Circuit board AOI detection result analysis method based on intelligent algorithm

Abstract

The invention discloses a circuit board AOI detection result analysis method based on an intelligent algorithm, and the core of the method is to carry out the high-precision synchronous collection, denoising normalization processing and time sequence alignment of a detection image and multi-dimensional process parameters, and extract joint features through an attention mechanism. A process parameter-detection result correlation distribution model under each process scene is established, model output abnormal drift detection and tracing reasoning are realized, the detection model and the process parameters are dynamically optimized in combination with interpretability AI, the method realizes efficient tracking, attribution and adaptive optimization of detection misjudgment caused by process parameter variation, and the detection accuracy is improved. And the stability of the production process of the SMT detection system is improved.

Description

Technical Field

[0001] This invention relates to the field of electronic manufacturing quality inspection and intelligent analysis technology, and in particular to a method for analyzing AOI inspection results of circuit boards based on intelligent algorithms. Background Technology

[0002] Currently, in the electronics manufacturing industry, especially in the printed circuit board (PCB) sector, Automated Optical Inspection (AOI) has become a crucial quality control method in manufacturing processes such as surface mount technology (SMT) and soldering. Modern AOI systems commonly employ technologies such as deep learning, feature extraction, and multimodal data analysis to perform high-precision detection, discrimination, and classification of PCB defects, thereby improving the intelligence and automation level of the manufacturing process. Mainstream detection and analysis solutions typically include image acquisition and processing, defect discrimination model training and application, and data statistics and attribution analysis for false positives and false negatives. Furthermore, to adapt to constantly changing manufacturing processes and complex production environments, the industry has gradually introduced process parameter-driven auxiliary analysis, statistical process control, and interpretable AI methods in recent years, continuously improving the adaptability and robustness of AOI inspection systems.

[0003] Currently, typical AOI intelligent analysis systems mainly rely on the following technical paths: First, high-precision image acquisition of key areas such as solder joints and components is achieved based on high-resolution industrial cameras; second, deep learning or traditional machine learning methods are used to denoise, extract features, and classify the image data, and output pass / defect judgment results; third, some systems introduce a two-dimensional model of "process parameters + inspection data", and preliminarily associate process setting conditions with inspection performance indicators (such as false positives and false negatives) through statistical regression analysis and cluster comparison; fourth, in the traceability analysis stage, offline review and simple rule correction are performed on typical false positives and abnormal false negatives based on information such as process batches and inspection points; fifth, in high-end applications, limited interpretable AI algorithms are used to assist in the analysis of error causes, or inspection parameters are optimized through expert experience.

[0004] While the methods described above have achieved some degree of automated inspection and quality traceability in the manufacturing process, most AOI (Automated Optical Inspection) misjudgment and missed detection analysis techniques still primarily focus on modeling inspection data under static process conditions or single batches and scenarios. In actual complex manufacturing processes, SMT process parameters (such as soldering temperature, humidity control, placement accuracy, and equipment settings) often exhibit range-bound, periodic, or even sudden changes due to environmental disturbances, equipment aging, or human error. This dynamic process fluctuation causes a coordinated drift between manufacturing process parameters and the input distribution of the inspection model, severely limiting the adaptability of existing static models and attribution algorithms to changes in misjudgment and missed detection rates. Common problems include:

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

[0006] (2) When process parameters are abnormal, the degradation pattern of detection performance in different manufacturing areas / processes is spatially heterogeneous, and the existing "one-size-fits-all" attribution strategy cannot accurately locate the main control factor.

[0007] (3) When AOI detection results show batch drift or continuous abnormality, the existing technology lacks dynamic analysis and model adaptive correction mechanism driven by process parameters, which can easily lead to misjudgment, missed detection and other quality problems being covered up for a long time, affecting the yield rate and control efficiency of the production line. Summary of the Invention

[0008] In order to solve the above-mentioned technical problems, the present invention provides a method for analyzing the AOI inspection results of circuit boards based on intelligent algorithms.

[0009] The technical solution of this invention is implemented as follows: a method for analyzing AOI inspection results of circuit boards based on intelligent algorithms, comprising:

[0010] S1: Acquire a synchronous time-series data set including detection image data and multi-dimensional process parameters, where the process parameters include welding temperature, humidity, mounting accuracy and process settings during the production cycle, as input data for subsequent feature processing.

[0011] S2: The collected detection image data and process parameter time series data are denoised, normalized and timestamp aligned respectively to ensure the consistency of multimodal feature data in time and dimension.

[0012] S3: Based on normalized detection image data and process parameter time series data, a joint feature vector reflecting the dynamic changes of the process is extracted using feature alignment and attention mechanisms, so as to take into account the multimodal characteristics of process and detection information.

[0013] S4: For multiple detection point locations and process operation labels, establish a process parameter-detection result correlation distribution model, and group the sub-models to correspond to different manufacturing areas and process scenarios.

[0014] S5: Based on the joint feature vector, the process parameter-detection result correlation distribution model is input into the dynamic drift detection algorithm to calculate the false positive rate and false negative rate statistics of the current batch and historical batches, and to identify abnormal drift in model output caused by process parameter disturbances.

[0015] S6: When an abnormal drift in the model output is detected and correlated with a specific process parameter range, the retraining or incremental learning process of the analysis model is automatically triggered to adapt to the detection state under the specific process parameters.

[0016] S7: Use interpretable AI algorithms to perform causal reasoning on the history of process parameter changes, joint feature weights, and the impact of model drift, and generate a ranking list of the dominant process factors for changes in the false positive rate and false negative rate of each batch to indicate the causal path.

[0017] S8: Based on the ranking list of dominant process factors, dynamically adjust the parameter structure of the analysis model. Specifically, this includes using targeted sub-models or changing the weights of the loss function when switching between process parameter ranges, in order to achieve adaptive optimization of the analysis model under multiple process scenarios.

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

[0019] Beneficial effects

[0020] The present invention provides a method for analyzing AOI inspection results of circuit boards based on intelligent algorithms, which has the following beneficial effects:

[0021] (1) This invention innovatively introduces multi-channel synchronous acquisition and temporal process parameter encoding, achieving high-precision alignment between detection image data and process parameters, providing a solid data foundation for feature fusion and correlation modeling. Based on feature alignment, multimodal feature extraction using convolutional neural networks and temporal recursive networks, and a cross-modal attention fusion mechanism, the influence of process parameter fluctuations on detection and discrimination is fully explored, significantly enhancing the model's sensitivity and adaptability to dynamic process changes. Multi-batch testing shows that compared with traditional single models, the drift sensitivity of the false positive rate is improved by more than 20%, and the drift response of the false negative rate is improved by more than 15%, ensuring stable output of cause analysis in complex scenarios such as process adjustments and high-variance batches.

[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] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0030] When used herein, the singular forms of “a,” “an,” and “the” may also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “comprising / including” or “having,” etc., specify the presence of the stated features, wholes, steps, operations, components, parts, or combinations thereof, but do not preclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts, or combinations thereof. Meanwhile, the term “and / or” as used in this specification includes any and all combinations of the associated listed items.

[0031] Please see Figure 1 As shown, a method for analyzing AOI inspection results of circuit boards based on intelligent algorithms includes:

[0032] S1: Acquire a synchronous time-series data set including detection image data and multi-dimensional process parameters, where the process parameters include welding temperature, humidity, mounting accuracy and process settings during the production cycle, as input data for subsequent feature processing.

[0033] S2: The collected detection image data and process parameter time series data are denoised, normalized and timestamp aligned respectively to ensure the consistency of multimodal feature data in time and dimension.

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

[0035] S4: For multiple detection point locations and process operation labels, establish a process parameter-detection result correlation distribution model, and group the sub-models to correspond to different manufacturing areas and process scenarios.

[0036] S5: Based on the joint feature vector, the process parameter-detection result correlation distribution model is input into the dynamic drift detection algorithm to calculate the false positive rate and false negative rate statistics of the current batch and historical batches, and to identify abnormal drift in model output caused by process parameter disturbances.

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

[0038] S7: Use interpretable AI algorithms to perform causal reasoning on the history of process parameter changes, joint feature weights, and the impact of model drift, and generate a ranking list of the dominant process factors for changes in the false positive rate and false negative rate of each batch to indicate the causal path.

[0039] S8: Based on the ranking list of dominant process factors, dynamically adjust the parameter structure of the analysis model. Specifically, this includes using targeted sub-models or changing the weights of the loss function when switching between process parameter ranges, in order to achieve adaptive optimization of the analysis model under multiple process scenarios.

