Circuit board quality tracing method based on big data

Abstract

The invention relates to a circuit board quality tracing method based on big data, and the scheme comprises the steps: carrying out the high-precision clock synchronization and labeling collection of the original data of a multi-node sensor, carrying out the time sequence normalization and denoising, constructing a multi-source error model, and carrying out the self-adaptive error compensation. And abnormal parameter clustering, process causal link automatic reconstruction and causal reasoning are realized. And finally, in combination with quality problem traceability and graphical visualization as well as model parameter adaptive adjustment of dynamic monitoring driving, precise traceability and efficient early warning of the quality problem in a cross-process and cross-batch scene are realized.

Description

Technical Field

[0001] This invention relates to the field of "big data-driven circuit board manufacturing process quality traceability and causal reasoning analysis technology", and more particularly to a circuit board quality traceability method based on big data. Background Technology

[0002] Currently, big data-driven traceability and causal reasoning analysis of circuit board manufacturing processes have become core technologies for intelligent manufacturing and high-reliability production systems. With the increasing complexity of electronic manufacturing processes and the continuous improvement of product precision and consistency requirements, data acquisition, intelligent analysis, and fault tracing in the manufacturing process are rapidly moving towards a high-dimensional, real-time, and end-to-end traceability technology landscape. Currently, mainstream technical solutions focus on multi-node data acquisition from industrial field sensor networks, historical-real-time data fusion of various process parameters in the production process, and statistical modeling and preliminary anomaly detection based on this data. With the development of big data and artificial intelligence technologies, causal analysis of multi-source process parameters, defect tracing modeling, and fault early warning algorithms for intelligent production lines have also been widely applied. The industry already has many circuit board quality monitoring technologies based on big data analysis and machine learning, such as statistical regression, shallow neural networks, and supervised classification, to achieve source identification and trend analysis of data anomalies in some production lines.

[0003] However, in practice, with the rapid expansion of the number of process nodes, the heterogeneity of sensor distribution, and the influence of objective factors such as network load, device local clock errors, and physical transmission delays, data acquisition at each process node often inevitably encounters problems such as acquisition delays, timestamp drift, local network congestion, and sudden synchronization failures. Because the local clocks of distributed acquisition devices in industrial settings are not perfectly synchronized, there are acquisition time deviations ranging from seconds to milliseconds between nodes, and this error exhibits spatial diffusion and dynamic variation. Furthermore, the different process labels, geographical layouts, and process stages of different acquisition nodes make time-series alignment of multi-source data extremely difficult.

[0004] The existing technologies generally have the following shortcomings: (1) There is a lack of dynamic compensation for the timing error of the process node acquisition. In a multi-node environment, the delay of data acquisition and the drift of the local clock of the equipment will produce the phenomenon of "timing error accumulation" in the causal chain analysis process, which will cause the time relationship between nodes to be distorted when the causal chain is reconstructed and the root cause positioning accuracy will decrease; (2) In the case of large-scale distributed acquisition, it is difficult to completely eliminate the timing misalignment of the process parameters of the whole process by relying on a single network synchronization protocol (such as independent NTP / PTP synchronization). In the case of abnormal system conditions, a large-scale time drift accumulation may also occur, which seriously affects the accuracy of causal reasoning and scheduling performance; (3) Most causal analysis models adopt post-event normalization or average callback for acquisition delay and timestamp offset, which can only partially correct the conventional deviation and is difficult to cope with the timing drift and comprehensive error coupling in large-scale, dynamic environments; (4) Some solutions lack flexible abnormal node identification, causal chain breakage and automatic timing consistency verification mechanisms, which result in insufficient traceability and model robustness in extreme process state switching and abnormal scenarios, and cannot quickly adapt to complex manufacturing processes with multiple processes, multiple batches and cross-equipment. Summary of the Invention

[0005] This application provides a circuit board quality traceability method based on big data, which aims to solve one of the problems or issues of the existing technology mentioned in the background.

[0006] This application provides a circuit board quality traceability method based on big data, which specifically includes: S1: Collect time-series data of multi-source process parameters during the circuit board manufacturing process, including real-time sensor output signals at each process node, and simultaneously record the geographical location information and process labels of the acquisition equipment to form a multi-dimensional time-series raw dataset.

[0007] S2: Normalize and denoise the timestamps of each process node in the multidimensional time series raw dataset, automatically correct the acquisition time series differences between devices based on the network time reference, and generate a standardized process parameter data stream.

[0008] S3: Based on standardized process parameter data streams, construct a multi-source error model to statistically evaluate timing errors caused by acquisition delays or timestamp drift, and generate adaptive timing error estimation vectors for each node.

[0009] S4: Input the timing error estimation vector and the standardized process parameter data stream into the adaptive timing error compensation algorithm to dynamically correct the timing arrangement order of the data at each process node and obtain a high-dimensional process parameter timing dataset after error compensation.

[0010] S5: For the high-dimensional process parameter time series dataset after error compensation, based on process labels and geographical location information, cluster adjacent abnormal parameters within a preset event-driven dynamic window, and automatically identify the temporal causal relationship characteristics between parameters.

[0011] S6: Input the causal relationship features of the parameters in the dynamic window into the causal reasoning modeling network, reconstruct the multi-level process causal links between the parameters through multi-dimensional causal chain reasoning, and mark the causal chain interruption points and time sequence abnormal nodes.

[0012] S7: Based on the reconstructed multi-level process causal link, automatically determine whether the current process anomaly point meets the preset defect tracing conditions. If it does, locate the defect source parameters and generate quality problem tracing decision factors.

[0013] S8: Input the quality problem tracing decision factors and causal link structure data into the visualization analysis system to generate a graphical defect tracing report in real time, and automatically output quality warning signals based on the consistency or anomaly of the causal chain time sequence.

[0014] S9: Monitors the dynamic changes of subsequent process parameters. If a process switch or change in environmental conditions is detected, it automatically adjusts the timing error estimation vector and causal chain inference model parameters to achieve adaptive traceability optimization for cross-process and cross-batch manufacturing scenarios.

[0015] This application provides a circuit board quality traceability method based on big data, which has the following beneficial effects: (1) This invention achieves time synchronization of multi-source sensor data at the second or even sub-second level through high-precision time synchronization mechanisms (such as PTP / NTP) and multi-node timestamp normalization, denoising, and calibration techniques, eliminating the data timing distortion problem caused by local clock drift and device asynchrony. Through the construction and dynamic correction of multi-source error models, the potential acquisition delay and timestamp drift between nodes can be suppressed within the error threshold. The maximum timing drift error is reduced from tens of milliseconds in traditional schemes to less than 1 millisecond, which greatly improves the timing consistency in the causal inference process and lays a high-resolution data foundation for the identification of process causal chains and the root cause of faults under complex processes.

[0016] (2) This invention introduces an event-driven dynamic window and a multi-dimensional feature adaptive clustering algorithm, which can adjust the data analysis window in real time for parameter mutations or abnormal fluctuations in the manufacturing process, and quickly locate and aggregate relevant abnormal parameter segments. Compared with fixed window or static acquisition methods, the dynamic window mechanism reduces the average abnormal response delay of the system by more than 40%, and can achieve second-level abnormal capture and dynamic tracing under variable process flow, batch switching or environmental disturbances, which greatly improves the online judgment and early warning capabilities of manufacturing quality.

[0017] (3) By adopting an adaptive timing error compensation and recursive optimization algorithm, the system can not only dynamically correct residual errors introduced by batch fluctuations, node drift, or environmental changes on the basis of the initial correction, but also continuously suppress instantaneous errors through a real-time feedback mechanism to ensure the high accuracy of the final data timing. Experimental results show that the robustness of this method to complex timing errors in dynamic process environments is improved by more than 15%, and the absolute error of node-level time drift after multiple rounds of correction is generally less than 0.1ms, which is far superior to existing static modeling and manual correction methods, effectively eliminating the interference of error accumulation on subsequent causal inference judgment.

