 5, unit: Celsius}.

[0121] Target System: Stores the identifier of the target device or system to which the instruction should be sent, including "MES,Station,03" or "SinteringFurnace,01". Priority: Defines the urgency level of the instruction.

[0122] The following is a detailed description of the implementation of the above content: This invention generates a quantitative risk index, Risk. final It is a continuous value that has been normalized to the interval [0,1].

[0123] When the risk index Risk finalThe closer to 1: the confidence level indicates that there is a greater defect inside the current patch inductor, and there is a higher failure probability under the preset working condition. This state indicates that the functional or reliability of the device is more serious, and higher and more decisive measures need to be taken, and the defect cause data has higher analysis value for macro process feedback.

[0124] When the risk index Risk final The closer to 0: the confidence level indicates that there is a smaller defect inside the current patch inductor. This state indicates that the quality of the device is more qualified, and it can safely enter the next production link.

[0125] The final value of the risk index Risk final is the result of the fusion calculation of multiple key input parameters through a nonlinear, hierarchical model. The influence of each parameter on the output conforms to the real physical and engineering logic.

[0126] The reference risk value Risk base is strongly positively correlated with the risk index Risk final . When other parameters remain unchanged, the increase of the reference risk value Risk base will directly raise the reference level of the risk index Risk final .

[0127] The basic severity S base is positively correlated with the risk index Risk final . The basic severity S base directly quantifies the observable physical severity of the defect at the current time. The more serious the defect, the greater the immediate impact on the performance of the device, and the higher the potential risk of failure. This positive correlation design is the basic logic of risk assessment.

[0128] The time dimension enhancement factor γt: this parameter has a nonlinear positive amplification effect on the risk index Risk final . When γt is greater than 1, it will significantly amplify the contribution of the basic severity S base to the final risk as a multiplier.

[0129] Static and stable micro-defects have controllable risks, but micro-defects that are rapidly deteriorating and expanding have greater future risks. The design of γt is to capture the dynamic change trend of such risks. By introducing the time dimension enhancement factor γt, the invention is no longer "static" risk assessment, but "forward-looking" risk prediction, which accurately maps the objective laws of defect evolution in materials science.

[0130] The spatial dimension correlation factor γs: this parameter also has a nonlinear positive amplification effect on the risk index Risk finalHas nonlinear positive amplification effect. When the spatial dimension correlation factor γs is greater than 1, the contribution of the basic severity S will also be amplified base Single internal thermal anomaly or surface stress anomaly, with diverse causes. But when the two are highly coincident in space, it constitutes strong evidence that they are derived from the same deep physical cause. The design of γs, the core of which is to quantify the causal correlation strength of this multi-physical field information.

[0131] To quantitatively verify the beneficial effects of the technical solutions described in the present application, the following comparative experiments are designed. The "layered spatiotemporal dynamic enhancement model" of the present application is compared with a "linear weighted model" representing the existing conventional technology. This experiment aims to verify the superiority of the risk assessment model of the present application compared with the conventional linear model by constructing five typical defect scenarios with clear physical meaning. The conventional linear model can only reflect the simple superposition of the category and current severity of the defect, while the model of the present application introduces two nonlinear enhancement factors γt and γs, aiming to more accurately identify those "hidden" defects with dynamic deterioration trend or strong causal correlation, which are more dangerous. The data of this experiment will clearly show that the present application can output more accurate and more warning value risk index when facing such complex defects; see Table 2 below for details: Table 2: Comparison of risk assessment model outputs under different defect scenarios ; Comparison method risk index: To construct a fair comparison benchmark, the "comparison method risk index" representing the conventional technology approach is defined here. The calculation logic is: multiply the "baseline risk value" by the preset "category weight" (0.5 in this example); at the same time, multiply the "basic severity" by the preset "severity weight" (0.5 in this example); finally, add the two products to get the risk index of the comparison method. This method represents a linear combination model that only considers the category and current severity of the defect.