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

[0041] Step S1: Acquire a synchronous time-series data set including detection image data and multi-dimensional process parameters, wherein the process parameters include welding temperature, humidity, mounting accuracy, and process settings within the production cycle, as input data for subsequent feature processing. Specifically, this includes:

[0042] S1.1: Configure the AOI inspection equipment for the circuit boards to be inspected so that each circuit board in the inspection batch can acquire its original inspection image data through a high-resolution image acquisition interface when it is inspected, and store it in a standardized image data format to ensure the uniformity of subsequent image feature preprocessing.

[0043] The AOI inspection equipment for the circuit boards under inspection is parameterized, setting the inspection resolution, lens parameters, and light source conditions to standardize the high-precision image data capture environment. A high-resolution image acquisition interface is used, with an industrial camera triggering and capturing images of each circuit board at a designated station on the inspection conveyor belt. This acquires raw inspection images reflecting microstructural details such as solder joints and mounting details, ensuring no overlap or missed inspections in single-board image acquisition. Furthermore, a standardized image data format conversion module (parameters set to 24-bit RGB three-channel or 8-bit grayscale, switching according to inspection requirements) unifies the format of the acquired raw inspection images, generating lossless files such as TIFF and PNG for comprehensive feature extraction. Metadata registration is performed using unique inspection batch and board identifiers (such as QR codes or RFID physical unique codes). Each image is logically associated with the batch number, board number, inspection station number, and timestamp, establishing a complete metadata chain covering the entire inspection process. Based on the standardized image data storage module, a redundancy check algorithm is invoked to perform damage detection and storage consistency verification on the generated image files, outputting anomaly marker signals for subsequent re-acquisition. Verified image data is written to a structured database or dedicated file storage system, archived in batches, and forms an image dataset with retrieval and efficient read / write capabilities. Through a series of standardized acquisition and storage processing methods, the original inspection images of each circuit board during the batch inspection process are transformed into structured, archived, and searchable standard inputs, providing a unified and data-quality-assured image input foundation for downstream multimodal feature extraction and process parameter synchronization.

[0044] For example, a high-resolution industrial camera (Basler acA1920-155um) is configured on an SMT production line, with a shooting resolution of 1920×1200, a frame rate of 30fps, a ring-shaped high-brightness LED light source, an exposure time of 20ms, and an aperture of F3.5. During batch inspection, a photoelectric sensor triggers the camera to capture images each time a board passes on the conveyor belt, using the GigE Vision triggering protocol to transmit image data in real time. The image acquisition module automatically converts the 1920×1200 24-bit raw image into PNG format and associates it with the 2D Data Matrix board code and batch number obtained by the barcode scanner, generating a filename with a format like batch20240416_board0001.png. A supporting database uses SQL Server to register and index the image and metadata. After acquisition, an MD5 checksum and a breakpoint retransmission mechanism are used for integrity verification. Broken frames, damaged or missing files are automatically re-acquired to ensure no data is lost in the batch. After this process, the detection task with batch number 20240416 achieved standardized acquisition and batch archiving of original images. The resolution of the original images on the single board, the integrity of the files, and the accuracy of metadata retrieval all reached 100%, establishing 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 to collect process parameters (including welding temperature, production environment humidity, mounting process accuracy and process setting value, etc.) synchronized with the inspection batch in real time. Through the continuous sampling algorithm of multi-dimensional sensor data stream, realize the associated structured storage of process parameters on the inspection time axis, and provide high-frequency process parameter raw data stream for subsequent time alignment processing.

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

[0047] High-precision distributed temperature 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), vision / laser displacement-type placement accuracy monitoring equipment (0.01mm resolution, sampling cycle synchronized with placement cycle) and process controller data reading interfaces (such as Modbus RTU protocol, real-time retrieval of current process setting parameters) are used to ensure a multi-dimensional process parameter acquisition chain that is fully synchronized with the inspection timeline of each circuit board.

[0048] Using a multi-channel data synchronous acquisition and control algorithm (parameter setting: clock alignment error of each channel is less than 10ms), the sensing data of each sensing node is scheduled in real time to generate a data sampling packet synchronized with the AOI detection trigger signal, and at the same time, the batch number and the board number with timestamp are recorded.

[0049] Furthermore, by using continuous sampling data stream encoding and buffer management algorithms, and following a multi-level index format of detection batch-board number-sensor channel-timestamp, the process parameter sample stream at each moment is associated and structured to generate a high-frequency, multi-dimensional process parameter raw data sequence.

[0050] The system employs data consistency and missing data monitoring algorithms to periodically check the integrity of process parameter streams in each channel. For data segments with packet loss, compensation is provided through interpolation estimation or resampling strategies to ensure that each detection moment corresponds to complete and multi-channel synchronized process parameter data points.

[0051] Through real-time collaboration, continuous sampling, and structured storage of the aforementioned multi-channel sensing nodes, the input of high-frequency process parameter raw data streams that precisely correspond to the time axis of the inspection batch and board is realized, providing a data foundation for subsequent time sequence alignment, feature fusion, and misjudgment root cause analysis.

[0052] For example, on a typical SMT production line, for an AOI inspection batch 20240416, a Siemens S7-1200 PLC control system is selected. The system uses the Modbus RTU protocol to acquire real-time data from a PT100 temperature sensor at the soldering station (specifically, acquiring the surface temperature of the soldering head every second, with an accuracy of ±0.1℃). An SHT35 humidity sensor is deployed at key nodes in the production area, obtaining a relative humidity reading every 2 seconds, with a monitoring accuracy of ±1.5%RH. Placement accuracy is achieved by a Keyence LJ-V7000 series laser displacement sensor, which immediately captures the position drift value after each action at the placement point and caches it in a local database via PLC communication. Process settings are issued through a Siemens HMI industrial touchscreen interface, with data acquisition automatically occurring within 0.5 seconds after each process parameter change. All sensor channels are uniformly aligned with the timestamps generated by the barcode scanning and photo taking at the AOI inspection station, with clock synchronization accuracy errors controlled within 5ms.

[0053] During the inspection process, the data acquisition controller and the AOI inspection main control PLC jointly schedule the data. Upon completion of each board inspection, a structured record containing [board number, batch number, inspection time, welding temperature, humidity value, mounting accuracy, and process setting parameters] is generated in the data buffer. For momentary packet loss from unit sensors, single-point anomalies are corrected through linear interpolation. Ultimately, the number of raw process parameter data sets acquired in a single batch corresponds one-to-one with the total number of boards inspected, with a missed acquisition rate of less than 0.01%. This high-frequency, high-completeness multi-channel raw process parameter data stream lays a high-quality data foundation for downstream process parameter synchronization with AOI inspection images, joint multimodal feature extraction, and batch misjudgment rate root cause analysis.

[0054] S1.3: Based on the main control platform of the AOI inspection system, a unified clock synchronization protocol is used to perform high-precision synchronization encoding of the timestamp of the AOI inspection image and the acquisition timestamp of the process parameters. By generating a unique time sequence code identifier for each inspection cycle, a strict correspondence between the inspection image data and the process parameter data within the acquisition batch is realized, so as to accurately input the synchronization data for the feature fusion module.

[0055] S1.4: The batch number and metadata registration of the synchronously acquired detection image data and process parameter data groups are performed. The acquired batch information and process parameter association table are written into the database in real time through the Manufacturing Execution System (MES) interface to ensure that each group of data has a unique ownership and traceability attribute in the subsequent processing stage, supporting the subsequent process parameter-driven misjudgment tracing analysis data tracking.

[0056] S1.5: After data acquisition is completed, execute data integrity verification and packet loss detection algorithms. For damaged or missing detection image data and process parameter data, use a predefined retry mechanism to reacquire data or mark anomalies to ensure the quality of the original data set finally used for feature extraction and model training, forming a highly reliable multimodal time series data input.

[0057] Step S2: The acquired detection image data and process parameter time-series data are subjected to denoising, normalization, and timestamp alignment processing to ensure the consistency of multimodal feature data in terms of time and dimensions. Specifically, this includes:

[0058] S2.1: Spatial domain denoising is performed on the detected image data. Based on median filtering, bilateral filtering or convolutional neural network adaptive denoising algorithms, image noise interference is removed to obtain noise-suppressed detected image data, laying a pure data foundation for subsequent feature normalization.

