A circuit board AOI detection result analysis method based on an intelligent algorithm

Abstract

The application discloses a circuit board AOI detection result analysis method based on an intelligent algorithm, and the core of the method is that high-precision synchronous collection, denoising normalization processing and time sequence alignment are performed on a detection image and multi-dimensional process parameters, a joint feature is extracted by using an attention mechanism, a process parameter-detection result correlation distribution model under each process scene is established, model output abnormal drift detection and traceability reasoning are realized, and the detection model and the process parameters are dynamically optimized in combination with an explainable AI, so that efficient tracking, attribution and self-adaptive optimization of detection misjudgment caused by process parameter variation are realized, and the stability of an SMT detection system production process is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic manufacturing quality detection and intelligent analysis, and particularly relates to a circuit board AOI detection result analysis method based on an intelligent algorithm. BACKGROUND

[0002] Currently, in the field of electronic manufacturing, especially in the field of printed circuit boards (PCB), automated optical inspection (AOI) has become a key quality control method in the manufacturing process of surface mount (SMT) and welding. Modern AOI systems generally use deep learning, feature extraction and multi-modal data analysis technologies to detect, distinguish and classify circuit board defects with high precision, in order to improve the intelligent and automated level of the manufacturing process. The mainstream detection and analysis scheme generally includes the collection and processing of detection images, the training and application of defect discrimination models, and the data statistics and attribution analysis for false positives and false negatives. In addition, in order to adapt to the changing manufacturing process and complex production environment, in recent years, the industry has gradually introduced process parameter driven auxiliary analysis, statistical process control and explainable AI methods to continuously improve the adaptability and robustness of AOI detection systems.

[0003] Currently, typical AOI intelligent analysis systems mainly rely on the following technical paths: first, high-precision detection image collection of key areas such as solder joints and components is realized based on high-resolution industrial cameras; second, deep learning or traditional machine learning methods are used to denoise, extract features and classify image data, and output qualified / defect judgment results; third, some systems introduce "process parameters + detection data" dual-dimensional modeling, and preliminarily correlate process setting conditions and detection performance indicators (such as false positives and false negatives) through statistical regression analysis, grouping comparison and other methods; fourth, in the traceability analysis stage, according to process batch, detection point and other information, typical false positive types and abnormal false negative phenomena are offline checked and simply corrected; fifth, in high-end applications, limited explainable AI algorithms are used to assist in analyzing error causes, or detection parameters are optimized through expert experience.

[0004] Although the above method realizes automatic detection and quality traceability of the manufacturing link to some extent, most AOI misjudgment and missed detection traceability analysis techniques mainly focus on static process conditions or single batch, single scene detection data modeling. In actual complex manufacturing process, SMT process parameters (such as welding temperature, humidity control, mounting precision, equipment setting value, etc.) often fluctuate in intervals, cycles or even bursts due to environmental disturbances, equipment aging or human error factors. Such dynamic process fluctuations will cause the coordinated drift of manufacturing process parameters and detection model input distribution, so that the existing static model and attribution algorithm have serious limitations in adapting to the changes of misjudgment and missed detection rate. Common problems include:

[0005] (1) The detection misjudgment / missed detection rate changes significantly nonlinearly with the fluctuation of process parameters, and the existing static model cannot timely discover and quantify the drift trend.

[0006] (2) When the process parameters are abnormal, the detection performance deterioration rules in different manufacturing areas / process scenes have spatial heterogeneity, and the existing "one-size-fits-all" attribution strategy cannot accurately locate the main factor.

[0007] (3) When the AOI detection result appears batch drift or continuous abnormality, the existing technology lacks a dynamic analysis and model self-adaptive correction mechanism driven by process parameters, which easily leads to long-term covering of misjudgment, missed detection and other quality problems, affecting the yield rate and control efficiency of the production line. SUMMARY

[0008] The present application provides a circuit board AOI detection result analysis method based on intelligent algorithm to solve the above technical problems.

[0009] The technical scheme of the present application is realized as follows: a circuit board AOI detection result analysis method based on intelligent algorithm, comprising:

[0010] S1: Collecting synchronous time sequence data sets including detection image data and multi-dimensional process parameters, wherein the process parameters include welding temperature, humidity, mounting precision and process setting value in the production cycle, as input data for subsequent feature processing.

[0011] S2: Denoising, normalizing and timestamp aligning the collected detection image data and process parameter time sequence data respectively to ensure the consistency of multi-modal feature data in time and dimension.

[0012] S3: Based on the normalized detection image data and process parameter time sequence data, using feature alignment and attention mechanism to extract joint feature vectors reflecting the dynamic changes of the process, so as to take into account the multi-modal characteristics of process and detection information.

[0013] S4: Establishing a process parameter-detection result correlation distribution model for multiple detection point positions and process procedure labels, and grouping the sub-models to correspond to different manufacturing areas and process scenarios respectively.

[0014] S5: Based on the joint feature vector, inputting the process parameter-detection result correlation distribution model into a dynamic drift detection algorithm, calculating the false positive rate and the false negative rate statistics of the current batch and the historical batch, and identifying the abnormal drift of the model output caused by the process parameter disturbance.

[0015] S6: When detecting the abnormal drift of the model output and correlating to a specific process parameter interval, automatically triggering the retraining or incremental learning process of the analysis model to adapt to the detection state under the specific process parameter.

[0016] S7: Using an interpretable AI algorithm to perform causal reasoning on the process parameter variation history, joint feature weight and model drift influence, generating a dominant process factor ranking list of the false positive rate and false negative rate changes of each batch to indicate the traceability path.

[0017] S8: According to the dominant process factor ranking list, dynamically adjusting the parameter structure of the analysis model, including using targeted sub-models or changing loss function weights when switching process parameter intervals, to realize adaptive optimization of the analysis model under multiple process scenarios.

[0018] S9: Real-time feedback of dynamic false positive traceability analysis results and process parameter tracking records to the manufacturing execution system, closed-loop implementation of process optimization suggestions and detection parameter configuration correction.

[0019] Beneficial effects

[0020] The circuit board AOI detection result analysis method based on intelligent algorithm provided by the application has the following beneficial effects:

[0021] (1) The application innovatively introduces multi-channel synchronous acquisition and time sequence process parameter coding, realizes high-precision alignment of detection image data and process parameters, and provides a solid data foundation for feature fusion and correlation modeling. Based on feature alignment, multi-modal feature extraction of convolutional neural network and time sequence recurrent network, and cross-modal attention fusion mechanism, the influence of process parameter fluctuation on detection discrimination is fully tapped, and the sensitivity and adaptability of the model to dynamic process variation are significantly enhanced. Multi-batch testing shows that compared with traditional single model, the drift sensitive detection of false positive rate is improved by more than 20%, the drift response of false negative rate is improved by more than 15%, and the stable output of traceability analysis in complex scenarios such as process adjustment and high variation batch is ensured.

[0022] (2) This invention constructs a dynamic drift detection algorithm, which, based on sliding window flow statistics of false positive and false negative rates and drift anomaly discrimination, can identify anomalies in the analysis model output caused by process disturbances in real time. Combining process parameter interval mapping with an automatically triggered incremental learning process, the model achieves real-time adaptive correction under abnormal process intervals. Through targeted sub-model switching and dynamic adjustment of loss function weights, the system can consistently output optimal detection and analysis capabilities during periods of process environment change. Actual engineering verification shows that in batches with sudden changes in process parameters, the convergence time of the analysis model parameters is shortened by 60%, and the peak false positive rate decreases by more than 30%, greatly improving the applicability and timeliness of the system on dynamic production lines.

[0023] (3) This invention automatically outputs the ranking of dominant process factors and the causal chain for each abnormal drift event through interpretable AI modules such as causal reasoning and SHAP / LIME, providing clear indications to process engineers and production decision-making systems of the process steps that need adjustment. Utilizing the historical weighting of process parameters and flexible causal chain topology, it is compatible with multi-process, multi-detection point, and complex batch grouping scenarios, and the analysis process has extremely high scalability and traceability. On the pilot manufacturing line, the accuracy of identifying dominant process parameters has been greatly improved, and batch-level automatic causal tracing and optimization suggestions can be pushed under multi-scenario mixed-line production conditions, greatly facilitating process management and rapid problem localization.

[0024] (4) This invention pushes the analysis results and process parameter tracking to the Manufacturing Execution System (MES) in real time, seamlessly completing a fully automated closed loop of "anomaly detection - cause tracing - process optimization suggestions - execution feedback". The system can correct production line parameter configurations in real time without human intervention, reducing manual inspection costs and shortening the response chain. Attached Figure Description

[0025] Appendix Figure 1 This is the main flowchart of a circuit board AOI inspection result analysis method based on intelligent algorithms;

[0026] Appendix Figure 2 This is a sub-flowchart of a method for analyzing AOI inspection results of circuit boards based on intelligent algorithms;

[0027] Appendix Figure 3 This is another sub-flowchart of a method for analyzing AOI inspection results of circuit boards based on intelligent algorithms. Detailed Implementation

[0028] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0029] The preferred embodiments of the present application will be described below with reference to the accompanying drawings. Those skilled in the art will appreciate that the embodiments are only used to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0030] As used herein, the singular forms "a", "an" and "the" include plural referents unless the context clearly dictates otherwise. It will be further understood that the terms "comprises", "comprising", "includes" and / or "including", or the like, when used in this specification, specify the presence of stated features, integers, steps, operations, components, parts, or combinations thereof, but do not preclude the presence or addition of one or more other features, integers, steps, operations, components, parts, or combinations thereof.

[0031] Please refer to Figure 1 As shown in the drawings, a circuit board AOI detection result analysis method based on an intelligent algorithm comprises:

[0032] S1: Collecting synchronous time sequence data sets including detection image data and multi-dimensional process parameters, wherein the process parameters include welding temperature, humidity, mounting accuracy and process setting values in the production cycle, serving as input data for subsequent feature processing.

[0033] S2: Respective denoising, normalization and timestamp alignment processing are performed on the collected detection image data and process parameter time sequence data, so as to ensure the consistency of multi-modal feature data in time and dimension.