[0018] (4) This invention integrates new causal modeling technologies such as high-dimensional parameter spatiotemporal labels and graph neural networks, realizing the automatic construction and optimization of multi-level and multi-parameter process causal chains throughout the entire manufacturing process. Compared with single-parameter or manual experience-based tracing, this solution supports full-link causal reasoning at the equipment level, process level, and space level, improving the accuracy of causal chain reconstruction by 10% to 20%. It can significantly shorten the defect tracing and parameter root cause localization cycle, and can accurately indicate abnormal nodes in the link anomaly and interruption links, avoiding misjudgment of tracing caused by time sequence distortion, and greatly improving the efficiency of manufacturing defect diagnosis and quality optimization. Attached Figure Description

[0019] Appendix Figure 1 This is the main flowchart of a circuit board quality traceability method based on big data.

[0020] Appendix Figure 2 This is a sub-flowchart of a circuit board quality traceability method based on big data.

[0021] Appendix Figure 3 This is another sub-flowchart of a big data-based circuit board quality traceability method. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following disclosure provides many different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. In addition, examples of various specific processes and materials are provided, but those skilled in the art will recognize the application of other processes and the use of other materials.

[0024] As attached Figure 1 As shown, this application provides a circuit board quality traceability method based on big data, specifically including: S1: Collect time-series data of multi-source process parameters during the circuit board manufacturing process, including real-time sensor output signals at each process node, and simultaneously record the geographical location information and process labels of the acquisition equipment to form a multi-dimensional time-series raw dataset.

[0025] S2: Normalize and denoise the timestamps of each process node in the multidimensional time series raw dataset, automatically correct the acquisition time series differences between devices based on the network time reference, and generate a standardized process parameter data stream.

[0026] S3: Based on standardized process parameter data streams, construct a multi-source error model to statistically evaluate timing errors caused by acquisition delays or timestamp drift, and generate adaptive timing error estimation vectors for each node.

[0027] S4: Input the timing error estimation vector and the standardized process parameter data stream into the adaptive timing error compensation algorithm to dynamically correct the timing arrangement order of the data at each process node and obtain a high-dimensional process parameter timing dataset after error compensation.

[0028] S5: For the high-dimensional process parameter time series dataset after error compensation, based on process labels and geographical location information, cluster adjacent abnormal parameters within a preset event-driven dynamic window, and automatically identify the temporal causal relationship characteristics between parameters.

[0029] S6: Input the causal relationship features of the parameters in the dynamic window into the causal reasoning modeling network, reconstruct the multi-level process causal links between the parameters through multi-dimensional causal chain reasoning, and mark the causal chain interruption points and time sequence abnormal nodes.

[0030] S7: Based on the reconstructed multi-level process causal link, automatically determine whether the current process anomaly point meets the preset defect tracing conditions. If it does, locate the defect source parameters and generate quality problem tracing decision factors.

[0031] S8: Input the quality problem tracing decision factors and causal link structure data into the visualization analysis system to generate a graphical defect tracing report in real time, and automatically output quality warning signals based on the consistency or anomaly of the causal chain time sequence.

[0032] S9: Monitors the dynamic changes of subsequent process parameters. If a process switch or change in environmental conditions is detected, it automatically adjusts the timing error estimation vector and causal chain inference model parameters to achieve adaptive traceability optimization for cross-process and cross-batch manufacturing scenarios.

[0033] Step S1: Collect time-series data of multi-source process parameters during the circuit board manufacturing process, including real-time sensor output signals at each process node, and simultaneously record the geographical location information and process labels of the acquisition equipment to form a multi-dimensional time-series raw dataset. Specifically, this includes: S1.1: Deploy multiple types of sensor acquisition devices at all process nodes in the manufacturing process, and obtain raw signal data of process parameters through industrial fieldbus protocol to obtain raw signals of multi-source process parameters such as temperature, humidity, current, and voltage.

[0034] The process nodes, as the objects of this step, cover the key nodes in the circuit board manufacturing process involving multiple types of process parameters such as temperature, humidity, current and voltage. Each node is equipped with a matching type of high-precision sensor acquisition device.

[0035] A multi-point parallel deployment strategy for the circuit board manufacturing environment is adopted, embedding temperature sensors (accuracy ±0.1℃), humidity sensors (accuracy ±2%RH), current sensors (range 0~10A, accuracy ±0.2%FS), voltage sensors (range 0~100V, accuracy ±0.5%FS) and other customized sensing units at each process node to achieve synchronous acquisition of high-frequency sampled raw signals at different physical locations.

[0036] To address the need for real-time transmission of sensor signal data, the system interfaces with the main control chip of the acquisition device via industrial fieldbus protocols (such as Modbus-RTU, CAN bus, EtherCAT protocol, etc.) to achieve unified reporting of raw process parameters and preliminary anomaly detection for each node.

[0037] The preprocessing function built into the signal acquisition module is used to perform amplitude filtering and quantization standardization on multiple types of raw sensor signals in one go, and output process parameters of different types and amplitudes into a standard and unified data format (such as 16-bit fixed-point representation) to ensure the compatibility of subsequent multi-node data fusion.

[0038] Furthermore, by assigning a data packet identifier to the original signal at each moment through the node controller, distributed data traceability and data source tracking are achieved, thereby improving the efficiency of data integrity verification in subsequent data governance processes.

[0039] Through the collaborative acquisition of multiple types of sensors and efficient transmission via fieldbus protocol, the high-frequency raw signals of multi-source process parameters distributed across various nodes are output in the form of structured data streams, enabling real-time parameter acquisition across multiple dimensions and nodes. This establishes a high-quality raw data source for full-stack process timing traceability and anomaly monitoring in the manufacturing process.

[0040] For example, on an actual automated circuit board assembly line, a PT100 high-precision temperature sensor (accuracy ±0.1℃), an SHT35 digital humidity sensor (accuracy ±1.5%RH), a LEM HO8-P Hall current sensor (range 0~8A), and a locally powered 24V voltage acquisition module are deployed in the reflow soldering process area. All acquisition devices communicate point-to-point with the central acquisition industrial control computer using the Modbus-RTU protocol at a baud rate of 9600bps. The sampling frequency is set to 50Hz per second, and each data packet contains a timestamp, node number, sensor type, and quantized data. After the sensor signals are processed by the acquisition module through second-order low-pass filtering and standardized quantization, they are output in a unified 16-bit signed integer format. During one hour of production line operation, all nodes output a total of 180,000 sets of raw data packets, with a data loss rate of less than 0.01%. Through verification and comparison, all abnormal nodes can be traced in real time, effectively supporting complete multi-node fine-grained traceability and subsequent clock synchronization, causal chain construction, and other full-process big data analysis.

[0041] S1.2: Based on the original process parameter signals, a high-precision clock synchronization mechanism is adopted to synchronize the local time of each sensor acquisition device with the network time reference in real time, so as to unify the acquisition timing reference of the original process parameter signals.

[0042] The input consists of raw signal data of process parameters collected from various types of high-precision sensor devices, involving high-frequency raw data streams including local device clock records.

[0043] A high-precision clock synchronization mechanism (based on Network Time Protocol (NTP) or Precision Time Protocol (PTP)) is employed to synchronize the local clocks of the sensor acquisition devices at each process node. Furthermore, a real-time calibration communication link with a unified network clock source is established in the embedded control system of each acquisition module through a network time synchronization manager, enabling periodic correction and updates of the device's local time and parameterizing the node's local clock deviation. Further, a system-level time synchronization deviation analysis algorithm is applied, dynamically sampling the clock drift curves of all nodes based on a preset acquisition period (e.g., 1 second) and a maximum allowable time drift threshold (e.g., 1 ms), to calculate the clock offset Δt for each node i. i The specific calculation formula is as follows:

[0044] in, Let be the local time of the i-th node. This is the reference global time under the Network Time Protocol.

[0045] The difference compensation module updates the timestamps of node-collected events in real time, uniformly mapping the collection time to a globally unique reference time sequence. Furthermore, a clock synchronization residual monitoring mechanism is used to perform adaptive error trigger verification of the timestamp synchronization effect within the collection period. If a synchronization error exceeding the threshold is detected, a secondary refinement calibration process of the local clock is automatically triggered.

[0046] Finally, the original process parameter signals, corrected to a unified time base, are output, providing a unified and highly consistent acquisition timing reference for subsequent addition of geographic information labels and integration of multi-node data.