[0059] Spatial domain denoising is performed on the acquired detection image data. The input is the set of raw detection image data obtained through a high-resolution industrial camera and a standardized acquisition process. A median filtering algorithm (parameters: filter window size 3×3 or 5×5) is used to perform pixel-level spatial median filtering on each detection image, effectively removing isolated outliers such as salt-and-pepper noise while preserving structural edge details. Furthermore, a bilateral filtering algorithm (parameters: spatial domain Gaussian kernel σ_s, pixel color difference Gaussian kernel σ_r) is used to preserve image edge information while reducing noise, suppressing Gaussian and mean-shift noise distributions, and obtaining a spatially smoother result with stronger edge protection. The weighted average of the image pixels is calculated using the following formula:

[0060]

[0061] Among them, I out (x) represents the pixel value of the output image, I(x) and I(y) are the pixel values ​​of position x and neighborhood y in the original image, respectively, Ω is the neighborhood window, and σ s σ represents the standard deviation of spatial distance. r W is the weighted standard deviation of pixel differences. p For normalized weights. Furthermore, based on convolutional neural network (CNN) adaptive denoising algorithms (such as DnCNN, U-Net structure, parameters: network depth, training loss function using mean squared error), images with known noise distributions are used as input. An end-to-end network learns the discriminative features of spatial noise and image content, achieving highly adaptive denoising for complex noise. During CNN denoising, the following loss function is used to train the parameters:

[0062]

[0063] Among them, L MSE For mean square error loss, To reduce noise, output the i-th pixel of the image. Let be the i-th pixel of the noise-free reference image, and N be the total number of pixels. Through the cascaded processing of the aforementioned physical model-based filtering method and data-driven neural network method, various spatial noise interferences are effectively eliminated, resulting in noise-suppressed detection image data with fully restored content and edge details. The denoising quality is quantitatively evaluated using performance indicators such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) of the denoised image, ensuring that the denoised detection image data meets the technical requirements for providing high-purity input to subsequent normalization and feature extraction stages.

[0064] For example, on an SMT assembly line, spatial domain denoising is applied to circuit board inspection images acquired at a resolution of 1920×1200 and in 24-bit RGB format. To address ADC acquisition noise and random laser point interference in the environment, the median filter window size is set to 5×5, the spatial standard deviation of the bilateral filter is σ_s = 2, and the pixel standard deviation is σ_r = 15. First, median filtering is performed in parallel on each channel of the original image, and then bilateral filtering is used to further suppress spatially uneven noise. For local areas with particularly complex high-frequency noise, a pre-trained DnCNN neural network model is used. The training dataset contains 40,000 real industrial defect images, the network has 17 layers, and the Adam optimizer and MSE loss are used, with 100,000 iterations. After noise reduction, the image PSNR was improved to 38.5dB and SSIM reached 0.982. Compared with the unreduced image, the defect edges were fully represented and the background noise was almost eliminated. This ensures that the input data for subsequent normalization, feature extraction and other steps are high-purity detection images without obvious interference, thereby reducing the false judgment rate and improving the robustness of the AOI detection process.

[0065] S2.2: Perform an adaptive outlier detection and interpolation correction algorithm on the process parameter timing data (including welding temperature, humidity, mounting accuracy and process setting values, etc.), and use sliding window statistics or multivariate timing consistency verification mechanism to automatically remove or correct outliers and output the outlier-corrected process parameter timing data.

[0066] S2.3: Based on the noise-suppressed detection image data obtained in S2.1, linear normalization or Z-Score normalization algorithms are used to uniformly scale the pixel values ​​of each channel to a preset range to eliminate the inconsistency in the units of the detection image data and obtain normalized detection image data with uniform scale.

[0067] S2.4: For the time series data of process parameters after anomaly correction output by S2.2, use the minimum-maximum normalization or standard fraction normalization method to uniformly process different physical quantities (such as temperature, humidity, position, etc.), and transform the process parameter data into a dimensionless standardized parameter sequence to adapt to the subsequent feature space fusion requirements.

[0068] S2.5: Based on the timestamps recorded in the detection cycle, perform time series synchronization and alignment processing on the normalized detection image data in S2.3 and the standardized process parameter time series data in S2.4. Using linear interpolation registration or dynamic time warping (DTW) algorithm, output the timestamp-aligned multimodal feature input group according to the common detection batch index, so as to achieve a precise one-to-one correspondence between process conditions and detection results at each moment, and provide a strictly time-consistent data foundation for the multimodal feature extraction stage.

[0069] Step S3: Based on the normalized detection image data and process parameter time-series data, a joint feature vector reflecting dynamic changes in the process is extracted using feature alignment and attention mechanisms to take into account the multimodal characteristics of both process and detection information. For example... Figure 2 As shown, it specifically includes:

[0070] S3.1: The normalized detection image data and process parameter time series data are subjected to feature dimension calibration processing. The standard feature alignment algorithm is used to map the feature dimensions, sampling intervals and time series labels of different data modalities one by one to obtain the feature alignment matrix, which provides the input basis for the consistency processing of multimodal feature space.

[0071] S3.2: Based on the feature alignment matrix, perform deep representation processing of the feature encoder, use convolutional neural networks (CNN) to extract spatial features from the detection image data, and use temporal recurrent networks (RNN / LSTM) to perform dynamic state modeling on the temporal data of process parameters to obtain detection spatial feature vectors and process temporal feature vectors.

[0072] Input normalized detection image data and standardized process parameter time series data after feature dimension calibration. The feature alignment matrix is ​​used as an index to indicate the feature mapping and sampling time series relationship of different modalities.

[0073] A convolutional neural network (CNN) (parameters: 16 layers, 3×3 kernel, ReLU activation function, 2×2 max pooling kernel, batch normalization, Dropout ratio 0.3) is used to automatically extract high-dimensional spatial features from detection image data. The input is a normalized multi-channel detection image. Through multi-level convolution and pooling operations, spatial features of edges, textures, structures, and defect regions are extracted step by step. The output is a high-dimensional feature vector after flattening the spatial features.

[0074] Furthermore, a temporal recurrent neural network (RNN) or a long short-term memory network (LSTM) (parameters: unidirectional LSTM, hidden layer dimension 64, input sequence length consistent with the sampling period, activation function tanh, dropout rate 0.2) is used to dynamically model the calibrated process parameter time-series data. The input is the standardized process parameter sequence within each sampling period, which is progressively fed into the LSTM unit after embedding mapping, continuously capturing and storing the temporal dependency features of parameter changes, forming a temporal state feature vector reflecting the dynamics of process operation.

[0075] By employing a feature alignment matrix, the spatial features output by the CNN are correlated with the temporal features output by the LSTM, ensuring a one-to-one correspondence between feature dimensions within the same detection period or sampling time, thus eliminating modality confusion errors.

[0076] By normalizing features, spatial and temporal features are standardized to the same numerical distribution range, which facilitates subsequent multimodal feature fusion and attention weight allocation.

[0077] The above algorithm is used to obtain high-dimensional detection space feature vectors and process time sequence feature vectors, providing an analytical basis for subsequent fusion analysis and improving process sensitivity and defect classification accuracy.

[0078] For example, in the AOI inspection scenario of circuit boards, a 16-layer ResNet-type CNN network is deployed for normalized inspection images with a resolution of 256×256 and 3 channels. The kernel stride is 1, and the length of the feature vector after max pooling is 1024. For normalized time-series data with a sampling period of 10ms for process parameters (soldering temperature, humidity, and mounting accuracy), a unidirectional LSTM network with 64 hidden units is input, and the final aggregated time-series feature vector is output. Cross-validation shows that the discrimination accuracy of the spatial feature vector is 92.5%, and the AUC of the LSTM time-series features for process anomaly detection is improved to 0.97. The two complement each other under the multimodal mapping of the feature alignment matrix, and the total feature dimension of the output is 1088 dimensions. After normalization, it is prepared for attention fusion processing. Validation results show that using the above feature encoding scheme, the model's sensitivity to changes in process parameters is improved by 22%, and the analysis of defect misjudgment and missed detection shows better adaptive ability.

[0079] S3.3: Input the detection space feature vector and process time series feature vector obtained in the previous step into the multimodal fusion module, call the feature cross-calculation algorithm based on the attention mechanism, calculate the correlation weight of different modal features in each time series segment, and adaptively aggregate according to the weight to obtain a cross-modal representative fusion feature vector to enhance process sensitivity.

[0080] The input consists of a normalized detection space feature vector and a process time sequence feature vector. The two sets of features have achieved a strict one-to-one correspondence between the time sequence label and the dimension through the aforementioned feature alignment matrix.

[0081] A multimodal fusion module (parameters: fusion layer type is cross-Attention, number of layers is 2-4, hidden dimension is 512, number of attention heads is 8) is used to concatenate spatial features and temporal features, construct a feature interaction input tensor, and realize the coupling and fusion of information from different modalities.

[0082] Furthermore, an attention-based feature crossover algorithm (parameters: query / key / value linear transformation matrix, softmax activation, residual connection, normalization) is used to calculate attention-related weights on the input detection space feature vector Q and process time series feature vectors K and V. The calculation method is as follows:

[0083]

[0084] Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k Using the key vector dimension, we can achieve correlation scoring of spatial-temporal features under different time segments.