[0034] S3: Based on the normalized detection image data and process parameter time sequence data, a joint feature vector reflecting process dynamic changes is extracted by using feature alignment and attention mechanism, so as to take into account the multi-modal characteristics of process and detection information.

[0035] S4: For multiple detection point positions and process procedure labels, a process parameter-detection result correlation distribution model is established, and sub-models are grouped to correspond to different manufacturing areas and process scenarios respectively.

[0036] S5: Based on the joint feature vector, the process parameter-detection result correlation distribution model is input into a dynamic drift detection algorithm, the misjudgment rate and the missed detection rate statistics of the current batch and the historical batch are calculated, and the model output abnormal drift caused by process parameter disturbance is identified.

[0037] S6: When the model output abnormal drift is detected and is related to a specific process parameter interval, the retraining or incremental learning process of the analysis model is automatically triggered to adapt to the detection state under the specific process parameter.

[0038] S7: Adopting the explainable AI algorithm to perform causal reasoning on the process parameter variation history, joint feature weight and model drift influence, and generate a dominant process factor ranking list of each batch misjudgment rate and missed detection rate change to indicate the traceability path.

[0039] S8: According to the dominant process factor ranking list, dynamically adjust the parameter structure of the analysis model, specifically including adopting a targeted sub-model or changing the loss function weight when switching the process parameter interval, to realize the adaptive optimization of the analysis model under multiple process scenarios.

[0040] S9: Real-time feedback of dynamic misjudgment traceability analysis results and process parameter tracking records to the manufacturing execution system, closed-loop realization of process optimization suggestions and detection parameter configuration correction.

[0041] The step S1: Collecting synchronous time series data sets including detection image data and multi-dimensional process parameters, wherein the process parameters include welding temperature, humidity, mounting precision and process setting value in the production cycle, as input data for subsequent feature processing. Specifically including:

[0042] S1.1: Configure the AOI detection equipment of the circuit board to be detected, so that each circuit board in the detection batch will have its original detection image data acquired through a high-resolution image acquisition interface when it is detected, and stored in a standardized format as image data to ensure the uniformity of subsequent image feature preprocessing.

[0043] The AOI detection equipment for the circuit board to be detected is parameterized configured, the detection resolution, lens parameters and light source conditions are set, the standardization of high-precision image data capture environment is realized. A high-resolution image acquisition interface is adopted, the circuit board to be detected is triggered and shot at each specified station of the detection conveying belt by an industrial camera, the original detection image reflecting the microstructure details of the welding points, mounting and the like is acquired, and the non-overlapping and non-missing detection of single board image acquisition is ensured. Further, through a standardized image data format conversion module (the parameter is set to 24bit RGB three channels or 8bit gray scale, which is switched according to the detection requirement), the format uniform storage of the obtained original detection image is realized, and lossless files such as TIFF, PNG and the like are generated for comprehensive feature extraction. The unique detection batch and the board identification (such as two-dimensional code or RFID physical unique code) are used for metadata registration, each image is logically associated with the batch number, the board number, the detection station number and the time stamp, and a complete metadata chain covering the whole detection process is established. Based on the image data standardized storage module, a redundancy check algorithm is called to detect the damage of the generated image file and check the storage consistency, and an abnormal marker signal is output for subsequent resampling; for the image data passing the check, the image data is written into a structured database or a special file storage system, is batch-archived, and an image data set with retrieval, efficient reading and writing ability is formed. Through a series of standardized acquisition and storage processing modes, the original detection image of each circuit board in the detection batch process is converted into a structured, archived and retrievable standard input, which provides a unified and data quality guaranteed image input basis for the downstream multi-modal feature extraction and process parameter synchronization.

[0044] Exemplary, a high-resolution industrial camera of Basler acA1920-155um is configured on a SMT production line, with a shooting resolution of 1920x1200 and a frame rate of 30fps, and is matched with a ring-shaped high-brightness LED light source, with an exposure time of 20ms and an aperture F value of 3.5. In the detection batch, a photoelectric sensor is used to trigger the camera to collect every time a board passes through the conveyor belt, and a trigger protocol GigE Vision is used to transmit image data in real time. The image acquisition module automatically converts the 1920x1200 24bit raw image into PNG format, and associates the 2D Data Matrix board code and batch number obtained by the code gun to generate a file name rule such as batch20240416_board0001.png. The supporting database realizes one-to-one registration and indexing of images and metadata with the help of SQL Server. After the collection is completed, the file MD5 checksum and breakpoint retransmission mechanism are used for integrity check, and the broken frame, damaged or lost files are automatically supplemented, ensuring that the group data is not missed. After this link processing, the detection task of batch batch20240416 realizes standardized image acquisition and batch archiving, and the single-board image resolution, file integrity and metadata retrieval accuracy are all 100%, which establishes a unified and reliable data starting point for subsequent process parameter synchronization and feature depth extraction.

[0045] S1.2: Configure multi-channel sensor nodes for the production line process parameter acquisition unit, and collect real-time process parameters (including welding temperature, production environment humidity, mounting process precision and process setting value, etc.) synchronized with the detection batch. Through the continuous sampling algorithm of multi-dimensional sensor data stream, the process parameters are stored in a structured manner on the detection time axis, providing high-frequency process parameter raw data stream for subsequent time alignment processing.

[0046] The production line process parameter acquisition unit is deployed and configured with multi-channel sensor nodes, covering process parameter sensing channels, including welding temperature sensing channels, production environment humidity sensing channels, mounting process precision acquisition channels, and process setting value reading channels.

[0047] High-precision distributed temperature measurement sensors (such as PT100 platinum resistance thermometers with a sampling frequency of 1Hz), industrial-grade digital humidity sensors (such as SHT35 with a sampling frequency of 0.5Hz), visual / laser displacement mounting precision monitoring equipment (resolution 0.01mm, sampling period synchronized with mounting cycle), and process controller data reading interface (such as ModbusRTU protocol, real-time pulling of current process setting parameters) are used to ensure the acquisition of multi-dimensional process parameters synchronized with the detection time axis of each circuit board.

[0048] The multi-channel data synchronous acquisition control algorithm (parameter setting: each channel clock alignment error less than 10ms) is used to schedule the sensing data of each sensing node in real time, generate data sampling package synchronized with AOI detection trigger signal, and record the current sampling batch number and board number with timestamp.

[0049] Further, through continuous sampling data stream encoding and buffer management algorithm, the process parameter sample stream at each moment is associated and structured according to the multi-level index format of detection batch-board number-sensing channel-timestamp, generating high-frequency, multi-dimensional process parameter raw data sequence.

[0050] The data consistency and missing monitoring algorithm is used to periodically detect the integrity of each channel process parameter stream. For the data section with packet loss, interpolation estimation or repeated sampling strategy is used for compensation to ensure complete and multi-channel synchronous process parameter data points at each detection moment.

[0051] Through the above multi-channel sensing node real-time cooperation, continuous sampling and structured storage, the high-frequency process parameter raw data stream input corresponding to the detection batch and board time axis is realized, which provides data basis for subsequent time sequence alignment, feature fusion and error trace analysis.

[0052] For example, in a typical SMT production line, for an AOI detection batch batch20240416, a Siemens S7-1200 PLC control system is selected, and the Modbus RTU protocol is used to collect the temperature of the welding station PT100 temperature sensor (specifically, the surface temperature of the welding head is collected every 1 second, with a temperature accuracy of ±0.1℃), and the SHT35 humidity sensor is arranged at the key node of the production area, and the relative humidity reading is obtained every 2 seconds, with a monitoring accuracy of ±1.5%RH. The mounting accuracy is captured by the laser displacement sensor based on Keyence LJ-V7000 series immediately after each action at the mounting point, and the position drift value is buffered in the local database through PLC communication. The process set value is issued through the Siemens HMI industrial touch screen interface, and the data reading frequency is set to automatically collect within 0.5 seconds after each process parameter change. All sensing channels are aligned with the timestamp generated by the AOI detection station's scan code + photographing moment, and the clock synchronization accuracy error is controlled within 5ms.

[0053] During the detection process, the acquisition controller and the AOI detection master PLC jointly schedule, and each board detection is completed in the data cache area to generate a structured record containing [board number, batch number, detection time, welding temperature, humidity value, mounting precision, process setting parameter]. For single sensor instantaneous packet loss, linear interpolation is used to correct single point anomaly. Finally, the number of raw process parameter data sets collected in a batch corresponds to the total number of board detection, and the missing rate is less than 0.01%. The high-frequency and high-integrity multi-channel process parameter raw data stream lays a high-quality data condition for downstream process parameter and AOI detection image synchronization, joint multi-modal feature extraction, and batch misjudgment rate traceability analysis.

[0054] S1.3: Based on the AOI detection system master control platform, the time stamp of the AOI detection image and the acquisition time stamp of the process parameter are high-precision synchronous coded by using a unified clock synchronization protocol, a unique time sequence coding identifier is generated for each detection period, and the strict correspondence between the detection image data and the process parameter data in the acquisition batch is realized, so as to accurately input synchronous data for the feature fusion module.

[0055] S1.4: The detection image data and the process parameter data set obtained by synchronous acquisition are numbered and registered with metadata, and the acquisition batch information and the process parameter association table are written into the database in real time through the manufacturing execution system (MES) interface, so that each data set has a unique attribution and traceability attribute in the subsequent processing link, and supports subsequent process parameter driven misjudgment traceability analysis data tracking.

[0056] S1.5: After data acquisition is completed, data integrity verification and packet loss detection algorithm are performed, and for damaged or missing detection image data and process parameter data, a pre-defined retry mechanism is used for data re-acquisition or abnormal marking, so as to ensure the quality of the raw data set used for feature extraction and model training, and form a high-reliability multi-modal time sequence data input.