[0047] Through a high-precision clock synchronization mechanism, the original signal times acquired by multiple nodes and asynchronously are transformed into a globally consistent reference time, enabling the collaborative fusion and high-resolution traceability of multi-source process parameter data on the time axis.

[0048] For example, all sensor acquisition devices deployed at each process node of the circuit board reflow soldering production line have built-in PTP (Precision Time Protocol) synchronization chips. The system uses a central time server as the main clock source and sends out a network-wide synchronization pulse signal every 10 seconds. The clock deviation measurement accuracy within each node is ±0.5ms, and the time synchronization tolerance upper limit is set to 1ms. When a node detects in real time... If the deviation is within 1.2ms, the local clock is automatically fine-tuned, and the deviation is converged to within ±0.5ms in the next sampling period through the PTP hardware compensation interface. After one hour of continuous acquisition, the timing error of all 180,000 synchronized data packets is less than 1ms, and the timing references of the original data of each node are completely aligned, achieving high-resolution data stream output with a unified timing reference. Subsequent verification by the spatial labeling and process integration modules confirmed that the consistency of the acquisition timing reached 100%. This method effectively solves the timing drift problem caused by local clock errors in multi-node process parameter acquisition, providing a highly consistent data foundation for critical causal chain analysis.

[0049] S1.3: Acquire the original process parameter signals with unified timing reference, and attach the geographical location information of the acquisition device to each data packet through the embedded acquisition control system to realize the digital identification of the spatial distribution of process nodes.

[0050] The original process parameter signal data, after high-precision clock synchronization correction, is used as input, which includes a multi-node high-frequency acquisition signal stream under a unified time reference.

[0051] An embedded acquisition and control system (parameters: main control chip model, data interface standard) is adopted. For each raw signal data packet of process parameters, the integrated geolocation sensing module (such as GNSS / GPS / BeiDou positioning module) is called to obtain the spatial coordinate information of the physical deployment location of the acquisition equipment in real time.

[0052] Furthermore, by using a spatial data binding algorithm (parameters: unique node hardware identifier, location data time window), a corresponding geographic location information tuple is attached to each data packet to realize a three-element digital association structure of data-device-space, and generate a multi-element data packet structure containing timestamp, node number, acquisition device ID, and spatial coordinates (such as longitude, latitude, and can be extended to three-dimensional elevation if necessary).

[0053] Furthermore, through a spatial distribution verification mechanism, the data packets of integrated spatial coordinates are checked for node location consistency. A spatial consistency threshold comparison algorithm is used to eliminate erroneous or drifting positioning values ​​caused by spatial hardware anomalies, ensuring the accuracy and traceability of geographic information.

[0054] Furthermore, a digital spatial distribution identifier generation strategy is adopted to assign a globally unique spatial distribution code to each process node. Based on spatial topology, each node is summarized into a data set with a clear spatial distribution hierarchy and topological relationship on the circuit board manufacturing process line, laying the foundation for subsequent spatial correlation analysis and process tracking.

[0055] Through the above series of digital spatial labels and integrated verification processes, the original signal data of process parameters are deeply bound with the geographical location information of the acquisition equipment, and the original data packets of process nodes with complete spatial distribution attribute recognition capabilities are output, realizing the digital and structured identification of the spatial distribution of each node.

[0056] For example, in the PT100 temperature acquisition nodes deployed on an automated circuit board manufacturing line, each acquisition device integrates a high-precision BeiDou positioning module (positioning accuracy better than 2 meters). The system synchronously acquires temperature data and device location information at a frequency of 10Hz. The main control chip STM32F407 receives NMEA-0183 format data output from BeiDou via a serial interface, parses longitude and latitude information in real time, and appends it to each temperature acquisition data packet. A single data packet contains a synchronization timestamp (e.g., 2024-06-12 10:15:23.640), node number (e.g., NODE-01), device ID (e.g., DEV-10001), raw temperature data (e.g., 23.7℃), and physical location (115.850123, 28.682543). By setting a spatial consistency threshold of ±5m, three sets of abnormal positioning data caused by production transfer and relocation are effectively eliminated during a day of production, and the spatial tag verification pass rate reaches 99.98%. This indicates that the above technology can efficiently and automatically integrate equipment geographic information into the process parameter acquisition stream, realize workshop-level spatial traceability and digital management of node distribution, and effectively support subsequent anomaly clustering and causal chain accurate tracing based on spatial attributes.

[0057] S1.4: For data streams with geographic location, automatically extract and record process tag information according to the process rules of different manufacturing stages to form process parameter tagged data with node identity and process stage.

[0058] S1.5: Based on the labeled data of process parameters, the data stream is structured and integrated according to the process node sequence and acquisition time sequence to construct a multi-dimensional time series raw dataset, providing high-dimensional and high-resolution basic data for subsequent time series normalization and causal inference.

[0059] Step S2: Normalize and denoise the timestamps of each process node in the multidimensional time-series raw dataset, automatically correct the acquisition timing differences between devices based on the network time reference, and generate a standardized process parameter data stream. Specifically, this includes: S2.1: Based on multi-source process parameter time series data, extract the original timestamp information of each process node and perform batch-level timestamp aggregation processing to achieve subsequent unified timeline management.

[0060] Based on the original dataset of multi-source process parameters integrated with high-precision clock synchronization and spatial identification, the real-time acquisition signals of each process node are used as input, including multi-node data streams under a unified time reference and complete process tag information.

[0061] A timestamp extraction algorithm (parameter settings: 16-bit or 32-bit clock counter bit width, sampling period matching standard) is used to parse the data stream of each process node unit and output the corresponding original timestamp set to form a multi-node original timestamp vector.

[0062] Furthermore, by using a batch aggregation processing method (parameter settings: batch identification rules, process stage window, maximum allowed time interval ΔT within each batch), the original timestamps of each node collected under the same production batch and the same process stage are summarized and grouped to generate a batch-level timestamp set, providing a data foundation for subsequent unified timeline management.

[0063] Furthermore, a time correlation verification algorithm (parameters: time continuity criterion, maximum allowable loss rate of 1%, preset abnormal jump threshold of 3 times the sampling period) is applied to perform integrity and validity checks on the timestamp set within the batch, filtering out duplicate, broken or non-compliant timestamp fragments, and outputting highly complete and compliant batch-level timestamp data.

[0064] Furthermore, by using a timestamp sequence alignment method between nodes (parameters: node number, process label sorting, and spatial coordinate prior), the aggregated timestamp data of each process node is sorted twice according to the process flow order and time progression relationship, generating a process node-timestamp list that supports unified time axis mapping, thus realizing the structured output of multi-node and multi-batch timestamp data.

[0065] Through the aforementioned algorithm chain, the batch extraction and aggregation of multi-node raw timestamp data can lay a high-resolution, centralized data foundation for subsequent time-series denoising, synchronization calibration, and unified sampling cycle management, thereby realizing batch-level structured management of manufacturing process time-series information.

[0066] For example, in an automated circuit board manufacturing line, temperature, humidity, current, and voltage signals at each process node are integrated with 32-bit local timers, outputting data packets with 16-bit high-precision timestamps, with a fixed sampling period of 20ms. The system aggregates all raw timestamps collected from the temperature process node, grouping them into batches of 100 circuit boards per production batch, based on the physical production line segment and process flow sequence, using the production batch number as the aggregation key. The maximum time interval ΔT for each batch is set to 2s, and a sliding window consistency check is applied to all timestamps within a group, detecting a single batch timestamp loss rate of no more than 0.3% and no cross-batch duplicate segments. After sorting by node number and process stage label, a batch-level process node-timestamp structured list is output. Before the introduction of denoising algorithms in subsequent steps, the batch-level timestamp data processed by this sub-step meets the unified axis management requirements, adapting to downstream time denoising and synchronization correction modules. Compared to the unaggregated, scattered timestamp stream, its structured rate increases from 72% to 99.6%.

[0067] S2.2: Apply time-series denoising algorithms (such as interference point filtering and moving mean smoothing) to the extracted original timestamp sequence to eliminate outliers and acquisition jitter noise, and obtain clean timestamp data.

[0068] S2.3: Compare the purification timestamp data with the network time base, and use network time calibration protocols (NTP / PTP, etc.) to synchronize and adjust the clock offset between devices to generate a unified standard timestamp sequence after alignment.