[0132] Comparing scenario two and scenario three: in scenario two, there exists a surface defect that is currently relatively severe (base severity = 0.50) but stable in nature (yt= 1.00). In scenario three, there exists an internal defect that appears to be mild (base severity = 0.20) but is rapidly deteriorating (yt= 1.80). The comparative method gives the exact same risk index (0.40) due to its inability to perceive dynamic changes, which does not conform to engineering practice as it completely overlooks the hidden great risk in scenario three. The model of the present invention, due to the introduction of yt, calculates a risk index as high as 0.78 for scenario three. This result, compared with the 0.40 of the comparative method, produces an absolute improvement of 0.38 on the normalized risk scale. More importantly, this improvement makes the risk assessment result directly jump from the low section of the “warning interval” to the high section of the “warning interval”, close to the boundary of the “danger interval”. This proves the effectiveness of the design of the time-dimension enhancement factor of the present invention, which can successfully identify and quantify the “developing” risk, achieving a technical progress from “static diagnosis” to “dynamic early warning”.

[0133] Comparing scenario two and scenario four: scenario four describes a compound defect with internal and external anomalies of comparable severity (base severity = 0.45) and highly overlapping spatial locations (ys= 1.90). The comparative method can only give a linearly superimposed risk value (0.68). The model of the present invention, due to the introduction of ys, interprets the strong correlation between internal and external signals as a sign of high risk, and calculates a risk index close to the failure boundary (0.98). This result, compared with the 0.68 of the comparative method, produces an absolute improvement of 0.30 on the normalized risk scale, and successfully pushes the risk assessment result from the “warning interval” to the “danger interval”. This proves the feasibility of the design of the spatial-dimension correlation factor of the present invention, which can improve the accuracy of diagnosis of complex and compound defects by quantifying the causal relationship of multi-modal information.

[0134] Scenario five describes a current severity of medium (base severity = 0.35), but at the same time in a fast deterioration (γt=1.50) and internal-external cause highly consistent (γs=1.80). Compared with scenario four, scenario five has a lower base severity (0.35 vs. 0.45), but additionally increases the dynamic deterioration trend (γt=1.50). The comparative method, due to the inability to perceive the change in the time dimension, even slightly reduces the risk index (0.63 vs. 0.68), completely misjudging the true direction of the risk. The model of the present application, under the synergistic amplification of γtand γs, its internal calculation of the intermediate risk value has far exceeded the saturation point, and the final output of the quantified risk index has reached the maximum value of 1.00. This shows that the model determines with the highest confidence that the defect is in a critical state of imminent failure. The comparison of this scenario with all other scenarios forms a complete chain of evidence, proving that the present application can identify the most hidden and most dangerous compound dynamic defects through the synergistic effect of space-time factors.

[0135] The final value of the quantified risk index of the present application is to drive specific and explicit automated control operations. The division of the output interval and the corresponding operation is based on a large number of experimental data ROC curve analysis, and is determined under the principle of maximizing defect detection rate while minimizing false positive rate. By analyzing the thousands of sample data used in the aforementioned experimental calibration, the risk index threshold value that can optimally distinguish between "good products", "products to be observed" and "definitely failed products" is determined; see Table 3 below for details: Table 3: Interval definition and corresponding operation ; Through the above division, the continuous and quantified risk index is converted into a set of discrete, executable, multi-level automated control strategies, fully achieving the technical goal of intelligent detection, accurate diagnosis, and closed-loop control.

[0136] Example two: please refer to Fig. 3 A patch inductor defect detection system based on machine vision, said system is used to execute the patch inductor defect detection method based on machine vision, comprising: a multi-physical field data acquisition unit for acquiring first data representing the internal heat conduction characteristics of the patch inductor; acquiring second data characterizing the surface microstructure of the patch inductor and the stress state; a spatial causal correlation diagnosis unit configured to perform correlation processing on the spatially registered first data and second data to identify a spatial causal correspondence between the internal thermal anomaly region indicated by the first data and the surface stress anomaly region indicated by the second data, and generate a defect physical cause diagnosis signal accordingly; a structured indication signal generation unit configured to generate a structured indication signal based on the defect physical cause diagnosis signal, the structured indication signal including at least a defect classification label and a quantitative risk index; a closed-loop linkage control execution unit configured to perform at least one of the following control operations based on the structured indication signal: performing a multi-level sorting operation on a single patch inductor based on its corresponding defect classification label and quantitative risk index; performing statistical analysis on the structured indication signal, and generating a process parameter adjustment instruction when it is determined that the frequency of occurrence of any category of defect classification label meets a preset condition, and sending the process parameter adjustment instruction to an upstream production equipment of the patch inductor.