[0057] The step S2: the acquired detection image data and process parameter time sequence data are respectively subjected to denoising, normalization and time stamp alignment processing, so as to ensure the consistency of multi-modal feature data in time and dimension. Specifically, it includes:

[0058] S2.1: The detection image data is subjected to spatial domain denoising processing, and based on the median filter, bilateral filter or convolutional neural network adaptive denoising algorithm, the image noise interference is removed, and the detection image data after noise suppression is obtained, laying a pure data foundation for subsequent feature normalization.

[0059] The collected detection image data is subjected to spatial domain denoising, and the input is the original detection image data set obtained through a high-resolution industrial camera and a standardized acquisition process. A median filter algorithm (parameters: filter window size 3x3 or 5x5) is used to perform pixel-level spatial median filtering on each detection image, effectively eliminating isolated abnormal points such as salt and pepper noise, and preserving structural edge details. Further, through a bilateral filter algorithm (parameters: spatial domain Gaussian kernel σ s, pixel color difference Gaussian kernel σ r), the image edge information is maintained while the noise is reduced, the Gaussian-type and mean shift-type noise distribution is suppressed, and a spatial smoothing result with stronger edge protection is obtained. The weighted average of the image pixels is calculated using the following formula:

[0060]

[0061] where I out (x) represents the pixel value of the output image, I(x) and I(y) are the pixel values of the original image positions x and y in the neighborhood, Ω is the neighborhood window, σ s is the spatial distance standard deviation, σ r is the pixel difference weighted standard deviation, W p is the normalized weight. Further, based on the convolutional neural network (CNN) adaptive denoising algorithm (such as DnCNN, U-Net structure, parameters: network depth, training loss function using mean square error), the image with known noise distribution is used as input, and the discriminant features of spatial noise and image content are learned through end-to-end network, to realize high adaptive denoising of composite noise. In the CNN denoising process, the following loss function is used to train the parameters:

[0062]

[0063] where L MSE is the mean square error loss, is the i-th pixel of the denoised output image, is the i-th pixel of the noise-free reference image, and N is the total number of pixels. After the cascade processing of the above-mentioned physical model-based filtering method and data-driven neural network method, various types of spatial noise interference are effectively eliminated, and the noise-reduced detection image data with fully restored content and edge details is obtained. The peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the denoised image are used to quantitatively evaluate the denoising quality, to ensure that the denoised detection image data meets the technical requirements of high-purity input to the subsequent normalization and feature extraction stages.

[0064] Exemplarily, in a SMT patch production line, the spatial domain denoising processing is applied to the circuit board detection image with a collection resolution of 1920x1200 and a 24-bit RGB format. For the ADC collection noise and random laser point interference in the environment, the median filter window size is set to 5x5, the bilateral filter spatial standard deviation σ_s is 2, and the pixel standard deviation σ_r is 15. The median filter is first performed on the original image channels in parallel, and then the bilateral filter is used to further suppress the spatially distributed uneven noise. For the local area with complex high-frequency noise, a pre-trained DnCNN neural network model is used, the training data set contains 40,000 real industrial defect images, the network layer is 17 layers, the Adam optimizer and MSE loss are used, and the iteration round is 100,000 times. After denoising, the image PSNR is improved to 38.5 dB, and the SSIM is 0.982. Compared with the image without denoising, the defect edge is complete, the background noise is almost eliminated, the subsequent normalization, feature extraction and other steps are ensured to be high-purity and interference-free detection image, and the AOI detection link is realized. The false rejection rate is reduced and the robustness is improved.

[0065] S2.2: Adaptive outlier detection and interpolation correction algorithm is performed on the process parameter time series data (including welding temperature, humidity, mounting precision and process setting value, etc.), and sliding window statistics or multivariate time series consistency checking mechanism is used to automatically remove or correct abnormal points, and the abnormal corrected process parameter time series data is output.

[0066] S2.3: Based on the noise suppressed detection image data obtained in S2.1, linear normalization or Z-Score standardization algorithm is used to uniformly scale the pixel values of each channel to a preset interval, so as to eliminate the inconsistency of the detection image data in dimension, and the normalized detection image data with uniform scale is obtained.

[0067] S2.4: For the abnormal corrected process parameter time series data output by S2.2, the minimum-maximum normalization or standard score normalization method is used to uniformly process different physical quantities (such as temperature, humidity, position, etc.), and the process parameter data is transformed into dimensionless standardized parameter sequence to adapt to the subsequent feature space fusion requirements.

[0068] S2.5: According to the time stamp recorded in the detection period, the normalized detection image data of S2.3 and the standardized process parameter time series data of S2.4 are executed time series synchronization alignment processing, linear interpolation registration or dynamic time warping (DTW) algorithm is used, and according to the common detection batch index, the time stamp aligned multi-modal feature input group is output, the accurate one-to-one correspondence of process conditions and detection results at each time is realized, and the strictly time series consistent data basis is provided for the multi-modal feature extraction stage.

[0069] The step S3: based on the normalized detection image data and the process parameter time series data, a joint feature vector reflecting the process dynamic change is extracted by using feature alignment and attention mechanism, so as to take into account the multi-modal characteristics of process and detection information. As shown in Figure 2 Specifically, it comprises:

[0070] S3.1: The feature dimension calibration processing is performed on the normalized detection image data and the process parameter time series data, the standard feature alignment algorithm is adopted to one-by-one map the feature dimensions, sampling intervals and time sequence labels of different data modalities, so as to obtain a feature alignment matrix, and to provide an input basis for the consistency processing of the multi-modal feature space.

[0071] S3.2: Based on the feature alignment matrix, the deep representation processing of the feature encoder is performed, the convolutional neural network (CNN) is respectively adopted to extract the spatial features of the detection image data, and the time sequence recurrent network (RNN / LSTM) is adopted to perform dynamic state modeling on the process parameter time series data, so as to obtain the detection spatial feature vector and the process time sequence feature vector.

[0072] The normalized detection image data and the standardized process parameter time series data subjected to the feature dimension calibration processing are inputted, and the feature alignment matrix is used as an index to respectively indicate the feature mapping and the sampling time sequence relationship of different modalities.

[0073] The convolutional neural network (CNN) (parameters: network layer number 16 layers, convolution kernel size 3x3, activation function ReLU, maximum pooling kernel 2x2, batch normalization, Dropout ratio 0.3) is adopted to realize the automatic extraction of high-dimensional spatial features of the detection image data. The input is the multi-channel detection image subjected to normalization, the edge, texture, structure and defect region spatial features are gradually extracted through multi-level convolution and pooling operations, and the output is the high-dimensional feature vector after the spatial feature flattening.

[0074] Further, the time sequence recurrent neural network (RNN) or long short-term memory network (LSTM) (parameters: one-way LSTM, hidden layer dimension 64, input sequence length consistent with the sampling period, activation function tanh, dropout rate 0.2) is adopted to dynamically model the calibrated process parameter time series data. The input is the process parameter sequence in each sampling period after standardization, which is gradually inputted into the LSTM unit after embedding mapping, continuously captures and stores the time sequence dependent features of parameter changes, and forms the time sequence state feature vector reflecting the process running dynamics.

[0075] The feature alignment matrix is adopted to associate the spatial features outputted by the CNN and the time sequence features outputted by the LSTM, so as to ensure that the feature dimensions at the same detection period or sampling time are one-to-one corresponding, and to eliminate the modal confusion error.

[0076] By feature normalization processing, the spatial features and the time sequence features are uniformly standardized to the same numerical distribution range, facilitating subsequent multi-modal feature fusion and attention weight distribution.

[0077] Through the above algorithm implementation, high-dimensional detection spatial feature vectors and process time sequence feature vectors are respectively obtained, providing an analysis basis for subsequent fusion analysis and improving process sensitivity and defect classification discrimination accuracy.

[0078] For example, in the circuit board AOI detection scene, for a normalized detection image with a resolution of 256x256 and 3 channels, a 16-layer ResNet type CNN network is deployed, the convolution kernel step is 1, and the feature vector length after maximum pooling is 1024. For the standardized time sequence data of the process parameters (soldering temperature, humidity and mounting precision) with a sampling period of 10ms, a one-way LSTM network is input, the number of hidden units is 64, and the final aggregated time sequence feature vector is output. Cross-validation shows that the discrimination accuracy of the spatial feature vector is 92.5%, the AUC of the LSTM time sequence feature for process anomaly detection is improved to 0.97, and the two complement each other in the multi-modal mapping of the feature alignment matrix, the total feature dimension is 1088, which is normalized and ready for Attention attention fusion processing. The verification result shows that the model sensitivity to process parameter changes is improved by 22% using the above feature encoding scheme, and the defect misjudgment and missed detection analysis shows better adaptive ability.

[0079] S3.3: Input the detection spatial feature vector and the process time sequence feature vector obtained in the previous step into the multi-modal fusion module, call the feature cross algorithm based on the attention mechanism (Attention), calculate the correlation weight of different modal features in each time sequence segment, and obtain the cross-modal representative fusion feature vector according to the weight adaptive aggregation to enhance the process sensitivity.

[0080] The input is the normalized detection spatial feature vector and the process time sequence feature vector, and the two groups of features have realized strict one-to-one correspondence of time sequence labels and dimensions through the aforementioned feature alignment matrix.

[0081] The multi-modal fusion module (parameters: fusion layer type is cross Attention, number of layers is 2-4 layers, hidden dimension is 512, and number of Attention heads is 8) is used to cascade the spatial features and the time sequence features, construct a feature interaction input tensor, and realize coupling fusion of different modal information.

[0082] Further, the input detection space feature vector Q and process timing feature vector K, V are calculated by an attention correlation weight based on an attention mechanism (Attention) cross feature algorithm (parameters: query / key / value linear transformation matrix, softmax activation, residual connection, normalization).

[0083]

[0084] wherein Q is a query matrix, K is a key matrix, V is a value matrix, d k is a key vector dimension, realizing the correlation score of space-time features under different timing segments.

[0085] Further, the process timing feature vector and the detection space feature vector are adaptively weighted in the fusion layer by the above correlation weight, generating a cross-modal representative fusion feature vector. The vector contains cross-modal feature weights allocated by Attention under each timing segment, which can reflect the modulation of the current process disturbance on the detection feature representation ability.