[0069] S2.4: Using a unified standard timestamp sequence, normalize the time matching relationship of the original process parameter data stream so that all process parameters are referenced by the same acquisition time, forming a normalized process parameter matrix.

[0070] S2.5: Based on the normalized process parameter matrix, a multi-node time consistency detection algorithm is used to correct and eliminate abnormal time records caused by acquisition synchronization delay, ensuring the time sequence consistency of the data and obtaining a preliminary standardized process parameter data stream.

[0071] S2.6: Input the preliminary standardized process parameter data stream into the end-to-end consistency verification module to perform multi-dimensional verification of the timing logic of each process node, and output the final standardized process parameter data stream to provide highly consistent input for the construction of multi-source error models and subsequent causal chain reconstruction.

[0072] Step S3: Based on the standardized process parameter data stream, a multi-source error model is constructed to statistically evaluate the timing errors caused by acquisition delays or timestamp drift, and to generate adaptive timing error estimation vectors for each node. For example... Figure 2 As shown, it specifically includes: S3.1: Perform preliminary statistical analysis on the timestamp data of all process nodes in the standardized process parameter data stream, and use a time-series distribution characteristic extraction algorithm to obtain the actual acquisition time distribution parameters of each process node to form a timestamp distribution characteristic parameter set.

[0073] S3.2: Based on the timestamp distribution feature parameter set, a data fusion modeling method is used to quantify the time reference offset between detection devices corresponding to each process node, construct the synchronization error feature matrix between devices, and provide synchronization error prior for the accuracy matching of subsequent error models.

[0074] Based on the set of timestamp distribution characteristic parameters obtained through preliminary statistical analysis, the actual acquisition time distribution parameters of each process node in the standardized process parameter data stream are used as input to cover all key process nodes and their corresponding acquisition devices.

[0075] The device time base difference mapping method (parameter settings: node number lookup table, distribution mean μ, variance σ, sampling interval Δt) is used to achieve preliminary quantification of the relative offset of time base between acquisition devices at different process nodes.

[0076] Furthermore, a multi-source data fusion modeling method is used (parameter settings: Kalman filter weights Wk, node spatial distance L). ij (Signal covariance threshold ε), integrate the distributed parameters of each node, and calculate the clock offset δ between devices. ij .

[0077] The synchronization error between devices is quantified using the following formula:

[0078] in, , Let be the mean timestamp distributions of node i and node j, respectively. This represents the synchronization offset between the two devices.

[0079] Furthermore, the synchronization offsets between all node pairs are integrated, and a matrix construction algorithm is used to generate a synchronization error feature matrix between devices. Specifically:

[0080] in, This represents the total number of process nodes. It is a symmetric synchronization error matrix.

[0081] Principal component analysis (PCA) was used to further reduce the dimensionality of the synchronization error feature matrix and extract the main synchronization error patterns, forming a prior feature subset for improving the accuracy of the subsequent error model.

[0082] A consistency confidence assessment algorithm (parameter: maximum synchronization error tolerance θ) is used to calibrate abnormal offset nodes and output the error feature confidence score V as a prior constraint index for the error model.

[0083] By using the above data fusion modeling and matrix feature quantization, the time base offset results between devices are output under a unified structure, providing a highly reliable a priori basis for subsequent modeling, parameter matching and compensation strategies of the multi-source error model, and realizing high-precision analysis of the synchronization relationship of all nodes in the process flow.

[0084] For example, in a circuit board mass production line, high-precision acquisition devices are configured for each process node, such as temperature, humidity, and current. Data fusion modeling is used to quantify the time base offset of five main process nodes. The average timestamp distribution of each node pair is as follows: Node 1 (245602.0ms), Node 2 (245604.5ms), Node 3 (245612.3ms), Node 4 (245600.2ms), and Node 5 (245611.9ms). Based on the above formula, the synchronization offset between each node is calculated. For example, node 1 to node 2 Nodes 3-4 Etc. Generate a 5×5 symmetric matrix. The diagonals are all 0. After PCA dimensionality reduction, the first three principal components have contributed 97.8% of the variance, indicating that the synchronization error is mainly concentrated in a few node pairs. The confidence rating algorithm calibrates that the offset between node 3 and node 4 is higher than the maximum tolerance θ=10ms, outputting a confidence score of V=0.99, suggesting that it needs to be included in subsequent model compensation. Finally, this step outputs the device synchronization error feature matrix after fusion modeling and matrix transformation, which can be used to guide the dynamic parameter setting of batch-level synchronization compensation strategy. In system verification, the maximum absolute error of synchronization offset decreased from the initial value of 15ms to 2.1ms, effectively improving the time axis accuracy and robustness of the causal chain inference model.

[0085] S3.3: Apply a dynamic detection algorithm for acquisition delay to the standardized process parameter data stream. By comparing the event trigger time of each node under the same process flow with the network reference clock, the acquisition delay feature quantity is calculated in real time to form a node-level acquisition delay statistical feature vector.

[0086] Dynamic detection of acquisition delay is performed on the standardized process parameter data stream to achieve real-time calculation of node-level acquisition delay characteristics.

[0087] An event trigger time comparison method (parameters: standardized acquisition timestamp, network reference clock, process flow node identifier) ​​is used to obtain the actual event trigger time of each node under the same process flow.

[0088] Furthermore, the trigger time of the distributed acquisition nodes is globally mapped by the network reference clock. The standard time axis generated after clock synchronization calibration is used to achieve precise alignment between the node trigger time and the unified time reference, and to eliminate pseudo-delay data caused by system jitter or local anomalies.

[0089] Furthermore, by applying the acquisition delay calculation formula, for each process node... Acquisition delay feature Quantification:

[0090] in, For nodes The actual data collection time, For nodes Theoretical or network benchmark time.

[0091] Furthermore, to address latency fluctuations in dynamic acquisition scenarios, a window sliding statistical algorithm (window width) is used. It is configurable (can be set according to the actual production line cycle time), performs real-time sampling of node-level acquisition delay features, calculates their mean, variance, extreme values ​​and other statistical characteristics, and forms an acquisition delay distribution descriptor.

[0092] Furthermore, a delay anomaly identification algorithm (parameters: delay threshold, maximum jitter amplitude, normalized scoring factor) is adopted to identify the delay statistical characteristics within the window, eliminate sudden abnormal data, and mark the delay interval that continuously deviates from the benchmark to ensure the representativeness and robustness of the statistical results.

[0093] Finally, the acquisition delay distribution descriptors of each node are summarized to form a node-level acquisition delay statistical feature vector, and the output is standardized to support downstream multi-source error modeling and adaptive timing error estimation.

[0094] The above-mentioned dynamic acquisition delay detection algorithm compares the actual acquisition time of each process node in the standardized process parameter data stream with the network reference clock in real time, and transforms it into a node-level, time-specific acquisition delay statistical feature vector, thereby achieving high-precision extraction of the timing delay features of multiple nodes in the process flow.

[0095] For example, in an automated circuit board manufacturing line, the network reference clock signal is set as the GPS time reference, and the data acquisition nodes include temperature, humidity, current, and voltage parameter units. The standard timestamp of the data packet for each process node is compared with the actual acquisition trigger time, using a formula... Data was collected from 1000 manufacturing cycles. Window sliding width. The sampling interval was set to 1 minute, the latency threshold to 2ms, and the maximum jitter amplitude to 10% of the sampling period. The mean latency, standard deviation, and 95% confidence extreme value intervals for each process node within the batch were calculated. Single-point sudden threshold exceedances were removed, and labels were added to the abnormal duration intervals to improve statistical robustness. Experimental results show that the mean acquisition latency of most nodes was controlled within 1.5ms, and the standard deviation was less than 0.2ms, with only brief deviations occurring during process transitions and network load peaks. The final output statistical feature vectors of acquisition latency for each node provide high-timeliness and high-resolution input for subsequent error model construction, improving the temporal accuracy and anomaly response capability of causal chain inference in the process flow.

[0096] S3.4: By integrating the synchronization error feature matrix between devices with the statistical feature vector of node-level acquisition delay, a multi-source error coupling modeling strategy is adopted to automatically construct a multi-source error model of the process flow and output the statistical probability distribution model of timestamp drift and delay components of each node.