[0137] It should be noted that all the calculation formulas in the present application file use regression analysis in machine learning algorithms, including but not limited to machine learning algorithms, to analyze the collected parameters in depth, identify their natural trends and mutual relationships. Professional software such as Python's Scikit-learn library or R language is used to automatically generate mathematical models that match the data. Then, the performance of the model is objectively evaluated through cross-validation and other methods, and combined with continuous feedback and optimization to ensure that the created formula truly reflects the inherent law of the data, thereby ensuring its effectiveness and accuracy. In all the calculation formulas in the present application, the parameters in each formula are processed by consistent range of dimensionless to ensure that different physical quantities are compared on the same scale; the dimensionless technique includes but is not limited to Min-Max-Normalization, Z-Score standardization; The algorithm method of the present application is realized as a Python program script. Before executing the core logic, the program will first execute a data loading module (for example, using the widely used pandas library in Python) configured to read the aforementioned spreadsheet file and load its contents into the program's working memory (for example, the DataFrame data structure). Subsequent algorithm steps will directly query and obtain the required configuration parameters from this memory data structure.

[0138] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A machine vision-based method for detecting defects in a patch inductor, the method comprising: The specific steps include: S1: obtaining first data representing internal heat conduction characteristics of the patch inductor; S2: obtaining second data representing surface microstructure and stress state of the patch inductor; S3: performing correlation processing on the spatially registered first data and second data to identify a spatial causal correspondence between the internal thermal anomaly area indicated by the first data and the surface stress anomaly area indicated by the second data, and to generate a defect physical cause diagnosis signal accordingly; S4: based on the defect physical cause diagnosis signal, generating a structured indication signal, the structured indication signal including at least a defect classification label and a quantitative risk index; S5: based on the structured indication signal, performing at least one of the following control operations: For a single patch inductor, based on its corresponding defect classification label and quantitative risk index, performing a multi-level sorting operation; Statistical analysis of the structured indication signal, when determining that the frequency of occurrence of any category in the defect classification label meets the preset condition, generating a process parameter adjustment instruction and sending it to the upstream production equipment of the patch inductor.

2. The method for detecting defects of patch inductors based on machine vision according to claim 1, characterized in that: S1 includes: applying a transient thermal excitation to the patch inductor; and, collecting and processing the temperature decay process of the patch inductor surface after the transient thermal excitation to generate the first data including at least a thermal relaxation time constant map.

3. The method of claim 2, wherein the method further comprises: S2 includes: irradiating the patch inductor with light of a predetermined polarization state; and, collecting and processing the scattered light of the patch inductor surface to generate the second data including at least a linear polarization degree map and a linear polarization angle map.

4. The method according to claim 3, wherein the method is characterized by: In S3, the correlation processing is realized by a pre-set defect physical cause inference engine, which is used to model the abnormal features extracted from the first data and second data to infer the physical causal relationship between the internal thermal anomaly area and the surface stress anomaly area.