[0086] Further, the fusion feature output is extracted by a multi-head Attention mechanism (parameters: 8 heads, 64 hidden dimensions per head) in parallel to extract the correlation information of different subspaces, and a final multi-modal joint feature tensor is generated. This tensor has high sensitivity to process and image feature changes.

[0087] Through feature aggregation and post-processing, the above fusion feature tensor is processed by a fully connected layer and layer normalization to output a cross-modal representative fusion feature vector with low redundancy and high process sensitivity, providing a key input for subsequent anomaly suppression and process drift quantification analysis.

[0088] Through the cross feature algorithm based on the attention mechanism (Attention), the original data of space and timing features are efficiently converted into fusion expression, realizing the adaptive enhancement of joint feature sensitivity under the background of process disturbance, and improving the recognition ability of the analysis model to misjudgment / misjudgment caused by dynamic process changes.

[0089] For example, in the SMT mounting circuit board AOI defect detection application, a convolution residual network is selected to output a 1024-dimensional image space feature vector and a long short-term memory network to output a 64-dimensional process parameter timing feature vector, which is fused into an Attention module. The Attention parameters are set to 3 layers, 512 hidden dimensions, and 8 multi-head Attention heads. The query matrix Q comes from the space feature, the key / value matrix K, V comes from the process timing feature, and the weighted output is calculated by the following formula in the fusion stage:

[0090]

[0091] In actual execution, under the periodic fluctuation of welding temperature and the change of mounting precision batch, the Attention algorithm gives high weight to the process parameter feature dimension, the output fusion feature tensor dimension is 512, and the feature contribution degree analysis shows that the fusion feature sensitivity to misjudgment rate is improved by 23%, and the sensitivity to missed detection rate is improved by 19%. The output cross-modal representative fusion feature vector is used by the backend model for process dynamic drift detection and misjudgment rate traceability, which significantly improves the response ability to complex process disturbance and the explanation of misjudgment / missed detection associated features. The AUC score of the final model joint feature reaches 0.977, and the attention weighted adaptive feature fusion effectively suppresses the detection sensitivity decline caused by process parameter disturbance in the actual batch data of the production line, realizing the high robustness analysis output of process-detection dual modal under multiple scenarios.

[0092] S3.4: Perform feature anomaly suppression and time series regularization processing on the cross-modal representative fusion feature vector, suppress the noise distribution of abnormal fusion features by introducing a specific regularization loss function, and maintain the smoothness of the feature change with the process disturbance. Finally, output the joint feature vector that meets the process dynamic distribution characteristics, providing input for subsequent associated distribution modeling and dynamic drift detection quantification.

[0093] The input is the cross-modal representative feature vector generated by the Attention cross mechanism fusion, and the feature dimension includes the numerical components of the detection space features and the process time series features aggregated under each time series segment.

[0094] The feature anomaly detection algorithm (parameters: abnormal value discrimination based on statistical distribution, threshold set to 3 times the standard deviation, abnormal distribution model is Gaussian Mixture Model GMM) is used to detect and label the abnormal feature components in the fusion feature vector that are distributed extremely and deviate from the overall trend.

[0095] Further, through feature anomaly suppression processing (parameters: threshold abnormal value replacement algorithm, replacement method uses local weighted average or interpolation correction), the identified abnormal features are replaced or suppressed softly without loss, and the fusion feature set after effective suppression of abnormal noise is obtained.

[0096] Further, the feature time series smoothing and regularization method (parameters: sliding window length is 3-5 times the corresponding detection period, smoothing strategy is first-order exponential weighted moving average or bidirectional Savitzky-Golay polynomial filtering) is used to realize the consistency constraint of the feature vector after anomaly suppression in the time series dimension, improve the smoothness of feature change, and suppress the inevitable time series mutation caused by process disturbance.

[0097] Further, a time regularization penalty term is introduced to construct a joint loss function, and the joint optimization of abnormal distribution constraint and time series smoothness is realized in the synthetic feature vector space through regularization algorithm. The joint loss function is defined as follows:

[0098] L total = L task + λ1L anom + λ2L smooth

[0099] Wherein, L task is the misjudgment / missed detection discriminant main loss, L anom is the feature anomaly suppression loss (such as weighted absolute deviation), L smooth is the time series smoothness regularization term (such as first-order difference sum of squares), λ1, λ2 are regularization weight parameters.

[0100] Through the loss function optimization, the feature distribution output that meets the abnormal feature noise suppression and adapts to the process dynamic disturbance is obtained.

[0101] Through the above algorithm processing, the high-dimensional representative feature vector obtained after Attention fusion is converted into a joint feature vector after abnormality suppression and time series regularization, realizing the smoothness constraint of feature distribution facing process dynamic change, and inputting high-confidence, multi-scene adaptive process-detection joint feature data for subsequent correlation distribution modeling and dynamic drift detection steps.

[0102] For example, in the AOI circuit board detection scene, the input is the 512-dimensional fusion feature output by the Attention module, and the Batch sampling batch length is 200. About 2.6% of the feature components are detected as abnormal noise by the Gaussian mixture model (the number of components is set to 3, and the discrimination threshold is 3σ), and the abnormal points are replaced by local interpolation weighted correction (sampling window 5). The feature sequence after abnormality suppression is smoothed by using sliding window exponential weight average (window width 20). The loss function weight is set to λ1=0.4, λ2=0.6, and the end-to-end training is realized by using Adam optimizer. The results show that the root mean square of feature time series mutation is reduced by about 57%, the abnormal distribution probability is reduced to 0.5%, and the joint feature smoothness is significantly improved. Under the condition of periodic disturbance of multi-batch process parameters, the feature vector processed by this step is input into the rear correlation modeling, the misjudgment detection AUC is improved by 2.9%, and the missed detection rate is reduced by 4.2%, effectively guaranteeing the stability and robustness output of the subsequent model in the dynamic process environment.

[0103] S3.5: Perform feature importance measurement and redundant feature screening on the output joint feature vector, select the most influential process-related sub-feature set for misjudgment and missed detection discrimination by using feature contribution analysis algorithms (such as Shapley value or information gain evaluation), and ensure that the final joint feature vector has high discriminability and high interpretability, providing accurate feature basis for the robustness and interpretability optimization of subsequent process parameter-detection result modeling.

[0104] The step S4: For multiple detection point positions and process procedure labels, a process parameter-detection result correlation distribution model is established, and the sub-models are grouped to correspond to different manufacturing areas and process scenarios. As shown in Figure 3

[0105] S4.1: Based on the joint feature vector obtained in the S3 step, the input data is aggregated according to the detection point physical position and the process procedure label, and a spatial distribution mapping method is used to form a grouped feature set for each detection point and its corresponding process procedure label, to support the subsequent point-by-point causal analysis between process parameters and detection outputs.

[0106] ​The input is the multimodal joint feature vector set output from step S3. This feature vector encompasses a high-dimensional fusion representation of detection spatial features and process temporal features, possessing normalization and temporal alignment attributes. A physical location index mapping algorithm (parameters: based on the actual detection point layout coordinate system on the PCB board, coordinate accuracy 0.01mm) is used to map and associate each detection point ID with its corresponding joint feature vector, ensuring the unique spatial label of the feature data. Furthermore, a process step label aggregation algorithm (parameters: the process step dictionary includes process number, section type, equipment identifier, etc., and label types cover multiple levels of processes such as solder paste printing, surface mount technology, and reflow soldering) is used to group the joint feature vectors of multiple detection points under the same process step label, achieving data segmentation based on the process flow dimension. Further, a spatial-process two-layer clustering method (parameters: spatial partition radius 3-5mm, process label similarity threshold set to 0.8) is used to pre-cluster the joint feature vector sets that are physically adjacent and have consistent process labels, generating multiple grouped feature sets to provide combined structural support for subsequent spatial distribution mapping. Furthermore, a spatial distribution mapping model (parameters: Gaussian radial basis function kernel, spatial distribution estimation grid resolution 1mm) is invoked to generate spatial distribution mapping relationships for each group feature set according to the spatial coordinate distribution of the detection surface, realizing the corresponding mapping of detection points in the physical block dimension. Further, for each spatial block and process step group, the group feature set structure is recorded and output, including detection point ID, process step label, physical coordinate interval, process label category, and its joint feature vector list, providing high-resolution input for subsequent fixed-point causal correlation analysis. Through a two-layer processing of spatial location index mapping and process step label aggregation, multimodal joint features are efficiently organized into well-distributed group feature sets, realizing spatial-process multidimensional correlation between process parameters and detection output under multiple detection points, multiple processes, and multiple manufacturing areas, providing an accurate foundation for subsequent process causal fixed-point analysis, region modeling, and dynamic drift detection.

[0107] Exemplarily, in the AOI detection line of the SMT mounting process, 32 detection points are set on the PCB, and the positions of the detection points are imported through CAD coordinates, with a spatial coordinate resolution of 0.01 mm. Each detection point is attached with a process procedure label in the detection batch, covering procedures such as patching, reflow soldering, and wave soldering. In actual operation, the batch to which the detection data belongs covers 2 hours, and the data amount is 12800 groups. Using a spatial indexing algorithm, detection points within 3 mm of each other on the same PCB are grouped. Using the process procedure label matching, the detection points with the “reflow soldering” label are further merged into the “reflow soldering group”. After spatial and process clustering, 12 group feature sets are formed, with a capacity of 700-1200 group joint feature vectors per group. A Gaussian radial basis distribution model is used to map the spatial distribution of each group feature block, and the spatial distribution map clearly reflects the high and low incidence areas of defects in the “reflow soldering” procedure. The output group feature set enables efficient retrieval of each process procedure-detection point block in the database. The actual effect shows that this grouping method provides a fixed-point basis for subsequent causal drift modeling, significantly improves the response capability to regional process abnormalities, and increases the spatial aggregation efficiency by 37%, with a group feature coverage rate of 100%.