[0097] The input consists of a standardized and multidimensional consistency-verified process parameter data stream, an inter-equipment synchronization error feature matrix, and a node-level acquisition delay statistical feature vector.

[0098] A multi-source coupling error statistical modeling method is adopted (parameters: total number of process nodes N, synchronization error characteristic matrix). Acquisition Delay Statistical Vector Fusion weight matrix This enables joint dynamic modeling of timing errors at multi-dimensional process nodes, and achieves the fusion and normalization of heterogeneous error data in a unified space.

[0099] Through the synchronization error feature matrix The column normalization process maps the synchronization error components between each process node pair to the [0,1] interval. The specific normalization calculation is as follows:

[0100] in, Let be the synchronization error components between node i and node j. , These are the minimum and maximum elements of the entire matrix, respectively.

[0101] Acquisition delay statistics vector Application data range normalization:

[0102] in, The acquisition delay statistics for node i are as follows: , Set global upper and lower bounds for the latency of all nodes.

[0103] Furthermore, an adaptive weighted feature fusion algorithm is adopted to set weights based on node importance. For the i-th node, the normalized synchronization error and the acquisition delay are fused into a coupled feature component. :

[0104] in, , Let i be the normalized weighting factor for the synchronization error and acquisition delay characteristics, satisfying .

[0105] Furthermore, all nodes Construct a multi-source coupling error feature matrix with a unified structure based on node linked lists and sequence grouping. , where n is the node number and t is the error feature sampling sequence number.

[0106] Using Gaussian mixture probability modeling, for The sequence is fitted with parameters. For the coupled feature samples at node n and time t, the mixture distribution number K is used to divide the samples, and the maximum likelihood estimation algorithm is used to optimize the model parameters, resulting in the following probability density function:

[0107] in, The mixed weights of the k-th distribution at the n-th node are... , These are the mean and variance, respectively.

[0108] Furthermore, using the Bayesian confidence interval criterion, the upper and lower bounds of the 99% confidence line are assessed for the statistical distribution of coupling errors at each process node. , Specifically:

[0109]

[0110] in The quantile factor is the quantile factor at the 99% confidence level under the standard normal distribution.

[0111] The probability density functions and confidence intervals of all nodes together form a multi-node time series error probability map. Out-of-bounds nodes are screened out using anomaly probability thresholds, and timestamp drift is output. and acquisition delay The statistical probability distribution model of the two major components ultimately forms a complete set of statistical parameters for multi-node time-series errors, providing a statistical basis for subsequent adaptive estimation vector generation and dynamic compensation strategies.

[0112] Through the aforementioned high-order data fusion, multi-level probability modeling, and anomaly threshold analysis, the original node pair synchronization and node-level delay characteristics are transformed into a statistical probability distribution model with engineering-determinable attributes. This enables comprehensive and quantitative modeling of cross-source and multi-node timing errors, effectively improving the accuracy of dynamic early warning of errors across all nodes in the process flow and the response speed of node anomaly identification.

[0113] For example, in an automated circuit board manufacturing line, standardized process parameter data streams cover 64 process nodes. After matrix normalization, the maximum absolute value of the synchronization error characteristic is 18ms, and the global mean of the acquisition delay statistical characteristic is 12.3ms, with a distribution range of [7.2ms, 16.1ms]. In terms of weight allocation, major nodes are assigned a weight of 0.85, and general nodes are assigned a weight of 0.45. Using weighted fusion, the coupled feature components of node 7 are calculated. .right The entire sequence is fitted with a mixture normal distribution. Let K=3, and the maximum likelihood method is used to iteratively converge to... , etc. With The upper and lower bounds of the confidence interval for node 7 are respectively The detection rate of abnormal nodes was 1.1%, and the output of the timestamp drift and delay probability distribution model of all nodes was improved by 9.2% after subsequent compensation. The average time for automatic identification of process abnormal points was less than 18 seconds.

[0114] S3.5: Based on the multi-source error model, an adaptive estimation algorithm is used to dynamically generate adaptive timing error estimation vectors for each process node, and to associate the numerical range of the standardized process parameter data stream in real time, so as to realize the characterization and output of timing error characteristics for different time periods and different process states.

[0115] Step S4: Input the timing error estimation vector and the standardized process parameter data stream into the adaptive timing error compensation algorithm to dynamically correct the timing arrangement order of the data at each process node, thereby obtaining a high-dimensional process parameter timing dataset after error compensation. Figure 3 As shown, it specifically includes: S4.1: Obtain the standardized process parameter data stream and the adaptive timing error estimation vector as the joint input of the adaptive timing error compensation algorithm, ensuring that the data stream and error estimation have a unified time base and parameter labels.

[0116] The input consists of a standardized high-dimensional process parameter data stream and adaptive timing error estimation vectors generated for each process node and process period. All data is bound to a unified time base and process parameter labels.

[0117] A precise data synchronization mechanism (parameters: network time base, process tag index) is adopted to achieve one-to-one correspondence calibration between the standardized process parameter data stream and the adaptive timing error estimation vector in terms of node number, time axis and tag dimension.

[0118] Furthermore, by using a joint data cache matching algorithm (parameters: cache depth M, node pool N, matching fault tolerance ε), the synchronous matching of standardized process parameter data streams from different sources with error estimation vector samples is achieved, and hash mapping is performed on all data indexes to form an efficient data retrieval structure.

[0119] Furthermore, a time-series index recalibration method (parameters: time-series consistency threshold δ, index correction factor γ) is adopted to jointly recalibrate the time-series labels of node-level data packets before the data stream is jointly input. This eliminates redundant or abnormal indexes induced by factors such as background noise and acquisition errors, thereby achieving complete alignment between the data stream and the error estimation vector in terms of process nodes and timing.

[0120] Furthermore, by utilizing input validation and consistency verification mechanisms (parameters: full node coverage condition, label integrity score λ), consistency is determined for the number of nodes, label sequence, and time-series index of the joint input data, screening for issues such as missed reports, duplication, and disordered order, and generating data and vector pairs that meet high consistency requirements.

[0121] Through the above steps, the standardized process parameter data stream and the adaptive timing error estimation vector are combined as highly consistent inputs to the adaptive timing error compensation algorithm, thus ensuring the basic parameters for subsequent global timing correction.

[0122] For example, in a high-speed circuit board automated testing production line, the input standardized process parameter data stream includes five key process nodes (temperature, humidity, soldering current, chip mounting suction, and appearance defects), with 50,000 data entries per node per batch, and a timestamp resolution of 0.1ms. An adaptive timing error estimation vector covers all parameters across "five nodes × thousands of batches × multiple process states," and after vector splitting, it is grouped by process tags (e.g., SMT, AOI, ICT, programming, packaging). A node index hash mapping algorithm (cache depth M=100, matching fault tolerance ε=5%) is used to fully retrieve and match data for each node, with an index recalibration threshold δ=0.5ms, eliminating mismatched data less than 0.1%. After verification, it is ensured that the parameters and error estimates for each node in each batch are a one-to-one correspondence, with no omissions or order discrepancies. Ultimately, the output can accurately align the data stream and error vector with the timing and labels, providing a high-quality and unambiguous input foundation for the adaptive timing error compensation algorithm. This supports core error correction capabilities such as global timing correction and precise compensation for instantaneous drift at subsequent nodes. System simulation shows that the mismatch rate is less than 0.01%, the false negative rate is zero, and the process node consistency score λ=1.00.

[0123] S4.2: Using an adaptive timing error compensation algorithm, global timing correction is performed on the timestamp information of each process node in the standardized process parameter data stream to correct timestamp drift caused by equipment distribution, acquisition delay, etc., and obtain a high-dimensional process parameter timing stream after preliminary calibration.

[0124] Based on the input conditions, the standardized process parameter data stream and the adaptive timing error estimation vector are jointly input into the adaptive timing error compensation algorithm module to ensure that the input data stream and error estimation parameters are aligned to a unified time reference and have complete node parameter labels.