[0085] Furthermore, using the aforementioned correlation weights, the process time-series feature vector and the detection space feature vector are adaptively weighted at the fusion layer to generate a cross-modal representative fused feature vector. This vector contains cross-modal feature weights assigned by Attention for each time segment, reflecting the modulation of the detection feature representation capability by the current process perturbation.

[0086] Furthermore, the fused feature outputs are processed in parallel through a multi-head attention mechanism (parameters: 8 heads, 64 hidden dimensions per head) to extract correlation information from different subspaces, and then converged to generate the final multimodal joint feature tensor. This tensor is highly sensitive to changes in process and image features.

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

[0088] By using a feature cross-validation algorithm based on the attention mechanism, the original data of spatial and temporal features are efficiently converted into a fused representation, thereby achieving adaptive enhancement of the sensitivity of joint features under the background of process disturbances and improving the ability of the analysis model to identify misjudgments / missed detections caused by dynamic process changes.

[0089] For example, in the AOI defect detection application of SMT (Surface Mount Technology) circuit boards, a 1024-dimensional image spatial feature vector output by a convolutional residual network and a 64-dimensional process parameter temporal feature vector output by a long short-term memory network are selected and fused into a cross-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 spatial features, and the key / value matrices K and V come from the process temporal features. In the fusion stage, the weighted output is calculated using the following formula:

[0090]

[0091] In actual implementation, under batch conditions of periodic fluctuations in welding temperature and variations in mounting accuracy, the Attention algorithm assigns high weights to the process parameter features, resulting in an output fused feature tensor with a dimension of 512. Feature contribution analysis shows that the fused features improve the sensitivity to false positives by 23% and the sensitivity to false negatives by 19%. The output cross-modal representative fused feature vector is used by the backend model for process dynamic drift detection and false positive attribution, significantly improving the responsiveness to complex process disturbances and the interpretability of false positive / false negative correlation features. The final model's joint feature AUC score reaches 0.977, and the attention-weighted adaptive feature fusion effectively suppresses the decrease in detection sensitivity caused by process parameter disturbances on actual production line batch data, achieving highly robust analysis output for process-detection dual modalities under variable scenarios.

[0092] S3.4: Perform feature anomaly suppression and temporal reregularization on the cross-modal representative fusion feature vector. By introducing a specific regularization loss function, suppress the noise distribution of abnormal fusion features and maintain the smoothness of feature changes with process disturbances. Finally, output a joint feature vector that conforms to the dynamic distribution characteristics of the process, providing input for subsequent correlation distribution modeling and dynamic drift detection quantization.

[0093] The input is a cross-modal representative feature vector generated by the Attention cross-linking mechanism. The feature dimension includes the numerical components of the aggregation of detection spatial features and process temporal features in each time segment.

[0094] A feature anomaly detection algorithm (parameters: outlier discrimination based on statistical distribution, threshold set to 3 times the standard deviation, and anomaly distribution model is Gaussian mixture model GMM) is adopted to detect and label anomalous feature components in the fused feature vector that are extremely distributed and deviate from the overall trend.

[0095] Furthermore, through feature anomaly suppression processing (parameters: subthreshold outlier replacement algorithm, replacement method is selected as local weighted average or interpolation correction), lossless replacement or soft suppression of identified abnormal features is achieved, and a fused feature set after the abnormal noise is effectively suppressed is obtained.

[0096] Furthermore, a feature temporal smoothing and re-regularization method is adopted (parameters: the sliding window length is 3-5 times the corresponding detection period, and the smoothing strategy is first-order exponential weighted moving average or bidirectional Savitzky-Golay polynomial filtering) to achieve consistency constraints on the feature vector after anomaly suppression in the temporal dimension, improve the smoothness of feature changes, and suppress temporal abrupt changes that are unavoidable due to process disturbances.

[0097] Furthermore, a temporal regularization penalty term is introduced to construct a joint loss function. Through a regularization algorithm, the joint optimization of outlier distribution constraints and temporal stationarity is achieved in the synthetic feature vector space. The joint loss function is defined as follows:

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

[0099] Among them, L task To determine the main loss in case of misjudgment / missed detection, L anom For the characteristic anomaly suppression loss (such as weighted absolute bias), L smooth λ1 and λ2 are the time-series smoothing regularization terms (such as the first-order sum of squared differences), and the regularization weight parameters are λ1 and λ2.

[0100] By optimizing the loss function, a characteristic distribution output that satisfies both abnormal noise suppression and adapts to dynamic process disturbances is obtained.

[0101] Through the above algorithm processing, the high-dimensional representative feature vector obtained after Attention fusion is transformed into a joint feature vector after anomaly suppression and temporal regularization, realizing the feature distribution fluency constraint for dynamic process changes, and inputting high-confidence, multi-scenario adaptable process-detection joint feature data for subsequent correlation distribution modeling and dynamic drift detection steps.

[0102] For example, in the AOI circuit board inspection scenario, the input is a 512-dimensional fused feature output from the Attention module, with a batch sampling length of 200. A Gaussian mixture model (with 3 components and a 3σ discrimination threshold) detects approximately 2.6% of the feature components as anomalous noise, which is then replaced by local interpolation weighted correction (sampling window 5). A sliding window exponential weighted averaging (window width 20) is used to temporally smooth the anomaly-suppressed feature sequence. The loss function weights are set to λ1 = 0.4 and λ2 = 0.6, and the model is trained end-to-end using the Adam optimizer. Results show that the root mean square of temporal mutations in the features is reduced by approximately 57%, the anomaly distribution probability is reduced to 0.5%, and the fluency of the joint features is significantly improved. Under conditions of periodic perturbations in process parameters across multiple batches, the feature vectors processed in this step are used for backend correlation modeling, resulting in a 2.9% improvement in the false positive detection AUC and a 4.2% decrease in the false negative rate, effectively ensuring the stability and robustness of the subsequent model output under dynamic process environments.

[0103] S3.5: Perform feature importance measurement and redundant feature screening on the output joint feature vector. Use feature contribution analysis algorithms (such as Shapley value or information gain evaluation) to select the set of process-related sub-features that have the greatest influence on the judgment of false positives and false negatives. This ensures that the final joint feature vector has high discriminative power and high interpretability, and provides accurate feature basis for the robustness and interpretability optimization of subsequent process parameter-detection result modeling.

[0104] Step S4: For multiple detection point locations and process step labels, establish a process parameter-detection result correlation distribution model, and group the sub-models to correspond to different manufacturing areas and process scenarios. For example... Figure 3 As shown, it specifically includes:

[0105] S4.1: Based on the joint feature vector obtained in step S3, aggregate the input data according to the physical location of the detection points and the process label. Use the spatial distribution mapping method to form a grouped feature set for each detection point and its corresponding process label to support the fixed-point causal analysis between subsequent process parameters and detection output.

[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] For example, in an AOI inspection production line for SMT assembly, 32 inspection points are set on the PCB board. The locations of the inspection points are imported via CAD coordinates, with a spatial coordinate resolution of 0.01mm for each point. Each inspection point is accompanied by a process step label in the inspection batch, covering processes such as surface mount, reflow soldering, and wave soldering. In actual operation, the batch to which the inspection data belongs covers 2 hours, with a data volume of 12,800 sets. A spatial indexing algorithm is used to group adjacent inspection points <3mm on the same PCB. Process step label matching is used to further group all inspection points tagged with "reflow soldering" into "reflow soldering groups". After spatial-process double-layer clustering, 12 group feature sets are formed, each with a capacity of 700-1200 joint feature vectors. A Gaussian radial basis distribution model is used to perform spatial distribution mapping on each group of feature blocks. The spatial distribution map clearly reflects the high-incidence and low-incidence areas of defect distribution under the "reflow soldering" process. The output group feature sets enable efficient retrieval of each process step-inspection point block in the database. Actual results show that this grouping method provides a fixed-point basis for subsequent causal drift modeling, significantly improves the response capability to regional process anomalies, increases spatial aggregation efficiency by 37%, and achieves 100% coverage of grouping features.

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

[0109] The input is a set of joint feature vectors obtained by grouping in step S4.1. Each group covers the physical location of the detection point, the process label, and normalized fused multimodal feature information.

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

[0111] Furthermore, using the GMM modeling method, the process parameter sub-vectors X in the grouped feature set are... proc With the detection result subvector Y det As input, fit the joint probability density function between the two, according to the following model:

[0112]

[0113] Where K is the number of mixture components, π k Let k be the weight of the k-th component. The mean is μk The covariance is ∑ k It follows a multivariate normal distribution.

[0114] Furthermore, the EM (Expectation-Maximization) algorithm is used for parameter estimation, with the E-step calculating the posterior probability and the M-step updating the mean μ. k , covariance ∑ k Weight π k The optimization is iterated until the log-likelihood converges or the maximum number of iterations is reached.