[0108] S4.2: Use a multivariate statistical modeling algorithm (such as Gaussian mixture model or variational Bayesian inference) to model the joint probability distribution of process parameters and detection results in each group feature set, obtain the process parameter-detection result association distribution model, and realize the preliminary quantitative description of the association between detection points and process scenarios.

[0109] The input is the joint feature vector set obtained by grouping in step S4.1, each group covering the physical location of the detection point, the process procedure label, and the normalized fused multi-modal feature information.

[0110] A multivariate statistical modeling algorithm (parameters: Gaussian mixture model GMM, number of components dynamically adaptive [2, 6], covariance type full, maximum iteration number 500, or variational Bayesian inference model, prior hyperparameter automatic calibration) is used to model the joint probability distribution of process parameters and detection outputs in each group feature set.

[0111] Further, by using the GMM modeling method, the process parameter sub-vector X proc and the detection result sub-vector Y det are input, the joint distribution probability density function between the two is fitted, and the model is as follows:

[0112]

[0113] where K is the number of mixed components, π k is the weight of the kth component, is the mean μk , covariance ∑ k of multivariate normal distribution.

[0114] Further, the EM (Expectation-Maximization) algorithm is used for parameter estimation, the E step calculates the posterior probability, and the M step updates the mean μ k , covariance ∑ k , weight π k , and iteratively optimize the log-likelihood to convergence or reach the maximum iteration number.

[0115] Further, for high-dimensional features and sample capacity, the variational Bayesian inference algorithm is called to adaptively regularize the Gaussian prior parameters and automatically select the optimal number of mixed components to prevent overfitting and improve the model generalization ability.

[0116] Through the above joint probability distribution modeling, the joint distribution model of process parameters-detection results is output for each grouped feature set, and the model parameters include component weight, grouped mean, covariance matrix, etc.

[0117] Further, through conditional probability inference of the joint distribution, the quantitative evaluation of the detection result abnormality (misjudgment / missed detection) probability under the specific value of the process parameter (such as temperature, humidity, mounting accuracy, etc.) can be obtained.

[0118] Through the joint probability modeling of multiple grouped feature sets, the spatial-process full coverage correlation modeling of process parameters-detection results under multiple detection points-multiple process scenarios is realized, which provides a high-resolution probability basis for subsequent process scenario sub-model division, feature importance extraction, and dynamic drift detection steps.

[0119] For example, for the spatial proximity detection point grouping in the "reflow soldering" process of the SMT patch production line, 1200 groups of joint feature vectors containing process parameters (soldering temperature range 230-250 degrees Celsius, line humidity 30-55% RH, mounting precision offset 0-0.05 mm) and detection output (misjudgment / missed label) are selected as samples. The GMM algorithm is used, the upper limit of the component is set to K=4, and the initialization strategy is K-means++. The maximum iteration of the EM algorithm is 250 per round, and the covariance matrix is fully standardized. In the actual running process, the number of components is automatically adjusted to 3, the convergence output mixed weight is [0.42, 0.36, 0.22], and the mean and covariance of each component are clearly separated to couple the mode of process parameters and detection results. According to the conditional probability formula, the conditional probability of the detection misjudgment rate is improved to 0.083 under the condition of soldering temperature < 235°C and humidity > 48% RH, and the conditional probability of the misjudgment rate is reduced to 0.024 under the process parameter interval (240-245°C, humidity 38-45% RH). Further, the variational Bayesian inference is used to model the large-capacity sample grouping (mounting process grouping, 3500 cases), and finally the best mixed component 2 is retained to prevent model overfitting under high-dimensional features, and the AUC of the model sensitivity to the process parameter interval is improved to 0.961. The final output of each grouping probability distribution model parameter realizes the spatial-process joint distribution modeling of process parameters and detection results, provides a solid foundation for sub-model division and subsequent drift detection, and realizes the quantifiable, inferable and traceable technical effect of the influence of dynamic process parameter disturbance on detection performance.

[0120] S4.3: The above process parameter-detection result correlation distribution model is divided into sub-models according to the detection point area attribute and the process procedure label, a hierarchical grouping logic is used to construct a process area / procedure scene mapping sub-model library, and it is ensured that each manufacturing area or process scene has a dedicated modeling object.

[0121] The input is each grouping joint feature set and its process parameter-detection result joint probability distribution model obtained by S4.2, and the grouping information covers the detection point physical area attribute and the process procedure label.

[0122] A region label and process procedure label double-index grouping algorithm (parameters: region division granularity 1-5 mm, process procedure label category covering all manufacturing process types) is used to realize spatial-process double-layer classification of all grouping feature sets.

[0123] Further, by the region attribute comparison method (parameters: PCB board region marking system, automatic matching of detection point XY coordinates and design process blocks), each grouping feature set is assigned to the physical manufacturing block where it is located, and region grouping classification is realized.

[0124] Further, based on the process procedure label consistency criterion (parameter: process procedure homogeneity threshold 0.9), the grouping features in the same region and the same process label are aggregated, and according to the process flow distribution, the reflow soldering area, the patch area, the wave soldering area and other process scene subsets are established respectively.

[0125] Further, the hierarchical grouping logic (parameter: priority setting is region priority, and then process procedure) is adopted to further subdivide the above-mentioned preliminary grouping, and independent sub-model division strategies are output for the inspection point intensive area and the multi-process intersection area, so as to ensure that the physical space overlapping area and the process intersection area have exclusive modeling objects.

[0126] Further, through the sub-model instantiation algorithm, the process parameter-detection result joint probability distribution sub-model is instantiated for each grouping (unique combination of sub-region and process label), and the model object is added to the process region / process scene mapping sub-model library, so as to finally realize the structured management and index archiving of the sub-model library.

[0127] Through the above-mentioned hierarchical grouping and sub-model instantiation processing, the grouping features and the joint probability distribution model are converted into a hierarchical sub-model group which is mapped by the region attribute and the process procedure label, each manufacturing region or process scene is configured with a unique corresponding modeling object, the scene refinement and model specialization of the process parameter-detection result correlation modeling are effectively realized, and the basic support for the analysis robustness under the manufacturing diversity and process disturbance is provided.

[0128] For example, in an SMT production line AOI detection management process covering patching, reflow soldering and wave soldering, the PCB board is assumed to be divided into 4 physical detection regions, and the region division granularity is 2 mm. 32 AOI detection points belong to the 4 physical partitions respectively, and each partition experiences patching, reflow soldering and wave soldering procedure labels in turn. The total amount of input data is 12800 groups, and 12 preliminary groupings are formed according to the region label and the process procedure label combination. Through the hierarchical grouping logic, the physical space overlapping (such as the boundary and the intersection area of the channel) and the multi-process intersection (such as the double-process maintenance interface area) detection points are output as separate groupings. Using the above-mentioned grouping results, a Gaussian mixture joint distribution sub-model instance is established for each grouping (the average sample size of each sub-model is 800-1200 groups). Finally, the sub-model library contains 16 independent sub-model instances, which cover all physical detection areas and process scenes. Compared with the ungrouped single model, the misjudgment rate of the hierarchical grouping sub-model is improved by 7.6%, the missed sensitivity is improved by 12.4%, and the abnormal detection capability of the boundary area and the process intersection point is enhanced by 20%, so as to realize the expected technical effects of model specialization and analysis refinement under the manufacturing partition / process scene diversity.

[0129] S4.4: Performance evaluation and feature importance ranking of each sub-model of the process parameter-detection result correlation distribution model, based on statistical effectiveness criteria (such as mutual information, information gain) to filter the process parameter subspace that significantly affects the false positive rate and the false negative rate, and to provide fine-grained feature reference for subsequent dynamic drift detection and model adaptation mechanism.

[0130] S4.5: Generate a unique mapping index of each sub-model and the corresponding process scenario and detection point area, and link the index structure with the time sequence process parameter tracking database to realize closed-loop management of data flow, model flow and process flow, and guarantee the traceability of abnormal cause path and data consistency of subsequent sub-model dynamic switching.

[0131] The step S5: based on the joint feature vector, input the process parameter-detection result correlation distribution model into the dynamic drift detection algorithm, calculate the false positive rate and the false negative rate statistics of the current batch and the historical batch, and identify the abnormal drift of the model output caused by the process parameter disturbance. Specifically, it includes:

[0132] S5.1: Group the input joint feature vector sequence and the process parameter-detection result correlation distribution model using batch identification and time stamp, divide the detection data into current detection batch and historical detection batch, and establish batch-level false positive rate and false negative rate statistical basic data set.

[0133] S5.2: Based on the correlation distribution model, execute the classification decision algorithm for each batch-level joint feature vector, respectively calculate the false positive rate (False Positive Rate) and the false negative rate (False Negative Rate) of the current detection batch and the historical detection batch, and output the batch-level false positive rate statistics and the false negative rate statistics, which provide parameter basis for subsequent dynamic analysis.

[0134] The input is the joint feature vector sequence grouped by batch, which includes normalized and time-aligned detection space features and process parameter time sequence features, and batch-level corresponding process parameter-detection result correlation distribution model instances.

[0135] Batch-level classification decision algorithm (parameters: select corresponding distribution model instance for each batch, model type can be Gaussian Mixture Model GMM or Variational Bayesian distribution; classification threshold is dynamically set according to the training data confusion matrix) is adopted to realize the defect category judgment of the input joint feature vector, and output the judgment label (such as "qualified", "false positive", "false negative", etc.).

[0136] Further, based on the above classification decision result, the false positive sample number actual negative sample total number false negative sample number and the total number of actual positive samples Batch quantitative analysis of defect recognition performance is realized.

[0137] The following false positive rate and false negative rate calculation formulas are used to generate the corresponding performance statistics for each batch:

[0138]

[0139] Wherein, N FP is the number of samples determined as false positives in the batch, N N is the total number of actual negative samples (non-defects) in the batch, N FN is the number of missed samples, N P is the total number of actual positive samples (defects).

[0140] For historical detection batches, the same classification determination algorithm and statistical scheme are used to obtain the false positive rate and false negative rate statistics of each historical batch, forming a historical control data set.