[0125] A global timing correction method is adopted (parameters: process node number set N, standardized timestamp sequence T, node adaptive timing error estimation vector Δt). est This enables the synchronous adjustment and correction of the current recorded timestamps of each process node in the standardized process parameter data stream.

[0126] Furthermore, for each process node i in the standardized process parameter data stream, the corresponding adaptive timing error estimate Δt is retrieved. est,i The following timing correction formula is applied:

[0127] in, Let be the corrected timestamp of node i. Let i be the original normalized timestamp of node i. The timing error is adaptively estimated.

[0128] Furthermore, using the above formula, the timestamp sequences of all process nodes N are batch-operated to generate a high-dimensional process parameter time stream after preliminary calibration, thereby achieving unified correction of timing misalignment caused by heterogeneous errors such as equipment spatial distribution, acquisition delay, and timestamp drift.

[0129] Furthermore, a global sorting algorithm between nodes is adopted (parameters: node number sequence, t). corr The corrected time-series data of each process node are sorted globally in a unique master sequence to ensure that the high-dimensional time-series flow is strictly arranged according to the actual occurrence order of the process flow.

[0130] The multi-dimensional consistency verification mechanism (parameters: node process label, geographical location information, and correction timestamp sequence) helps to determine whether there are potential anomalies such as residual time sequence overlap or order reversal in the corrected data stream, and marks the detected violation records to facilitate downstream recursive optimization.

[0131] By using an adaptive timing error compensation algorithm, the standardized process parameter data stream is globally time-corrected according to the adaptive timing error of each node, and transformed into a high-dimensional process parameter timing stream after preliminary calibration, thereby improving the timing consistency and causal chain drive of the process flow.

[0132] For example, in a real-world circuit board mass production line, assume the input standardized process parameter data stream contains five process nodes, numbered 1-5, corresponding to original standardized timestamps of [245602.3, 245604.5, 245612.1, 245600.2, 245611.8] ms. The node-level adaptive timing error estimation vector Δt... est The values ​​are [1.9, 2.4, 12.0, 0.1, 2.2] ms respectively. Apply the timing correction formula: For node 1, the corrected timestamp is 245602.3-1.9=245600.4ms.

[0133] For node 2, 245604.5 - 2.4 = 245602.1 ms.

[0134] For node 3, 245612.1 - 12.0 = 245600.1 ms.

[0135] For node 4, 245600.2 - 0.1 = 245600.1 ms.

[0136] For node 5, 245611.8 - 2.2 = 245609.6 ms.

[0137] The above formula is executed in batches to obtain a five-node corrected timestamp sequence [245600.4, 245602.1, 245600.1, 245600.1, 245609.6] ms. The corrected timestamps are then globally sorted. Nodes 3 and 4 both have a corrected timestamp of 245600.1 ms. Node 3 is sorted before Node 4 based on node number and process tag priority. A multi-dimensional consistency verification mechanism ensures that the corrected timing stream does not exhibit duplicate nodes or reversed order across process stages, achieving timing consistency. This correction process reduces the timing drift error between nodes from a maximum of 12.0 ms to less than 1 ms, effectively supporting process causal chain reasoning and defect localization at high timing resolution, and providing a precise calibration basis for subsequent recursive optimization and robustness verification.

[0138] S4.3: Based on the high-dimensional process parameter time flow after preliminary calibration, the timing deviation residuals between each process node are dynamically calculated. The adaptive timing error estimation vector is recursively optimized through a real-time error feedback mechanism to achieve secondary correction of transient errors.

[0139] Obtain the high-dimensional process parameter time stream after preliminary calibration, which includes the corrected timestamp sequence of each process node and its corresponding data vector after standardization and adaptive initial timing error compensation.

[0140] A node-level time series deviation residual dynamic calculation method is adopted (parameters: full set of nodes N, standardized timestamp sequence T, calibrated timestamp sequence T). corr This allows for one-to-one calculation of the timing residuals between each process node after preliminary calibration, and extraction of the actual remaining synchronization offset between nodes.

[0141] Furthermore, the mean, variance, and extreme values ​​of the time-series deviation residuals at each node level within the window are calculated in real time using a residual dynamic recursion algorithm (parameters: recursion window width w, historical residual storage H, recursion learning rate α), forming the transient response characteristics of time-series errors between nodes.

[0142] Furthermore, a real-time error information feedback mechanism (parameters: recursive minimum residual threshold ε, residual convergence criterion γ) is adopted. The dynamic characteristics of the aforementioned time-series deviation residuals are jointly input with the original adaptive time-series error estimation vector. The vector weights are dynamically adjusted through a recursive feedback correction formula to achieve secondary optimization of the node-level error estimation. The correction formula is as follows:

[0143] in, For nodes A new adaptive timing error estimate. Based on the previous round of estimation results, The current window node The timing residuals, These are the parameters for recursive learning.

[0144] Furthermore, through multi-round recursive correction and residual convergence criterion determination (parameters: maximum number of recursive rounds R, convergence tolerance η), it automatically determines whether the time series residual sequence has entered the steady state region, applies an adaptive fine-tuning mechanism (parameter: node-level buffer factor β) to non-converged nodes, and continuously adjusts until the residual converges or reaches the maximum number of rounds.

[0145] Through the above adaptive recursive feedback optimization method, the transient error of the high-dimensional process parameter timing flow after preliminary calibration is dynamically fed back to the error estimation vector, forming a highly stable adaptive timing error estimation vector after recursive optimization, thereby achieving accurate secondary correction of second-level timing errors induced by node-level occasional drift, batch transients and environmental disturbances.

[0146] For example, in a critical process timing correction scenario of a high-speed circuit board manufacturing line, the high-dimensional process parameter timing flow after initial calibration covers five main process nodes, with standardized timestamp sequences of [245600.4, 245602.1, 245600.1, 245600.1, 245609.6] ms. The theoretical order of the process nodes is 1-5. Using a recursive window width w=5 batches, a historical residual storage depth H=30, and a recursive learning rate α=0.2, the actual remaining synchronization offset of adjacent nodes within the window is calculated. The residual between node 1 and node 2 is 1.7 ms, the residual between node 2 and node 3 is -2.0 ms, and the residual between node 4 and node 5 is 9.5 ms. The initial error estimation vector [1.9, 2.4, 12.0, 0.1, 2.2] ms is recursively corrected using the recursive feedback correction formula. The new estimate for node 2 is (1-0.2)×2.4+0.2×1.7=2.22ms. This process is repeated multiple times, with a maximum of 6 recursive rounds (R=6) and a convergence tolerance of η=0.1ms. The residuals of most nodes are reduced to less than 0.08ms. Node 5 continues to have a small positive residual, which is continuously fine-tuned using a buffer factor β=0.05. Ultimately, the temporal drift error of the recursively optimized error estimation vector is less than 0.09ms across all nodes, and the residual sequence enters a steady-state region. The average correction cycle delay is less than 4ms. After this secondary correction, the maximum absolute value of the node timestamp error in subsequent causal chain inference is less than 0.10ms, significantly improving the temporal resolution and anomaly sensitivity of quality tracing.

[0147] By using a recursive optimization adaptive timing error estimation vector feedback mechanism, the preliminary calibration data from the previous step is transformed into dynamically corrected, highly stable error estimation parameters. This enables rapid suppression and precise dynamic correction of transient timing drift errors at each node in the high-dimensional process of circuit board manufacturing, laying a highly consistent foundation for error compensation closed-loop control and downstream process anomaly detection.

[0148] S4.4: The recursively optimized adaptive timing error estimation vector is input again into the adaptive timing error compensation algorithm to perform multiple rounds of dynamic iterative adjustment on the high-dimensional process parameter timing flow, and converge to obtain the final error-compensated high-dimensional process parameter timing dataset, ensuring that the timing arrangement of each data point conforms to the actual process flow to the greatest extent.

[0149] S4.5: Perform integrity and consistency verification on the high-dimensional process parameter time series dataset after final error compensation. Use the data consistency criterion module to detect missing, duplicate or non-converged nodes, and output a high-dimensional process parameter time series dataset with high time series accuracy and causal inference stability, providing quality assurance for subsequent abnormal parameter clustering and causal chain discrimination.