5. The method of claim 4, wherein the method further comprises: The defect physical cause inference engine is a graph neural network model; the graph neural network model is configured to: take the thermal anomaly points in the first data and the stress anomaly points in the second data as nodes of a graph; construct edges of the graph based on spatial proximity between nodes and pre-set material mechanics constraints; and output a quantitative evaluation result of the spatial causal correspondence by propagating node information on the graph. The correlation processing of the defect physical cause inference engine includes:

6. The method of claim 5, wherein the method further comprises: a candidate association pair generation stage for screening out candidate anomaly pairs with spatial causal association from all thermal anomaly points and stress anomaly points based on spatial proximity; and a causal confidence evaluation stage, wherein the graph neural network model only constructs and propagates information for the candidate anomaly pairs, and finally forms a causal confidence score; The smaller the causal confidence score is, the weaker the defect physical cause inference engine judges the physical causal association between the internal thermal anomaly and the surface stress anomaly. The causal confidence evaluation stage further includes:

7. The method according to claim 6, wherein the method is characterized by: determining an internal defect severity measure for each thermal anomaly point based on the first data; ​ In the information propagation process of the graph neural network model, the information propagated from the thermal anomaly point node to the stress anomaly point node is modulated by a dynamic causal correlation weight; The dynamic causal correlation weight is dynamically generated by a preset nonlinear mapping function based on the internal defect severity metric of the thermal anomaly point node.

8. The method according to claim 7, wherein the method is characterized by: In S4, the defect classification label is selected from a preset label set, and the label set at least includes: "internal structural defect", "surface processing defect", and "endogenous and exogenous composite defect"; The quantified risk index is a normalized value representing the failure probability of the patch inductor under the preset working condition; The generation of the quantified risk index is realized by a two-stage hierarchical risk fusion model, which is specifically configured as: In the first stage, based on the defect classification label, a reference risk value representing the prior hazard level of the defect classification label category is determined from a preset defect risk reference library; In the second stage, based on the first data and the second data corresponding to the patch inductor, a dynamic severity metric representing the severity of the current physical state of the defect classification label is determined, and the dynamic severity metric is used to nonlinearly modulate the reference risk value to generate a final quantified risk index; the dynamic severity metric includes normalized surface stress severity and normalized internal defect severity.

9. The method of claim 8, wherein the method further comprises: The dynamic severity metric further includes a time dimension enhancement factor; The dynamic risk escalation module is configured to continuously monitor the change trend of the internal thermal anomaly feature indicated by the first data over time; When it is determined that the internal thermal anomaly feature presents a deterioration trend, the value of the time dimension enhancement factor is dynamically increased; The time dimension enhancement factor is used to amplify the modulation amplitude of the dynamic severity metric to the reference risk value in the nonlinear modulation process in the second stage; The dynamic severity metric further includes a spatial dimension correlation factor; The spatial consistency evaluation module is configured to calculate the spatial distance between the geometric center of the internal thermal anomaly region indicated by the first data and the geometric center of the surface stress anomaly region indicated by the second data; Based on the spatial distance, the value of the spatial dimension correlation factor is determined, wherein the smaller the spatial distance, the greater the value of the spatial dimension correlation factor; The spatial dimension correlation factor is used to further amplify the modulation amplitude of the dynamic severity metric to the reference risk value in the nonlinear modulation process in the second stage.

10. A machine vision based patch inductor defect detection system, characterized by: The system is used to perform the machine vision-based patch inductor defect detection method of any one of claims 1-9, comprising: A multi-physical field data acquisition unit is configured to acquire first data representing the internal thermal conduction characteristics of the patch inductor; Second data representing the surface microstructure and stress state of the patch inductor is acquired; a spatial causality correlation diagnosis unit configured to perform correlation processing on the first data and the second data that are spatially registered, to identify a spatial causality correspondence between an internal thermal anomaly region indicated by the first data and a surface stress anomaly region indicated by the second data, and to generate a defect physical cause diagnosis signal accordingly; a structured risk signal generation unit configured to generate a structured indication signal based on the defect physical cause diagnosis signal, the structured indication signal including at least a defect classification label and a quantified risk index; a closed-loop linkage control execution unit configured to perform at least one of the following control operations based on the structured indication signal: performing a multi-level sorting operation on a single patch inductor based on its corresponding defect classification label and quantified risk index; performing statistical analysis on the structured indication signal, and when determining that the frequency of occurrence of any one of the defect classification labels meets a preset condition, generating a process parameter adjustment instruction and sending it to an upstream production device of the patch inductor.

Images