[0141] Through batch-level classification determination and statistical quantity extraction, the false positive rates FPR curr ,{FPR hist,i} and false negative rates FNR curr ,{FNR hist,i} of the current batch and historical batches are obtained, which establish a complete data basis for subsequent dynamic drift detection and abnormal traceability analysis.

[0142] Through the above algorithm processing, the joint feature vector sequence and distribution model instance of the previous step are converted into batch-level false positive rate and false negative rate statistics, which provide accurate basic parameters for modeling the time sequence characteristics of the dynamic analysis model, process parameter drift response and adaptive performance optimization.

[0143] For example, in the reflow soldering process scenario, the input batch is production batch 20240430_X001, the batch detection data contains 1280 joint feature vectors, the corresponding distribution sub-model uses Gaussian mixture probability model (component number K=3), and the detection discrimination label is divided into: 1160 qualified, 60 false positives, 20 missed, and the remaining 40 are real defects based on the determination algorithm. After statistics, the actual negative samples positive samples false positive samples missed samples Using the above formulas, we get:

[0144]

[0145] The historical control data set selects the determination results of the past 10 batches of the same process and the same block to obtain the average of {FPR_hist,i} as 0.0381 and the average of {FNR_hist,i} as 0.2947. Thus, the misjudgment rate and the missed detection rate of the current batch increase relative to the historical average, which provides an objective statistical benchmark for subsequent dynamic drift detection and process traceability analysis. The batch-level FPR and FNR statistics and their time series are output, achieving the expected technical effect of batch quantization evaluation of detection performance under dynamic process disturbance.

[0146] S5.3: Fuse the batch-level misjudgment rate statistics and the missed detection rate statistics of the current batch and the historical batch, and dynamically generate the misjudgment rate drift feature sequence and the missed detection rate drift feature sequence using the sliding window method and the moving mean and variance calculation scheme, to realize the induction of batch time series statistical characteristics.

[0147] The input is the batch-level misjudgment rate (False Positive Rate, FPR) and missed detection rate (False Negative Rate, FNR) statistics generated for the current detection batch and the historical detection batch, which are extracted through a classification determination algorithm and an associated distribution model, and have been archived according to the detection batch timestamp and batch number.

[0148] The sliding window method (parameters: window length w is dynamically set according to the actual batch number of the production line, recommended w = 5-10) is used to realize the time series analysis of the historical batch misjudgment rate and missed detection rate sequence.

[0149] Further, the moving mean calculation algorithm (parameters: the mean updating method is weighted moving average, and the historical sequence weighting factor a = 0.6-0.9) is used to calculate the moving mean of the window misjudgment rate and missed detection rate sequence from batch t to t-w+1 FPR (t), μ FNR (t), reflecting the center tendency of the model output under process dynamic disturbance.

[0150] Further, the moving variance calculation algorithm (parameters: variance recursive updating method) is used to calculate the moving variance of the misjudgment rate and missed detection rate sequence in the window from batch t to t-w+1 to quantify the fluctuation intensity of the misjudgment and missed detection rates with batch drift.

[0151] The following recursive formula is used to realize the moving mean calculation:

[0152]

[0153] wherein, μ X (t) is the mean in the window up to the current batch t, X t-i is the statistics of the t-i batch, and w is the sliding window width.

[0154] The flow variance recursive formula is used:

[0155]

[0156] wherein, is the variance within the current batch t window, and X refers to FPR or FNR.

[0157] Further, by the time sequence recursion of the above mean and variance, the false positive rate drift feature sequence and the false negative rate drift feature sequence systematically reflects the influence trend and fluctuation distribution of process state changes on model false positive and false negative statistics values as the production batch advances.

[0158] By difference or normalization processing, the above drift feature sequence is converted into batch-to-batch drift measurement indicators, which facilitate subsequent dynamic drift detection algorithms for significant discrimination and anomaly detection.

[0159] Through sliding window and flow statistics processing, the historical and current data of batch-level false positive rate and false negative rate are converted into time sequence drift feature indicators reflecting the dynamic evolution of process disturbance and model performance, achieving the goal of batch statistical induction analysis of detection model output under process parameter dynamic changes.

[0160] For example, in the SMT patch production line AOI detection process, according to the actual production batch number, the production line outputs one detection batch every 2 hours (numbered B001-B020), and the FPR and FNR of each batch are collected through classification judgment. The original FPR sequence is {0.041, 0.036, 0.039, 0.045, 0.049,...}, and the FNR sequence is {0.27, 0.25, 0.31, 0.33, 0.30,...}.

[0161] The sliding window length w is set to 5. For the t = 10th batch, the flow mean is calculated using the FPR sequence of the previous 5 batches

[0162] {0.036, 0.039, 0.045, 0.049, 0.041}

[0163]

[0164] The flow variance is calculated using the same group of data:

[0165]

[0166] All batches repeat the above sliding window processing, and dynamically output the drift feature sequence. Actual detection found that the FPR drift mean of some batches (such as B015) increased for a short time, and the variance increased significantly, indicating that process disturbance caused the model output to drift statistically, providing high-resolution drift feature input for subsequent abnormal drift detection and model adaptation. Finally, through this step, the false alarm and missed alarm timing trend induction and batch abnormal response ability are significantly improved, laying the foundation for model dynamic robustness optimization in complex process scenarios.

[0167] S5.4: For the generated false alarm rate drift feature sequence and the missed alarm rate drift feature sequence, a dynamic drift detection algorithm (such as CUSUM, Page-Hinkley, or a Bayesian-based drift identification mechanism) is used to compare the significance difference of the drift features in the current batch and the historical interval, to determine whether there is a statistical abnormal drift phenomenon of the model output.

[0168] S5.5: When the dynamic drift detection algorithm determines that there is a statistical abnormal drift, the detected abnormal batch is associated with the corresponding process parameter disturbance record for analysis, and a pairing report of the abnormal drift event and the associated process parameter interval is generated as the input basis for the model adaptive optimization and traceability analysis process.

[0169] The step S6: When the model output abnormal drift is detected and associated with a specific process parameter interval, the retraining or incremental learning process of the analysis model is automatically triggered to adapt to the detection state under the specific process parameter. Specifically, it includes:

[0170] S6.1: Map the abnormal model output probability distribution output by the dynamic drift detection algorithm and the normalized process parameter time series features to the process parameter interval to determine the specific process-driven interval of the model drift, providing a basis for subsequent targeted retraining of the process parameter interval.

[0171] S6.2: Based on the process parameter interval mapping result, automatically generate a process parameter-detection image joint feature sample set, and use a time series distribution sampling algorithm to select representative training samples to take into account the current process state and the historical drift trend of the model.

[0172] The input is the process parameter abnormal interval mapped by the dynamic drift detection algorithm and the batch-level normalized detection image, process parameter joint feature vector data set. The joint sample pool construction method (parameters: target process interval number, historical drift label, batch index range) is adopted to dynamically filter out all joint feature samples in the current abnormal process parameter interval, including detection image space features and corresponding process parameter time series features. Further, through the time series distribution sampling algorithm (parameters: window length w, drift significance threshold δ, historical / current batch minimum sample size n_thr), representative samples are randomly / layered extracted within the current process interval, and the consistency of the process parameter disturbance state, model output drift degree and detection label (qualified, misjudgment, missed detection, etc.) distribution covered by the samples is ensured. The process parameter floating sampling weight distribution method is adopted. For the interval where the process parameter change rate is higher than the average drift level, the sampling weight is increased to enhance the representativeness of the samples in this interval, and the weight-adjusted sample distribution is obtained. Further, by comparing the historical and current process parameter disturbance level, the matching sampling method is adopted to supplement several historical process interval representative samples for the newly generated training sample set, ensuring that the new sample pool reflects the current drift process characteristics and is compatible with the historical robustness of the model. Through the sample consistency discrimination algorithm (such as distribution consistency test KS-test or KL divergence evaluation), it is confirmed that the selected sample set meets the representativeness of the target process interval and the historical drift interval in feature distribution and label distribution, and the samples with abnormal distribution or inconsistent labels are removed, and the final process parameter-detection image joint feature training sample set is output. Through the above algorithms and processing methods, the detection samples under the current and historical typical process parameter intervals are scientifically included in the training set, and the distribution and representativeness of the samples are strictly controlled, realizing the reasonable selection of incremental training basic samples and the enhancement of drift sensitivity.

[0173] Exemplarily, in the SMT mounting process scene, it is detected that the FPR mean value of the 20240507_B015 batch is improved, and it is determined that the process parameter interval [temperature: 235-245℃, humidity: 40-45%] is an abnormal drift interval, the target parameter w=8, δ=0.05, n_thr=500, and a total of 1280 sets of joint feature samples in the sampling window. Using stratified sampling algorithm, 640 sets of samples are extracted from the drift significant batch in the current abnormal process interval, and 320 sets of samples are supplemented according to the historical misjudgment rate weight in the historical stable process interval (temperature: 230-235℃), and the remaining 320 sets are selected from the overlapping transition interval according to the process change rate. All samples are labeled with detection judgment labels, and the KS value of the sample set is less than 0.1, and the KL divergence is lower than the set threshold 0.02, confirming the consistency of the distribution. Finally, 1280 sets of process parameter-detection image joint feature sample sets covering the current drift process parameter interval and the historical representative state are formed, providing a high-quality training basis for the incremental training of the subsequent analysis model and the improvement of the misjudgment traceability ability. The application effect shows that after incremental learning, the FPR of the model for the abnormal process parameter interval decreases by 14.7%, and the FNR decreases by 8.2%, effectively suppressing the detection and judgment abnormalities caused by process dynamic disturbance.

[0174] S6.3: For the screened process parameter-detection image joint feature sample set, the model parameter update is performed by using a multi-modal feature incremental learning framework, and the self-adaptive decision-making ability of the model to the process dynamic change condition is enhanced by integrating the new process interval features.

[0175] S6.4: In the model incremental learning process, the process parameter sensitivity change of the joint feature vector space before and after the update is compared, the robustness evaluation algorithm is used to analyze the optimization and improvement range of the misjudgment rate and the missed detection rate statistical indicators after retraining, and the effect closed loop under the process condition is ensured.