[0150] Step S5: For the high-dimensional process parameter time-series dataset after error compensation, based on process labels and geographical location information, cluster adjacent abnormal parameters within a preset event-driven dynamic window, and automatically identify the temporal causal relationship characteristics between parameters. Specifically, this includes: S5.1: Perform joint pairing of process labels and geographic location information on the high-dimensional process parameter time series dataset after error compensation, and use a label aggregation mechanism to generate parameter groups with geographic attributes and process stage characteristics to obtain process parameter clusters with spatiotemporal identifiers.

[0151] S5.2: Based on the process parameter cluster with spatiotemporal identifiers, an event-driven dynamic window is constructed. The abnormal fluctuation of key process parameters triggers the opening of the window. A threshold discrimination algorithm is used to monitor the time-series data stream in real time to dynamically locate the abnormal parameter clusters.

[0152] The input conditions are a high-dimensional process parameter time series dataset after error compensation and process parameter clusters generated by matching process labels and geographical location information. Each parameter cluster has a clear process stage identifier, equipment spatial location and complete time series index.

[0153] An event-driven dynamic window creation method is adopted (parameter: key process parameter screening set S). key (Initial window width Δt0, window adaptive expansion factor k, trigger threshold set Θ), for each process parameter cluster with spatiotemporal identifier, sequentially retrieve all parameters in S... key Based on batch number and node attributes, the key process parameters within the window are initially determined to be the monitoring targets of the dynamic window. Through high-precision time series extraction and process stage segmentation, the complete time series of each key process parameter within the window is obtained within a specified time period.

[0154] Furthermore, an abnormal fluctuation detection algorithm is applied (parameter: moving average window W). avg , criterion for change magnitude ΔP, threshold for parameter anomaly discrimination Θ i The difference between the current value and the moving average of each key parameter is calculated in real time, and the event trigger is determined by the following mathematical relationship:

[0155] in, For the first Key process parameters at time The absolute fluctuation amplitude, This is the observed value of the parameter at this moment. When a critical parameter is identified as having an abnormal transition, a dynamic window is created, and the time-series parameter stream of that segment is included in the monitoring.

[0156] An adaptive window expansion algorithm based on parameter clustering (parameters: current window width Δt, clustering statistic σ, maximum window width Δt) is used. max The density of outlier parameters within the time series dimension is dynamically evaluated, and the current clustering degree is derived using the following formula:

[0157] in, This represents the number of parameter points identified as abnormal within the current window. This is the window width. If... The window automatically moves to the right and increases in width. until or .

[0158] Furthermore, a real-time threshold discrimination mechanism is utilized (parameter: anomaly clustering threshold p). anomaly Minimum aggregation point n min (Parameter index filtering rules) are used to determine the abnormal attributes of each parameter point within the window segment, extract the abnormal parameter index and corresponding time label within each cluster segment, and achieve high-precision determination of abnormal parameter cluster segments in the dynamic window.

[0159] Then, through hierarchical screening of process nodes and joint comparison of multi-dimensional features (parameters: process stage level L, geographical location matching accuracy γ), geo The parameter time overlap (ρ) is used to perform spatial and process label level verification on the critical section to ensure that the abnormal clustering section corresponds accurately with the actual process flow and equipment spatial layout.

[0160] The above method transforms the high-dimensional process parameter time series dataset after error compensation into reasonably demarcated anomalous parameter clustering segments. Through multi-level causal logic reasoning, including real-time event triggering, adaptive window expansion, clustering degree discrimination, and segment selection, it ultimately provides a highly consistent spatiotemporal structured input for subsequent multidimensional clustering and time series correlation analysis of anomalous parameters.

[0161] For example, on a high-speed, multi-process circuit board manufacturing line, the input is a compensated high-dimensional process parameter time-series dataset (such as SMT, AOI, ICT sections). Each batch of parameter records contains 50,000 records, and the geographical location information is mapped to equipment numbers D01 to D12. The initial dynamic window width Δt0 = 30ms, the window adaptive expansion factor k = 1.5, and the maximum window width Δt max =150ms. Filter set S using key parameters. key (Covering six parameters including welding current and adsorption pressure), sliding average window W avg =7, Abnormal fluctuation threshold Θ i =2σi minimum threshold n for cluster point discrimination min =6. A sudden increase in welding current was detected in the SMT section at t=4089ms, and real-time calculations were performed. If the value is greater than Θ1=3.2A, a window is immediately triggered, initially covering 30ms of data. (The text then abruptly shifts to a different topic: abnormal clustering within the window.) The window was shifted to the right and expanded to 75ms, ultimately extracting 14 abnormal parameter point indices. Combined with the geographical location D07 and the SMT process node, the abnormal segment was determined to completely correspond to the actual placement segment. Subsequent multidimensional clustering and causal chain reconstruction used this segment as input, achieving an overall cluster segment detection rate of 97.8%, an average response latency of less than 4ms, and an average window adjustment frequency of 2.5. Its robustness and real-time performance were significantly better than the static window strategy.

[0162] S5.3: Multidimensional feature clustering algorithm is used to cluster the time series data of process parameters in the dynamic window. K-means or density peak clustering methods are used to identify adjacent abnormal clusters between parameters, realize the spatial aggregation of abnormal process parameters, and output the abnormal parameter clusters in the dynamic window.

[0163] For processing high-dimensional process parameter time series datasets based on event-driven dynamic windows, this step aims to focus on the automatic spatial aggregation of outlier parameters, providing structured, high-confidence input for the correlation and causal feature analysis of outlier parameters.

[0164] The input conditions are: a high-dimensional process parameter time series that has been output by the previous steps (S5.1, S5.2), aggregated with error compensation and tag geographic features, and restricted to a single event-driven dynamic window.

[0165] A parameter representation mechanism combining multi-feature normalization mapping and heterogeneous data input (parameters: feature dimension set F = {process label, equipment number, time series index, geographic coordinates, process stage, physical parameter value}) is adopted. Standard fraction transformation is applied to each parameter data point within the window to achieve numerical normalization. The mapping function is as follows:

[0166] in, For parameters In features The normalization result is as follows: These are the original parameters. , These are the sample mean and standard deviation of this feature, respectively.

[0167] The K-means clustering algorithm (parameters: number of clusters K, automatically determined by the data volume within the window and parameter dissimilarity analysis; maximum clustering iteration step L) is used to aggregate the above normalized parameter space. Initial centroid values ​​are generated heuristically using K-means++, and Euclidean distance is used as the distance metric. The definition is as follows:

[0168] in, and Parameters and In the Normalized components under dimensional features, total dimension.

[0169] Furthermore, the centroid coordinates are iteratively optimized. To implement the class labeling of samples within a window The automatic convergence satisfies the minimization of the sum of squared errors:

[0170] in, For the first A collection of sample classes.

[0171] Moving to the next level, the density peak clustering method is used (parameter: local density kernel width). Distance threshold Furthermore, on the candidate clusters output by K-means, density-distance joint constraints are used to identify outlier clusters: The local density is calculated as follows:

[0172] in, For the sample The local density.

[0173] Calculate the minimum distance from each point to a point with higher density:

[0174] In the polymerization results, density The largest point and the greatest distance The points together serve as cluster centers.

[0175] By progressively introducing K-means high-dimensional spatial grouping into density peak identification through the aforementioned stages, decoupling of complex spatiotemporal anomaly parameter clusters is achieved. The quality evaluation of various clusters is then gradually improved using silhouette coefficients. Davies-Bouldin Index Auxiliary constraints:

[0176]

[0177] in, The average distance between samples within the cluster. The average distance from the sample to the nearest other cluster. For the first Average distance between classes This represents the distance between the centers of the two clusters.

[0178] The selection of the optimal clustering scheme makes maximum, The result with the smallest parameters.

[0179] Through the five-level progressive process of multidimensional feature normalization, spatial clustering, density recursion, and cluster evaluation, the final output is a cluster of process parameter anomalies with high confidence within the spatiotemporal range of the dynamic window. This provides a solid foundation for the comprehensive quantification of high-dimensional anomalies in subsequent time-series dependencies and causal discrimination. Furthermore, the spatial grouping results directly correspond to equipment location information and process stage labels, meeting the engineering requirements for automated detection and anomaly localization in high-dimensional scenarios.