[0115] Furthermore, for high-dimensional features and sample size, the variational Bayesian inference algorithm is invoked to adaptively regularize the Gaussian prior parameters and automatically select the optimal number of mixture components to prevent overfitting and improve the model's generalization ability.

[0116] By modeling the joint probability distribution as described above, a joint distribution model of process parameters and detection results is output for each group feature set. The model parameters include component weights, group mean, covariance matrix, etc.

[0117] Furthermore, by inferring the conditional probability of the joint distribution, a quantitative assessment of the probability of abnormal detection results (false positives / false negatives) under specific values ​​of process parameters (such as temperature, humidity, mounting accuracy, etc.) can be obtained.

[0118] By jointly probabilistically modeling multiple sets of grouped features, spatial-process full-coverage correlation modeling of process parameters-detection results under multiple detection points and multiple process scenarios is realized, providing a high-resolution probabilistic foundation for subsequent steps such as process scenario sub-model division, feature importance extraction, and dynamic drift detection.

[0119] For example, for spatial proximity detection points in the "reflow soldering" process of an SMT assembly line, 1200 sets of joint feature vectors containing process parameters (soldering temperature range 230-250 degrees Celsius, line humidity 30-55% RH, placement accuracy offset 0-0.05mm) and detection outputs (false positive / false negative labels) were selected as samples. The GMM algorithm was used, with a component upper limit K=4 and an initialization strategy of K-means++. The EM algorithm had a maximum of 250 iterations per round, and the covariance matrix was fully standardized. During actual operation, the number of components was automatically adjusted to 3, and after convergence, the mixed weights [0.42, 0.36, 0.22] were output. The mean and covariance of each component clearly separated the coupling mode between process parameters and detection results. For the joint distribution model, based on the conditional probability formula, the conditional probability of the false positive rate increased to 0.083 in the region of welding temperature < 235℃ and humidity > 48%RH, while the conditional probability of the false positive rate decreased to 0.024 in the process parameter range (240-245℃, humidity 38-45%RH). Further variational Bayesian inference was used to model a large-capacity sample group (mounting process group, 3500 cases), ultimately retaining the optimal mixture component 2 to prevent overfitting under high-dimensional features. The AUC of the model's sensitivity to the process parameter range increased to 0.961. Finally, the probability distribution model parameters for each group were output, realizing spatial-process joint distribution modeling of process parameters and detection results. This provides a solid foundation for sub-model division and subsequent drift detection, achieving quantifiable, inferable, and traceable technical effects on the impact of dynamic process parameter perturbations on detection performance.

[0120] S4.3: The above process parameter-test result association distribution model is divided into sub-models based on the test point area attributes and process operation labels. A hierarchical grouping logic is used to construct a process area / process scene mapping sub-model library to ensure that each manufacturing area or process scene has a dedicated modeling object.

[0121] The input consists of the joint feature set of each group obtained through step S4.2 and its joint probability distribution model of process parameters and detection results. The group information covers the physical area attributes of the detection points and the process operation labels.

[0122] A dual-index grouping algorithm using region labels and process label (parameters: region division granularity 1-5mm, process label category covers all manufacturing process types) is adopted to achieve spatial-process dual-layer classification of the entire grouped feature set.

[0123] Furthermore, by using a regional attribute comparison method (parameters: PCB board regional marking system, automatically matching the XY coordinates of the detection point with the design process block), each group feature set is assigned its physical manufacturing block, thereby achieving grouping and classification within the region.

[0124] Furthermore, based on the consistency criterion of process operation labels (parameter: process operation homogeneity threshold 0.9), the group feature sets of the same region and the same process label are aggregated, and process scenario subsets such as reflow soldering area, surface mount area, and wave soldering area are established according to the process flow distribution.

[0125] Furthermore, a hierarchical grouping logic is adopted (parameter: priority is set to region first, then process step) to further refine the above preliminary grouping into fine-grained subclasses. For areas with dense inspection points and areas where multiple processes intersect, an independent sub-model partitioning strategy is output to ensure that areas with overlapping physical spaces and process intersections have dedicated modeling objects.

[0126] Furthermore, through the sub-model instantiation algorithm, a joint probability distribution sub-model of process parameters and detection results is instantiated for each group (a unique combination of sub-region and process label), and the model object is added to the process region / process scenario mapping sub-model library, thus realizing the structured management and index archiving of the sub-model library.

[0127] Through the above hierarchical grouping and sub-model instantiation process, the grouping features and joint probability distribution model are transformed into a hierarchical sub-model group that is dually mapped by regional attributes and process labels. Each manufacturing region or process scenario is configured with a unique corresponding modeling object, which effectively realizes the scenario refinement and model specialization of process parameter-test result correlation modeling, and provides basic support for the analysis robustness to cope with manufacturing diversity and process disturbances.

[0128] For example, in an SMT production line AOI inspection management process covering surface mount technology (SMT), reflow soldering, and wave soldering, the PCB board is assumed to be divided into 4 physical inspection areas with a granularity of 2mm. 32 AOI inspection points belong to these 4 physical areas, each area sequentially undergoing SMT, reflow soldering, and wave soldering process labels. The total input data is 12,800 sets, which are combined with the area labels and process labels to form 12 initial groups. Through hierarchical grouping logic, inspection points with overlapping physical spaces (such as boundary and channel intersections) and multiple process intersections (such as dual-process repair interface areas) are output as separate groups. Using the above grouping results, a Gaussian mixture joint distribution sub-model instance is established for each group (with an average sample size of 800-1200 sets per sub-model). Finally, the sub-model library contains 16 independent sub-model instances, covering all physical inspection areas and process scenarios. Compared with the ungrouped single model, the hierarchical grouped sub-model improves the false positive rate identification AUC by 7.6%, the false negative sensitivity by 12.4%, and the anomaly detection capability at boundary areas and process intersections by 20%, achieving the expected technical effects of model specialization and analysis refinement under the diversity of manufacturing zones / process scenarios.

[0129] S4.4: The performance of each sub-model of the process parameter-detection result correlation distribution model is evaluated and the importance of features is ranked. Based on statistical validity criteria (such as mutual information and information gain), the process parameter subspace that has a significant impact on the false positive rate and false negative rate is selected, providing a refined feature reference for subsequent dynamic drift detection and model adaptation mechanism.

[0130] S4.5: Generate a unique mapping index between each sub-model and the corresponding process scenario and detection point area, and link this index structure to the time-series process parameter tracking database to realize closed-loop management of data flow, model flow and process flow, and ensure the dynamic switching of subsequent sub-models, traceability of anomaly tracing paths and data consistency.

[0131] Step S5: Based on the joint feature vector, the process parameter-detection result correlation distribution model is input into the dynamic drift detection algorithm to calculate the false positive rate and false negative rate statistics of the current batch and historical batches, and to identify abnormal drift in model output caused by process parameter disturbances. Specifically, this includes:

[0132] S5.1: For the input joint feature vector sequence and the process parameter-detection result correlation distribution model, the data is grouped using batch identifiers and timestamps, and the detection data is divided into current detection batches and historical detection batches to establish a batch-level statistical basis dataset for false positive rate and false negative rate.

[0133] S5.2: For each batch-level joint feature vector, execute a classification and judgment algorithm based on the correlation distribution model, and calculate the false positive rate and false negative rate of the current detection batch and the historical detection batches respectively. Output batch-level false positive rate statistics and false negative rate statistics to provide parameter basis for subsequent dynamic analysis.

[0134] The input is a sequence of joint feature vectors grouped by batch. The features include normalized and time-aligned detection spatial features and time-series features of process parameters, as well as batch-level process parameter-detection result correlation distribution model instances.

[0135] A batch-level classification algorithm is adopted (parameters: each batch selects a corresponding distribution model instance, the model type can be Gaussian Mixture Model (GMM) or Variational Bayesian Distribution; the classification threshold is dynamically set according to the confusion matrix of the training data) to determine the defect category of the input joint feature vector and output the determination label (such as "qualified", "false positive", "missed detection" etc.).

[0136] Furthermore, based on the above classification results and the joint feature vectors of the current detection batch, the number of misclassified samples at the batch level is counted. Actual total number of negative samples Number of missed samples and the actual total number of positive samples Achieve batch-based quantitative analysis of defect identification performance.

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

[0138]

[0139] Where, N FP N represents the number of samples in this batch that were incorrectly identified. N N represents the total number of actual negative samples (non-defective) in this batch. FN N represents the number of missed samples. P This represents the total number of actual positive samples (defects).

[0140] For historical testing batches, the same classification and statistical scheme is used to obtain the false positive rate and false negative rate statistics for each historical batch, forming a historical comparison dataset.