[0176] S6.5: Based on the output of the corrected analysis model and its adjustment history, an analysis model parameter change log is generated in real time, the model adaptability state under the current process parameter interval is labeled, and traceable evidence chain is provided for subsequent traceability explanation and further optimization.

[0177] The step S7: using an interpretable AI algorithm to perform causal reasoning on the process parameter change history, joint feature weight and model drift influence, and generating a dominant process factor sorting list of the misjudgment rate and the missed detection rate change of each batch to indicate the traceability path. Specifically, it includes:

[0178] S7.1: Synchronize the historical process parameter variation sequence with the detection false positive rate and false negative rate time series statistics, and use time series correlation analysis methods (such as Granger causality test) to obtain a preliminary causal relationship set between process parameter dynamic changes and model output trends, in order to extract the main control correlation factors of the process parameter variation history.

[0179] S7.2: Take the process parameter variation history as input, combine the joint feature weight distribution, call the explainable AI algorithm (such as SHAP value decomposition or LIME local explanation), calculate the marginal contribution of multi-dimensional process parameter features in the model false positive rate and false negative rate output, and obtain the causal influence mapping of joint feature weight and process parameter.

[0180] S7.3: Take the output abnormal drift events identified by the model drift detection algorithm as a criterion, based on the above process parameter-false positive rate / false negative rate marginal contribution, execute the causal reasoning engine (such as Bayesian Network inference based on causal graph model), and perform traceability analysis on the main process parameter factors affecting the model output drift.

[0181] The input data is the abnormal drift event set identified by the model drift detection algorithm, the process parameter-detection output marginal contribution degree statistical vector, the normalized process parameter history variation sequence, the joint feature weight distribution, and the false positive rate and false negative rate time series indicators of each detection batch.

[0182] An abnormal drift event positioning method (parameters: abnormal batch marker B_{drift}, drift significance threshold δ) is used to identify and number the output abnormal drift batch, and extract the process parameter history and model output of the corresponding batch interval.

[0183] Through the contribution weight matching algorithm (parameters: process parameter set P, marginal contribution vector C_{P→FPR / FNR}), the influence of each process parameter in the false positive rate / false negative rate output abnormal batch is measured, and the key process parameter subset P^* with a contribution degree greater than the threshold α is selected.

[0184] Further, a causal graph modeling method is used to construct a directed causal relationship graph G=(V,E) based on the abnormal drift batch process parameter history and detection output label, where the nodes V include P^* and the false positive rate / false negative rate output, and the edges E represent significant causal correlation.

[0185] Through the Bayesian Network structure learning algorithm (parameters: structure score function S, prior confidence β), the conditional probability distribution between the process parameters and FPR / FNR in the abnormal drift batch is modeled, and the causal chain structure inference is realized.

[0186] The maximum a posteriori inference mechanism is used to perform probability inference on the critical path of the causal graph G, identify the master process parameter factor set Q causing the model output drift, and output the conditional probability contribution of each process parameter to the occurrence of the drift event:

[0187]

[0188] wherein P(Drift FPR / FNR |P i ) represents the conditional probability of the misjudgment or missed detection of the abnormal drift event occurring under the given condition of the process parameter P i .

[0189] The conditional probability ranking list of the key process parameters Q is integrated to generate a master factor contribution matrix indicating the traceability path.

[0190] Through the above algorithm processing, the input abnormal drift detection result, process parameter marginal contribution, and process parameter historical variation data are converted into process parameter-model output drift causal links, and multi-factor traceability analysis is realized.

[0191] For example, for the SMT mounting batch 20240507_B015, the dynamic drift detection algorithm determines that the batch index B_{drift} = B015, the FPR drift significance δ = 0.042, and the corresponding process parameter interval [temperature: 237°C, humidity: 43%] is extracted. Through SHAP marginal contribution calculation, the process parameter subset with a contribution greater than 0.17 is selected: temperature P1, humidity P2, and mounting precision P3. Using the process parameter history record and the current FPR / FNR sequence, a causal relationship graph G is constructed, and a Bayesian network structure learning algorithm is used to train the probability network with a structure score function S = Bayesian Information Criterion (BIC) and a prior confidence β = 0.99.

[0192] The maximum a posteriori probability inference is used to calculate the conditional probability of the master process parameter under the FPR drift, wherein:

[0193] P(Drift FPR |P1 = 237) = 0.77

[0194] P(Drift FPR |P2 = 43) = 0.65

[0195] P(Drift FPR |P3 = 89%) = 0.22

[0196] The output causal link indicates that temperature is the main factor and humidity is the secondary influencing factor. Finally, a dominant process parameter factor weight ranking list Q = [P1(0.77), P2(0.65), P3(0.22)] is generated, indicating the traceability path of the FPR abnormal drift. The application results show that after this causal reasoning step, the main process parameters affecting the drift event are clearly distinguished and ranked, providing strong data support for subsequent model adaptive adjustment and manufacturing parameter optimization.

[0197] S7.4: The process parameter variation history, joint feature weight decomposition result and model drift causal chain of different time series batches are weighted and aggregated, and a dominant process factor weight list of the false rejection rate and the missed detection rate change of each batch is generated by applying a weight ranking and normalization processing algorithm to ensure the consistency and comparability of the ranking results under multiple batches and multiple distribution conditions.

[0198] S7.5: Based on the dominant process factor weight list, an interpretable AI algorithm recommendation mechanism is used to automatically generate the traceability path structure of each batch, including the subset of main control process parameters and their logical link to the change of false rejection rate and missed detection rate, clearly indicating the process parameters that need to be paid attention to and adjusted as the output of traceability analysis and the input basis for subsequent model adjustment.

[0199] The step S8: According to the dominant process factor ranking list, the parameter structure of the analysis model is dynamically adjusted, specifically including using a targeted sub-model or changing the loss function weight when the process parameter interval is switched, to realize the adaptive optimization of the analysis model under multiple process scenarios. Specifically, it includes:

[0200] S8.1: Based on the dominant process factor ranking list output by the interpretable AI algorithm, the dominant influencing factor set corresponding to each process parameter interval is extracted as the input condition for model adjustment, and the mapping relationship between each process parameter interval and the change trend of false rejection and missed detection rate is determined through correlation analysis.

[0201] S8.2: For the process parameter interval corresponding to the dominant influencing factor set, based on the process parameter-detection result correlation distribution model, the current recommended targeted sub-model structure is automatically retrieved to optimize the adaptive response capability of the model in this process parameter interval, and to realize the model switching output for specific process environment.

[0202] S8.3: Based on the switching to the targeted sub-model structure, for the process parameter interval boundary or overlapping region, a loss function weight self-adjustment strategy is applied to dynamically update the model loss function weight parameters, to realize the optimization output of the transition interval false rejection and missed detection trade-off capability.

[0203] Within the process parameter interval switched to the targeted sub-model structure, input includes process parameter data of the current detection batch, joint feature vector and dominant process factor weight list.

[0204] The process parameter interval boundary identification algorithm (parameter: interval boundary threshold ε) is adopted to realize dynamic comparison between the actual value of the process parameter and the interval critical value, and to mark the detection batch statistical characteristics in the transition interval or overlapping area.

[0205] Further, through the dynamic loss function weight adjustment algorithm (parameters: dominant process factor weight Q, false positive rate and false negative rate target weight β_{FPR},β_{FNR}), the loss function weight for false positive rate and false negative rate in the analysis model is adaptively adjusted according to the weight contribution of the dominant factor in the current process parameter overlapping area.

[0206] The following weighted loss function form is adopted:

[0207] L=β FPR ·L FPR +β FNR ·L FNR

[0208] Wherein, L is the total loss function, L FPR is the false positive loss component, and L FNR is the false negative loss component.

[0209] Further, through the parameter dynamic optimization strategy (parameter: Q=[q_1,q_2,...,q_n]), the influence weight of each component of the dominant process parameter on β FPR and β FNR is analyzed in real time to realize the following adaptive adjustment:

[0210]

[0211] Wherein, γ is the step coefficient, and ΔFPR i and ΔFNR i are the marginal contribution of the i-th dominant process parameter to the false positive rate and false negative rate change in the overlapping interval, respectively.

[0212] Monte Carlo sampling or adaptive back propagation algorithm is applied to batch update the loss weight parameters of the analysis model, and the convergence of the loss curve is discriminated, and the optimal loss weight configuration is output.

[0213] Through the above processing mode, the model weight when the current process parameter is in the interval overlap or critical value is converted into a highly adaptive loss function parameter, which realizes the trade-off optimization of the model to the false positive rate and false negative rate in the process transition interval, and significantly improves the robustness and accuracy of the model in complex dynamic process environment.

[0214] Exemplary, in SMT soldering batch 20240507_B037, there is an overlap between the current process parameter temperature interval [235℃, 240℃] and the humidity interval [42%, 46%], and the overlapping area batch detection data is input into the targeted sub-model with the dominant process factor weight list Q = [temperature (0.80), humidity (0.62), mounting accuracy (0.18)]. According to the dynamic loss function weight adjustment algorithm, β_{FPR} = 0.5, β_{FNR} = 0.5, and the step coefficient γ = 0.05 are initialized. According to the FPR and FNR marginal contribution of each dominant process parameter in the current batch, ΔFPR_temperature = +0.03, ΔFPR_humidity = +0.01, ΔFNR_temperature = -0.01, and ΔFNR_humidity = +0.02, the weight adjustment is calculated as follows:

[0215]

[0216] After adaptive adjustment of the model weight, the loss convergence curve is stable, the FPR of the overlapping area is reduced to 3.6%, and the FNR is controlled at 2.4%, which is better than the performance before weight adjustment (FPR 5.1%, FNR 4.0%), and the detection performance in the process transition interval is optimized.

[0217] S8.4: According to the dynamic process parameter tracking record, the change of the process parameter interval is continuously monitored, the opportunity for adaptive adjustment of the current model parameter or structure is determined through linkage with historical model performance statistics (such as batch misjudgment rate, average / variance of missed detection rate), and the optimal model structure or weight configuration is output.