[0141] By classifying and extracting statistics at the batch level, the false positive rate (FPR) for the current batch and historical batches is obtained. curr ,{FPR hist,i} and false negative rate (FNR) curr ,{FNR hist,i Batch-level performance indicators, such as those mentioned above, establish a complete data foundation for subsequent dynamic drift detection and anomaly root cause analysis.

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

[0143] For example, in a reflow soldering process scenario, the input batch is production batch number 20240430_X001. The batch inspection data contains 1280 sets of joint feature vectors. The corresponding distribution sub-model adopts a Gaussian mixture probability model (number of components K=3). Based on the judgment algorithm, the inspection and discrimination labels are divided into: 1160 qualified pieces, 60 misjudged pieces, 20 missed pieces, and the remaining 40 pieces with real defects. Statistically, the actual negative samples of the current batch... Positive samples Misjudged samples Missed samples Using the above formula, we can calculate:

[0144]

[0145] The historical comparison dataset selected the judgment results of the same batches and blocks using the same process from the past 10 batches. The mean value of {FPR_hist,i} was 0.0381, and the mean value of {FNR_hist,i} was 0.2947. Therefore, the false positive rate and false negative rate of this batch are higher than the historical averages, providing an objective statistical benchmark for subsequent dynamic drift detection and process tracing analysis. The batch-level FPR and FNR statistics and their time series are output, achieving the expected technical effect of batch-level quantitative evaluation of detection performance under dynamic process disturbances.

[0146] S5.3: Integrate the batch-level misclassification rate statistics and missed detection rate statistics of the current batch and historical batches, and use the sliding window method and the mean and variance calculation scheme to dynamically generate the misclassification rate drift feature sequence and the missed detection rate drift feature sequence, so as to realize the summarization of the batch time series statistical characteristics.

[0147] The input consists of batch-level false positive rate (FPR) and false negative rate (FNR) statistics generated for the current testing batch and historical testing batches, respectively. These statistics are extracted using a classification algorithm and an association distribution model and are archived according to the testing batch timestamp and batch number.

[0148] The sliding window method (parameter: window length w is dynamically set according to the actual number of batches on the production line, w = 5–10 is recommended) is used to achieve time-series local analysis of the historical batch false positive rate and false negative rate sequence.

[0149] Furthermore, using a moving average calculation algorithm (parameters: weighted moving average for mean update, historical sequence weighting factor α = 0.6–0.9), the moving average μ of the false positive and false negative rates for each window within batch t to t-w+1 is calculated. FPR (t), μ FNR (t) reflects the central trend of the model output under process dynamic disturbances.

[0150] Furthermore, using a flow variance calculation algorithm (parameter: variance recursive update method), the flow variance is calculated in real time for the false positive rate and false negative rate sequence within the window from batch t to t-w+1. The intensity of fluctuations in the quantification of false positives and false negatives with batch drift.

[0151] The following recursive formula is used to calculate the moving average:

[0152]

[0153] Where, μ X (t) represents the mean value within the window up to the current batch t, ​​X t-i Let be the statistics for the ti-th batch, and w be the width of the sliding window.

[0154] The flow variance recursive formula is used:

[0155]

[0156] in, X represents the variance within the current batch window t, where X refers to FPR or FNR.

[0157] Furthermore, by using the time-series recursion of the above mean and variance, a false positive rate drift feature sequence is dynamically generated. With the drift characteristic sequence of the false negative rate The system reflects the trend and fluctuation distribution of the impact of changes in process status on the statistical values ​​of model misjudgment and missed detection as production batches progress.

[0158] By using differential or normalization processing, the above drift feature sequences are transformed into batch-to-batch drift metrics, which facilitates subsequent dynamic drift detection algorithms for significance discrimination and anomaly detection.

[0159] By using sliding window and flow statistics processing, the historical and current data of batch-level false positive and false negative rates are transformed into time-series drift characteristic indicators that reflect the dynamic evolution of process disturbances and model performance, thus achieving the goal of batch statistical inductive analysis of the detection model output under dynamic changes in process parameters.

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

[0161] Set the sliding window length w = 5, and for the t = 10th batch, use the FPR sequence of the first 5 batches.

[0162] Calculate the moving average of {0.036, 0.039, 0.045, 0.049, 0.041}:

[0163]

[0164] Calculate the flow variance using the same set of data:

[0165]

[0166] The sliding window process was repeated for all batches, dynamically outputting the drift feature sequence. Actual testing revealed that the mean FPR drift of some batches (such as B015) increased temporarily, with a significant increase in variance, indicating that process disturbances caused batch-specific statistical drift in the model output. This provided high-resolution drift feature input for subsequent anomaly detection and model adaptation. Ultimately, this step significantly improved the ability to summarize the time-series trends of false positives and false negatives and the batch anomaly response capability, laying the foundation for dynamic robustness optimization of the model under complex process scenarios.

[0167] S5.4: For the generated false positive rate drift feature sequence and false negative rate drift feature sequence, use a dynamic drift detection algorithm (such as CUSUM, Page-Hinkley or Bayesian drift identification mechanism) to compare the significant differences of drift features in the current batch and historical intervals, and determine whether there is a statistically abnormal drift phenomenon in the model output.

[0168] S5.5: When the dynamic drift detection algorithm determines that there is statistical anomaly drift, it performs correlation analysis between the detected abnormal batch and the corresponding process parameter disturbance record, and generates a pairing report of abnormal drift event and associated process parameter range, which serves as the input basis for model adaptive optimization and cause analysis process.

[0169] Step S6: When an abnormal drift in the model output is detected and correlated with a specific process parameter range, the retraining or incremental learning process of the analysis model is automatically triggered to adapt to the detection state under the specific process parameters. Specifically, this includes:

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

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

[0172] The input consists of a process parameter anomaly range mapped by the dynamic drift detection algorithm, batch-level normalized detection images, and a joint feature vector dataset of process parameters. A joint sample pooling method (parameters: target process range number, historical drift marker, batch index range) is used to dynamically select all joint feature samples within the current abnormal process parameter range, including spatial features of the detection image and temporal features of the corresponding process parameters. Further, a temporal distribution sampling algorithm (parameters: window length w, drift significance threshold δ, minimum historical / current batch sample size n_thr) is used to randomly / stratifiedly extract representative samples within the current process range, ensuring consistency in the distribution of process parameter disturbance states, model output drift degree, and detection labels (qualified, false positive, missed detection, etc.) covered by the samples. A floating sampling weight allocation method for process parameters is adopted; for ranges where the rate of change of process parameters is higher than the average drift level, the sampling weight is increased to enhance the representativeness of the samples in this range, resulting in a weighted sample distribution. Furthermore, by comparing the historical and current process parameter perturbation levels, a matching sampling method is used to supplement the newly generated training sample set with several representative samples from historical process intervals. This ensures that the new sample pool reflects both the current drift process characteristics and the historical robustness of the model. Through sample consistency discrimination algorithms (such as the KS-test or KL divergence assessment), it is confirmed that the selected sample set satisfies the representativeness of the target process interval and historical drift intervals in terms of feature distribution and label distribution. Samples with abnormal distributions 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, detection samples under current and historical typical process parameter intervals are scientifically incorporated into the training set, and the distribution and representativeness of the samples are strictly controlled, achieving reasonable selection of incremental training base samples and enhanced drift sensitivity.

[0173] For example, in the SMT placement process, an increase in the average FPR of batch 20240507_B015 was detected, and the process parameter range [temperature: 235-245℃, humidity: 40-45%] was determined to be an abnormal drift range. The target parameters were w=8, δ=0.05, and n_thr=500, with a total of 1280 sets of joint feature samples within the sampling window. A stratified sampling algorithm was used to extract 640 sets of samples from the batches with significant drift within the current abnormal process range. An additional 320 sets of samples were collected from the historically stable process range (temperature: 230-235℃) based on the historical misjudgment rate weight. The remaining 320 sets were selected from the overlapping transition range based on the process change rate weighting. All samples were labeled with detection judgment tags, and the KS value of the sample set was verified to be less than 0.1, and the KL divergence was lower than the set threshold of 0.02, confirming the consistency of the distribution. Ultimately, a joint feature sample set of process parameters and detection images covering the current drift process parameter range and historical representative states was formed, providing a high-quality training foundation for the incremental training of the subsequent analysis model and the improvement of its ability to trace the causes of misjudgments. The application results show that after incremental learning, the model's FPR for this abnormal process parameter range decreased by 14.7% and FNR decreased by 8.2%, effectively suppressing detection anomalies caused by dynamic process disturbances.

[0174] S6.3: For the selected process parameter-detection image joint feature sample set, the model parameters are updated using a multimodal feature incremental learning framework. By integrating new process interval features, the model's adaptive decision-making ability to dynamic process change conditions is enhanced.

[0175] S6.4: During the incremental learning process of the model, compare the changes in the sensitivity of process parameters in the joint feature vector space before and after the update, and use a robustness evaluation algorithm to analyze the improvement of the statistical indicators of false positive rate and false negative rate after retraining, so as to ensure the closed loop of the effect under the process conditions.

[0176] S6.5: Based on the corrected analysis model output and its adjustment history, generate an analysis model parameter change log in real time, and mark the model adaptability status under the current process parameter range, providing a traceable chain of evidence for subsequent cause analysis and further optimization.

[0177] Step S7: An interpretable AI algorithm is used to perform causal reasoning on the historical changes in process parameters, joint feature weights, and the impact of model drift, generating a ranking list of dominant process factors for changes in false positive and false negative rates for each batch, to indicate the causal path. Specifically, this includes:

[0178] S7.1: Synchronously pair the historical process parameter variation sequence with the time series statistics of detection misjudgment rate and missed detection rate, and use time series association analysis methods (such as Granger causality test) to obtain a preliminary set of causal relationships between the dynamic changes of process parameters and the trend of model output, so as to extract the main control correlation factors of the history of process parameter variation.

[0179] S7.2: Using the history of process parameter changes as input, combined with the joint feature weight distribution, call interpretable AI algorithms (such as SHAP value decomposition or LIME local interpretation) to calculate the marginal contribution of multidimensional process parameter features in the model's misclassification rate and false negative rate output, so as to obtain the causal influence mapping between joint feature weights and process parameters.

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

[0181] The input data consists of the set of abnormal drift events identified by the model drift detection algorithm, the statistical vector of the marginal contribution of process parameters to detection output, the normalized historical variation sequence of process parameters, the joint feature weight distribution, and the time series indicators of the false positive rate and false negative rate of each detection batch.

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

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

[0184] Furthermore, a causal graph modeling method is adopted to construct a directed causal relationship graph G=(V,E) based on the history of process parameters and detection output labels of abnormal drift batches. The node V includes P^* and false positive / false negative output, and the edge E represents a significant causal relationship.

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

[0186] Using a maximum a posteriori inference mechanism, probabilistic reasoning is performed on the critical path of the causal graph G to identify the set of key process parameter factors Q that cause model output drift, and the conditional probability contribution of each process parameter to the occurrence of the drift event is output:

[0187]

[0188] Among them, P(Drift) FPR / FNR |P i ) indicates that in process parameter P i Given conditions, the conditional probability of misjudging or missing abnormal drift events.

[0189] By integrating the conditional probability ranking list of key process parameter Q, a principal factor contribution matrix is ​​generated to indicate the causal path.

[0190] Through the above algorithm, the input abnormal drift detection results, marginal contribution of process parameters, and historical variation data of process parameters are transformed into a causal link of process parameter-model output drift, realizing multi-factor tracing analysis.

[0191] For example, for SMT placement batch 20240507_B015, the dynamic drift detection algorithm determines that at batch index B_{drift}=B015 and FPR drift significance δ=0.042, the corresponding process parameter range [temperature: 237℃, humidity: 43%] is extracted. Based on SHAP marginal contribution calculation, a subset of process parameters with a contribution greater than 0.17 is selected: temperature P1, humidity P2, and placement accuracy P3. Using historical process parameter records and the current FPR / FNR sequence, a causal relationship graph G is constructed. A Bayesian network structure learning algorithm is used, with the structure score function S = Bayesian Information Criterion (BIC) and a prior confidence β=0.99, to train the probabilistic network.

[0192] Using maximum a posteriori probability inference, the conditional probabilities of the master process parameters under FPR drift are calculated, where:

[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 chain indicates that temperature is the primary controlling factor, and humidity is a secondary influencing factor. The final generated list of dominant process parameter factors by weight is Q = [P1(0.77), P2(0.65), P3(0.22)], indicating the causal path of FPR anomaly drift. Application results show that this causal reasoning step clearly distinguishes and ranks the main process parameters affecting the drift event, providing strong data support for subsequent model adaptive adjustment and manufacturing parameter optimization.

[0197] S7.4: Weighted aggregation of process parameter variation history, joint feature weight decomposition results and model drift causal chain for different time series batches, and weighted sorting and normalization processing algorithms are applied to generate a list of dominant process factor weights for the changes in false positive rate and false negative rate of each batch, so as to ensure the consistency and comparability of the sorting results in multiple batches and multiple distribution scenarios.

[0198] S7.5: Based on the list of dominant process factors and using an interpretable AI algorithm recommendation mechanism, automatically generate the cause-finding path structure for each batch, including a subset of key process parameters and their logical links to changes in false positive and false negative rates, clearly indicating the process parameters that need to be prioritized and adjusted, serving as the output of cause-finding analysis and the input basis for subsequent model adjustments.

[0199] Step S8: Based on the ranking list of dominant process factors, dynamically adjust the parameter structure of the analysis model. Specifically, this includes using targeted sub-models or changing the weights of the loss function when switching between process parameter ranges, to achieve adaptive optimization of the analysis model under multiple process scenarios. Specifically, this includes:

[0200] S8.1: Based on the ranking list of dominant process factors output by the interpretable AI algorithm, extract the set of dominant influencing factors corresponding to each process parameter range as input conditions for model adjustment, and clarify the mapping relationship between each process parameter range and the trend of false positive and false negative rates through correlation analysis.

[0201] S8.2: For the process parameter range corresponding to the set of dominant influencing factors, based on the process parameter-detection result correlation distribution model, automatically retrieve the currently recommended targeted sub-model structure to optimize the adaptive response capability of the model under the process parameter range and realize model switching output for specific process environments.

[0202] S8.3: Based on the already switched to the target sub-model structure, for the boundaries or overlapping areas of the process parameter range, the loss function weight self-adjustment strategy is applied to dynamically update the model loss function weight parameters, thereby optimizing the output of the ability to balance misjudgment and missed detection in the transition range.

[0203] Within the process parameter range that has been switched to the target sub-model structure, the input includes the process parameter data of the current test batch, the joint feature vector, and the list of dominant process factor weights.

[0204] A 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 statistical characteristics of the detection batch in the transition interval or overlapping area.

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

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

[0207] L = β FPR ·L FPR +β FNR ·L FNR

[0208] Where L is the total loss function, L FPR To avoid misjudging the loss component, L FNR This represents the loss due to missed detection.

[0209] Furthermore, through a dynamic parameter optimization strategy (parameters: Q = [q_1, q_2, ..., q_n]), the effect of each component of the dominant process parameter on β is analyzed in real time. FPR and β FNR The influence weights are adjusted adaptively as follows:

[0210]

[0211] Where γ is the step size coefficient, ΔFPR i With ΔFNR i These represent the marginal contributions of the i-th master control process parameter to the changes in false positive rate and false negative rate in the overlapping interval.

[0212] Apply Monte Carlo sampling or adaptive backpropagation algorithms to update the loss weight parameters of the analysis model in batches, determine the convergence of the loss curve, and output the optimal loss weight configuration.

[0213] By using the above processing method, the model weights when the current process parameters are in overlapping intervals or at critical values ​​are transformed into highly adaptive loss function parameters, thereby achieving a trade-off optimization between false positive rate and false negative rate in the process transition interval, significantly improving the robustness and accuracy of the model in complex dynamic process environments.

[0214] For example, in SMT soldering batch 20240507_B037, the current process parameter temperature range [235℃, 240℃] and humidity range [42%, 46%] overlap. The batch detection data in the overlapping area is input into a targeted sub-model with a weighted list of dominant process factors Q = [temperature (0.80), humidity (0.62), placement accuracy (0.18)]. Based on the dynamic loss function weight adjustment algorithm, β_{FPR} = 0.5, β_{FNR} = 0.5, and step size coefficient γ = 0.05 are initialized. According to the marginal contribution of each dominant process parameter FPR and FNR 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 weights, the loss convergence curve stabilized, and the FPR in the overlapping region was reduced to 3.6%, while the FNR was controlled at 2.4%. This performance was better than that without weight adjustment (FPR 5.1%, FNR 4.0%), achieving optimal detection performance in the process transition range.

[0217] S8.4: Based on dynamic process parameter tracking records, continuously monitor the changes in process parameter ranges, and determine the timing for adaptive adjustment of current model parameters or structure by linking with historical model performance statistics (such as batch misjudgment rate, mean / variance of missed detection rate), and output the optimal model structure or weight configuration.

[0218] S8.5: An incremental learning mechanism is adopted to perform linkage verification between the model output after each model switch or loss function weight adjustment and the real-time process parameters. By feeding back residual scalars or drift statistics, the adaptive effect of the analysis model in the new process parameter range is further corrected and analyzed to obtain the optimized process parameters-model structure-detection result coupled matching output.