[0218] S8.5: An incremental learning mechanism is adopted to link the model output after each model switching or loss function weight adjustment with real-time process parameters, and through feedback residual scalar or drift statistics, the adaptive effect of the analysis model in the new process parameter interval is further modified, and the optimized process parameter-model structure-detection result coupling matching output is obtained.

[0219] The step S9: The dynamic misjudgment traceability analysis result and the process parameter tracking record are fed back to the manufacturing execution system in real time, and the process optimization suggestion and detection parameter configuration modification are realized in a closed loop. Specifically, it includes:

[0220] S9.1: Based on the dominant process factor sorting result output by the causal reasoning module, the dynamic misjudgment traceability analysis result and the process parameter tracking record of the current batch are obtained to prepare the standardized feedback message.

[0221] S9.2: The results of dynamic misjudgment tracing analysis and process parameter tracking records are standardized and encapsulated using a data communication protocol compatible with Manufacturing Execution System (MES) to ensure the formatting and integrability of the feedback content.

[0222] S9.3: Standardized encapsulated data packets are pushed to the data receiving interface of the manufacturing execution system in real time for highly reliable industrial information transmission via industrial Ethernet or OPC UA bus.

[0223] S9.4: Based on the process parameter optimization rules of the manufacturing execution system, automatically generate a list of process optimization suggestions based on the received dominant process factors ranking and detection anomaly results, including process parameter range adjustments, equipment inspection plans, and suggestions for correcting detection parameter configurations.

[0224] S9.5: The process optimization suggestion list is pushed to process engineers and production line control units through the manufacturing execution system to realize adaptive adjustment of production formula parameters and real-time correction of detection parameter configuration, and finally complete the closed-loop control process of dynamic process parameters-detection analysis results-process optimization suggestions.

[0225] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0226] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart algorithm-based circuit board AOI detection result analysis method, characterized in that, The method comprises the following steps: S1: Collecting a synchronous time series data set comprising detection image data and multi-dimensional process parameters as input data for subsequent feature processing; S2: Denoising, normalizing and timestamp aligning the collected detection image data and process parameter time series data respectively; S3: Based on the normalized detection image data and process parameter time series data, using feature alignment and attention mechanism to extract joint feature vectors reflecting process dynamic changes; S4: For multiple detection point positions and process procedure labels, establish a process parameter-detection result correlation distribution model, and group the sub-models, respectively corresponding to different manufacturing areas and process scenarios; S5: Based on the joint feature vector, input the process parameter-detection result correlation distribution model into the dynamic drift detection algorithm, calculate the misjudgment rate and missed detection rate statistics of the current batch and historical batches, and identify the model output abnormal drift caused by process parameter disturbance; S6: When detecting the abnormal drift of the model output and correlating to a specific process parameter interval, automatically trigger the retraining or incremental learning process of the analysis model to adapt to the detection state under specific process parameters; S7: Using an interpretable AI algorithm to perform causal reasoning on process parameter variation history, joint feature weight and model drift influence, and generate a dominant process factor ranking list of the misjudgment rate and missed detection rate changes of each batch; S8: According to the dominant process factor ranking list, dynamically adjust the parameter structure of the analysis model. 2.The smart algorithm-based circuit board AOI detection result analysis method according to claim 1, characterized in that, The step S8 further comprises: S9: Real-time feedback of dynamic misjudgment traceability analysis results and process parameter tracking records to the manufacturing execution system, closed-loop implementation of process optimization suggestions and detection parameter configuration correction. 3.The smart algorithm-based circuit board AOI detection result analysis method of claim 1, wherein, The step S1 specifically comprises: Configure the AOI detection equipment of the circuit board to be detected, so that each circuit board in the detection batch will have its original detection image data acquired through a high-resolution image acquisition interface when it is detected, and stored in a standardized format; Configure multi-channel sensing nodes to the production line process parameter acquisition unit to collect real-time process parameters synchronized with the detection batch; Based on the AOI detection system host platform, use a unified clock synchronization protocol to perform high-precision synchronization encoding on the timestamps of the AOI detection images and the acquisition timestamps of the process parameters; Batch number and metadata registration are performed on the detection image data and process parameter data sets obtained by synchronous acquisition, and the acquisition batch information and process parameter correlation table are written into the database in real time through the manufacturing execution system interface; After data acquisition is completed, perform data integrity verification and packet loss detection algorithm, and use a pre-defined retry mechanism to perform data resampling or mark as abnormal for damaged or missing detection image data and process parameter data.

4. The smart algorithm-based circuit board AOI detection result analysis method according to claim 3, characterized in that, The process parameters include soldering temperature, production environment humidity, mounting process precision and process setting value.

5. The smart algorithm-based circuit board AOI detection result analysis method according to claim 1, characterized in that, The step S3 specifically comprises: Perform feature dimension calibration processing on the normalized detection image data and process parameter time series data, use a standard feature alignment algorithm to one-to-one map the feature dimensions, sampling intervals and time series labels of different data modalities, and obtain a feature alignment matrix; Based on the feature alignment matrix, a deep representation processing of a feature encoder is performed, a spatial feature of the detection image data is extracted by using a convolutional neural network, and a dynamic state modeling of the process parameter time series data is performed by using a time series recurrent network, to obtain a detection spatial feature vector and a process time series feature vector; The detection spatial feature vector and the process time series feature vector are input into a multi-modal fusion module, a feature cross algorithm based on an attention mechanism is called, a correlation weight of different modal features under each time series segment is calculated, and a cross-modal representative fusion feature vector is adaptively aggregated according to the weight; The cross-modal representative fusion feature vector is subjected to feature anomaly suppression and time series regularization processing, a specific regularization loss function is introduced to suppress the noise distribution of abnormal fusion features and maintain the fluency of the features changing with process disturbance, and finally a joint feature vector meeting the process dynamic distribution characteristics is output. The joint feature vector is subjected to feature importance measurement and redundant feature screening, and a feature contribution degree analysis algorithm is used to select a process-related sub-feature set having the most influence on misjudgment and missed detection discrimination.

6. The smart algorithm-based circuit board AOI detection result analysis method according to claim 5, characterized in that, The process-related sub-feature set selected in the step S3 is specifically: The joint feature vector output after feature fusion is selected by using a Shapley value or information gain feature contribution degree algorithm to select a feature subset most sensitive to discrimination misjudgment rate and missed detection rate, so as to enhance the explanation and discrimination of the model to the process disturbance.

7. The smart algorithm-based circuit board AOI detection result analysis method according to claim 1, characterized in that, The step S4 specifically includes: Based on the joint feature vector obtained in the step S3, input data is aggregated according to detection point physical positions and process procedure labels, a spatial distribution mapping method is used to form a grouping feature set of each detection point and corresponding process procedure label; A multivariate statistical modeling algorithm is used to model a joint probability distribution of the process parameters and the detection results in each grouping feature set, to obtain a process parameter-detection result correlation distribution model; The process parameter-detection result correlation distribution model is divided into sub-models according to detection point region attributes and process procedure labels, and a hierarchical grouping logic is used to construct a process region / procedure scene mapping sub-model library; The performance of each sub-model of the process parameter-detection result correlation distribution model is evaluated and the feature importance is sorted, and a statistically effective criterion is used to screen a process parameter subspace having a significant influence on misjudgment rate and missed detection rate; A unique mapping index of each sub-model and corresponding process scene and detection point region is generated, and the unique mapping index is connected with a time series process parameter tracking database. 8.The smart algorithm-based circuit board AOI detection result analysis method of claim 1, wherein, The step S5 specifically includes: The input joint feature vector sequence and the process parameter-detection result correlation distribution model are grouped by using batch identification and time stamp, the detection data is divided into a current detection batch and a historical detection batch, and a batch-level misjudgment rate and missed detection rate statistical basic data set is established; For each batch-level joint feature vector, a classification judgment algorithm is executed based on the correlation distribution model, the misjudgment rate and the missed detection rate of the current detection batch and the historical detection batch are respectively counted, and a batch-level misjudgment rate statistical quantity and a missed detection rate statistical quantity are output. The batch-level false rejection rate statistics and the missed detection rate statistics of the current batch and the historical batches are fused, a sliding window method and a moving mean and variance calculation scheme are adopted, and a false rejection rate drift feature sequence and a missed detection rate drift feature sequence are dynamically generated; For the generated false rejection rate drift feature sequence and the missed detection rate drift feature sequence, a dynamic drift detection algorithm is adopted to compare the significant differences of the drift features in the current batch and the historical interval, and to determine whether there is a statistical abnormal drift phenomenon of the model output; When the dynamic drift detection algorithm determines that there is a statistical abnormal drift, the detected abnormal batch is associated with the corresponding process parameter disturbance record for analysis, and a pairing report of the abnormal drift event and the associated process parameter interval is generated. 9.The smart algorithm-based circuit board AOI detection result analysis method of claim 8, wherein, The dynamic false rejection rate and missed detection rate drift detection specifically comprises: adopting a combination of a sliding window, a moving mean and a variance, combining a CUSUM or Bayesian drift recognition method to determine statistical abnormal drift.

10. The smart algorithm-based circuit board AOI detection result analysis method according to claim 1, characterized in that, The step S6 specifically comprises: The abnormal model output probability distribution output by the dynamic drift detection algorithm is mapped to the process parameter interval with the normalized process parameter time sequence features; Based on the process parameter interval mapping result, a process parameter-detection image joint feature sample set is automatically generated, and a time sequence distribution sampling algorithm is used to select representative training samples; The process parameter-detection image joint feature sample set after screening is used to update the model parameters by using a multi-modal feature incremental learning framework; During the model incremental learning process, the process parameter sensitivity changes of the joint feature vector space before and after updating are compared, and a robustness evaluation algorithm is used to analyze the optimization and improvement amplitude of the false rejection rate and the missed detection rate statistical indicators by retraining; Based on the corrected analysis model output and its adjustment history, an analysis model parameter change log is generated in real time, and the model adaptability state under the current process parameter interval is labeled.

Images