[0219] Step S9: The results of dynamic misjudgment root cause analysis and process parameter tracking records are fed back to the manufacturing execution system in real time, achieving closed-loop process optimization suggestions and correction of detection parameter configurations. Specifically, this includes:

[0220] S9.1: Based on the ranking results of the dominant process factors output by the causal reasoning module, obtain the dynamic misjudgment tracing analysis results and process parameter tracking records of the current batch in order to prepare standardized feedback messages.

[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 method for analyzing AOI inspection results of circuit boards based on intelligent algorithms, characterized in that, Includes the following steps: S1: Collect a synchronous time-series data set including detection image data and multi-dimensional process parameters as input data for subsequent feature processing; S2: Denoise, normalize, and align the collected detection image data and process parameter time series data respectively; S3: Based on normalized detection image data and time series data of process parameters, a joint feature vector reflecting the dynamic changes of the process is extracted using feature alignment and attention mechanisms; S4: For multiple detection point locations and process operation labels, establish a process parameter-detection result correlation distribution model, and group the sub-models to correspond to different manufacturing areas and process scenarios. S5: Based on the joint feature vector, the process parameter-detection result correlation distribution model is input into the dynamic drift detection algorithm to calculate the false positive rate and false negative rate statistics of the current batch and historical batches, and to identify abnormal drift in model output caused by process parameter disturbances. S6: When an abnormal drift in the model output is detected and correlated with a specific process parameter range, the retraining or incremental learning process of the analysis model is automatically triggered to adapt to the detection state under the specific process parameters. S7: Use interpretable AI algorithms to perform causal reasoning on the history of process parameter changes, joint feature weights and model drift effects, and generate a ranking list of the dominant process factors for changes in the false positive rate and false negative rate of each batch. S8: Dynamically adjust the parameter structure of the analysis model according to the sorting list of the dominant process factors.

2. The method for analyzing AOI inspection results of circuit boards based on intelligent algorithms according to claim 1, characterized in that, Following step S8, the following is also included: S9: Real-time feedback of dynamic misjudgment root cause analysis results and process parameter tracking records to the manufacturing execution system, realizing closed-loop process optimization suggestions and detection parameter configuration correction.

3. The method for analyzing AOI inspection results of circuit boards based on intelligent algorithms according to claim 1, characterized in that, Step S1 specifically includes: Configure the AOI inspection equipment for the circuit boards to be inspected so that each circuit board in the inspection batch can have its original inspection image data acquired through a high-resolution image acquisition interface and stored in a standardized image data format when it is inspected. Configure multi-channel sensor nodes in the production line process parameter acquisition unit to collect process parameters synchronized with the test batch in real time; Based on the main control platform of the AOI inspection system, a unified clock synchronization protocol is used to perform high-precision synchronization encoding of the timestamp of the AOI inspection image and the acquisition timestamp of the process parameters. The batch number and metadata of the synchronously acquired detection image data and process parameter data are registered, and the acquired batch information and process parameter association table are written into the database in real time through the manufacturing execution system interface. After data acquisition is completed, data integrity verification and packet loss detection algorithms are executed. For damaged or missing detection image data and process parameter data, data reacquisition or anomaly marking is performed through a predefined retry mechanism.

4. The method for analyzing AOI inspection results of circuit boards based on intelligent algorithms according to claim 3, characterized in that, The process parameters include welding temperature, production environment humidity, mounting process accuracy, and process setting values.

5. The method for analyzing AOI inspection results of circuit boards based on intelligent algorithms according to claim 1, characterized in that, Step S3 specifically includes: The normalized detection image data and process parameter time series data are subjected to feature dimension calibration. The standard feature alignment algorithm is used to map the feature dimensions, sampling intervals and time series labels of different data modalities one by one to obtain the feature alignment matrix. Based on the feature alignment matrix, deep representation processing of the feature encoder is performed. Convolutional neural networks are used to extract spatial features from the detection image data, and a temporal recursive network is used to perform dynamic state modeling on the temporal data of process parameters to obtain detection spatial feature vectors and process temporal feature vectors. The detection space feature vector and the process time sequence feature vector are input into the multimodal fusion module, and the feature cross-calculation algorithm based on the attention mechanism is called to calculate the correlation weight of different modal features in each time segment. The cross-modal representative fusion feature vector is obtained by adaptive aggregation according to the weight. The cross-modal representative fused feature vector is subjected to feature anomaly suppression and temporal reregularization. By introducing a specific regularization loss function, the noise distribution of abnormal fused features is suppressed, and the smoothness of feature changes with process disturbances is maintained. Finally, a joint feature vector that conforms to the dynamic distribution characteristics of the process is output. The joint feature vector is subjected to feature importance measurement and redundant feature screening. The feature contribution analysis algorithm is used to select the set of process-related sub-features that have the greatest influence on the judgment of false positives and false negatives.

6. The method for analyzing AOI inspection results of circuit boards based on intelligent algorithms according to claim 5, characterized in that, The selection of the process-related sub-feature set in step S3 is specifically as follows: The joint feature vector output after feature fusion is used to select the feature subset that is most sensitive to the misclassification rate and the false negative rate using the Shapley value or information gain feature contribution algorithm, so as to enhance the model's ability to interpret and discriminate process disturbances.

7. The method for analyzing AOI inspection results of circuit boards based on intelligent algorithms according to claim 1, characterized in that, Step S4 specifically includes: Based on the joint feature vector obtained in step S3, the input data is aggregated according to the physical location of the detection points and the process label. The spatial distribution mapping method is used to form a grouped feature set for each detection point and its corresponding process label. Multivariate statistical modeling algorithms are used to perform joint probability distribution modeling of process parameters and test results in each group feature set to obtain the process parameter-test result correlation distribution model. The process parameter-detection result association distribution model is divided into sub-models based on the detection point area attributes and process operation labels, and a process area / process scene mapping sub-model library is constructed using hierarchical grouping logic. The performance of each sub-model of the process parameter-detection result correlation distribution model is evaluated and the importance of features is ranked. Based on the statistical validity criterion, the process parameter subspace that has a significant impact on the false positive rate and the false negative rate is selected. Generate a unique mapping index between each sub-model and the corresponding process scenario and detection point area, and link the unique mapping index to the time-series process parameter tracking database.

8. The method for analyzing AOI inspection results of circuit boards based on intelligent algorithms according to claim 1, characterized in that, Step S5 specifically includes: For the input joint feature vector sequence and the process parameter-detection result correlation distribution model, the data is grouped using batch identifiers and timestamps, and the detection data is divided into current detection batches and historical detection batches to establish a batch-level statistical basic dataset for false positive rate and false negative rate. For each batch-level joint feature vector, a classification and judgment algorithm is executed based on the correlation distribution model. The false positive rate and false negative rate of the current detection batch and the historical detection batch are calculated respectively, and the batch-level false positive rate statistics and false negative rate statistics are output. By integrating batch-level misclassification rate statistics and false negative rate statistics of the current batch and historical batches, and using the sliding window method and the mean and variance calculation scheme, the misclassification rate drift feature sequence and false negative rate drift feature sequence are dynamically generated. For the generated false positive rate drift feature sequence and false negative rate drift feature sequence, a dynamic drift detection algorithm is used to compare the significant differences of the drift features in the current batch and the historical interval to determine whether there is a statistically abnormal drift phenomenon in the model output. When the dynamic drift detection algorithm determines that there is statistical anomaly drift, it performs correlation analysis between the detected abnormal batch and the corresponding process parameter disturbance record, and generates a pairing report of abnormal drift event and associated process parameter range.

9. The method for analyzing AOI inspection results of circuit boards based on intelligent algorithms according to claim 8, characterized in that, The dynamic false positive rate and false negative rate drift detection specifically involves using a combination of sliding window, flowing mean and variance, combined with CUSUM or Bayesian drift identification methods to identify statistical anomalies.

10. The method for analyzing AOI inspection results of circuit boards based on intelligent algorithms according to claim 1, characterized in that, Step S6 specifically includes: The probability distribution of the abnormal model output by the dynamic drift detection algorithm is mapped to the time series characteristics of the normalized process parameters. Based on the process parameter interval mapping results, a joint feature sample set of process parameters and detection images is automatically generated, and a time-series distribution sampling algorithm is used to select representative training samples. For the selected process parameter-detection image joint feature sample set, the model parameters are updated using a multimodal feature incremental learning framework; During the incremental learning process of the model, the changes in the sensitivity of process parameters in the joint feature vector space before and after the update are compared, and the robustness evaluation algorithm is used to analyze the improvement of the statistical indicators of false positive rate and false negative rate after retraining. Based on the corrected analysis model output and its adjustment history, a log of analysis model parameter changes is generated in real time, and the model's adaptability status under the current process parameter range is marked.

Images