[0180] For example, in the AOI node of a high-speed PCB multi-segment production line, a single batch Six parameters (process label, equipment number, process stage, timestamp, welding temperature, optical anomalous rate), normalized according to standard scores, and K-means adaptive clustering number. Iteration step K-means++ initialization. After Euclidean distance clustering, 5 preliminary categories are generated, with most normal process intervals concentrated in the main cluster. The cluster centers focus on the time period [4800, 4830] ms of the AOI process stage, equipment D03, and welding temperature above the mean. Then, the density peaking algorithm... Three high-density anomaly core clusters were aggregated. The core anomaly clusters were partitioned at the AOI node and D03 device at [4809,4824] ms; the profile coefficients DBI=0.91. The final output is the parameter index corresponding to the cluster label. The anomaly parameter detection accuracy reaches 96%, the false negative rate is within 2%, the spatial positioning accuracy is 100%, and the matching degree of downstream time-series correlation feature analysis is improved by 12%.

[0181] S5.4: For the clustering of abnormal parameters in the dynamic window, apply time-series correlation analysis algorithms (such as lag correlation coefficient analysis) to calculate the time dependence characteristics between process parameters and form a preliminary time-series correlation causal feature set between parameters.

[0182] S5.5: Based on the preliminary time-series related causal feature set, combined with historical parameter dependency patterns and process label sequence information, the direct or indirect time-series causal relationship features between parameters are automatically inferred and strengthened through causal discovery algorithms (such as Granger causality test or structural equation model-based methods), generating a complete time-series causal relationship feature set of process parameters within a dynamic window.

[0183] Step S6: Input the causal correlation features of the parameters within the dynamic window into the causal inference modeling network, reconstruct the multi-level process causal links between parameters through multi-dimensional causal chain inference, and simultaneously identify causal chain breakpoints and time-series anomaly nodes. Specifically, this includes: S6.1: The causal relationship features of parameters within the event-driven dynamic window are processed into feature vectors. Causal embedding technology is used to map discrete indicators, process labels, and geographic location information into high-dimensional causal feature vectors to obtain a data input format suitable for causal inference modeling networks, providing high-dimensional causal feature input for subsequent multi-level process causal link construction.

[0184] S6.2: Based on causal feature vectors, the time-series graph neural network module is used to perform multi-level causal correlation modeling on the causal features within the dynamic window to generate a preliminary multi-dimensional process parameter causal link structure, realizing the causal structure mapping from the original process parameter causal correlation features to the preliminary causal link structure.

[0185] For high-dimensional parametric causal feature vector inputs processed by causal embedding, containing process labels, geographical location information, and temporal dependency features, possessing structured and normalized characteristics, a temporal graph neural network modeling method is adopted (parameter: node feature dimension d). in Number of hidden layer units h, time step T, aggregation function type agg type This ensures that high-dimensional causal features within all dynamic windows are efficiently mapped in the network structure: causal, temporal, and spatial dependencies of each process parameter node are modeled in parallel.

[0186] Furthermore, a dynamic graph generation algorithm is used to generate process parameter nodes and timing-dependent edges (parameters: node determination rules, edge weight function w). e (i,j), temporal Lagrange constraint λ seq The system automatically constructs a multidimensional parametric graph G = (V, E), with process parameters as nodes and temporal correlation and spatial proximity as edges. Edge weights are calculated as follows:

[0187] in Let i and j be the time-series correlation coefficients. Geographical distance This is a label consistency factor for the process stage. These are the weighting coefficients. The weights of causal edges can be adaptively adjusted in real time based on temporal consistency and spatial mapping accuracy.

[0188] Furthermore, utilizing the message passing mechanism of graph neural networks (parameters: aggregation round number l, activation function a(x)), node embeddings are updated within each aggregation round through recursive embedding aggregation of node neighborhoods. :

[0189] in Aggregation functions used for the l-th layer, such as adaptive weighted mean, max pooling, or graph attention.

[0190] Based on the high-dimensional representation of node embedding, further decoding is performed using the causal chain structure method with process parameters (parameter: maximum depth d). max (Minimum edge weight γ, hierarchical attribution marker β), combining the aggregation score between nodes and structural connectivity, recursively traverse the causal graph, extract parameter pairs with edge weights not less than γ, and label the causal direction and link level:

[0191] During the decoding process, the data above is used as the main thread to form a multi-level causal chain, with the causal direction of each pair of nodes as the main thread.

[0192] Furthermore, through causal consistency scoring indicators Consistency checks are performed on all preliminary link sets, and the results are calculated using the Pearson correlation coefficient and Granger time-series correlation score, as follows:

[0193] in The Pearson correlation coefficient between the process parameters is given. For Granger causality fractions, For normalized weights.

[0194] Links that do not meet the consistency threshold are pruned to remove weakly correlated causal paths. The complete process described above enables the adaptive generation of high-dimensional causal chain structures, providing a spatially hierarchical multi-dimensional causal chain structure for causal chain integrity verification and subsequent defect root cause localization, thereby improving inference accuracy, temporal consistency, and structural robustness.

[0195] For example, in a multi-batch, multi-segment circuit board manufacturing process, the input is a 1024-dimensional causal feature vector of process parameters within a dynamic window (including SMT segment parameters, equipment number, outlier index, timing lag, geographic coordinates, process stage, etc.). The feature dimension d of the configured graph neural network node is then used. in =32, hidden layer unit number h=128, time step T=8. Implement ternary edge weight parameters. Edge weight threshold γ=0.82, maximum chain depth d max =5. In actual model execution, the average convergence score for node embedding was 0.89, with 12 main causal chains and 46 branch causal chains outputting, an average link depth of 3.8 layers, and a Pearson-Granger consistency score of 5. With an average value of 0.93 and a noise link elimination rate of 96.2%, the causal chain hierarchy is highly consistent with the actual process flow, and the subsequent defect tracing subsystem improves the accuracy of reconstruction by 15%. Ultimately, the output preliminary multidimensional process causal chain structure effectively supports the automatic quantification and time-series tracing of causal relationships between process parameters in complex dynamic process scenarios.

[0196] S6.3: Input the multi-dimensional process parameter causal link structure into the causal chain integrity verification mechanism, and identify the interruption points and information break segments in the causal transmission through the chain transmission consistency judgment algorithm, so as to form a multi-level process causal link node set with initial anomaly identification.

[0197] S6.4: Perform timing error evaluation on a multi-level process causal link node set with initial anomaly identification based on an adaptive timing consistency analysis method. Compare the actual timing of nodes in the causal chain inference process with the theoretical timing in the standardized process parameter timing dataset to automatically identify timing anomaly nodes in the causal chain and realize the consistency identification between the causal link structure and the actual timing.

[0198] S6.5: Perform a causal knowledge graph update operation on the multi-level process causal link structure and its abnormal node labeling results after time-series consistency analysis, and generate the final multi-level process causal link data containing node causal attributes, link abnormal points and time-series abnormal annotations, providing process causal reasoning network output for subsequent defect tracing and quality decision factor generation.

[0199] Step S7: Based on the reconstructed multi-level process causal chain, automatically determine whether the current process anomaly point meets the preset defect tracing conditions. If it does, locate the defect source parameters and generate quality problem tracing decision factors. Specifically, this includes: S7.1: For each process anomaly point in the reconstructed multi-level process causal link, based on the parameter causal correlation characteristics output by the causal reasoning modeling network, and according to the preset defect tracing conditions, a set of target anomaly points that meet the process characteristic anomaly criteria is obtained.

[0200] S7.2: For the set of target anomalies that meet the preset tracing conditions, based on the parameter causal link structure, the link recursive analysis algorithm is applied to trace back layer by layer from top to bottom to locate the source parameter node of the corresponding causal chain in order to extract the potential defect source parameters.

[0201] S7.3: Based on the extracted defect source parameters, a linkage analysis mechanism of process tags and geographical location information is applied to cross-determine the actual process stage attributes of abnormal nodes and generate multi-dimensional parameter tags for defect sources with process identifiers and spatial markers.

[0202] Obtain the defect source parameter nodes and their parameter sets located based on the link recursive analysis algorithm, and adopt a process tag-location dual mapping mechanism (parameters: node parameter ID, process tag index, spatial location information) to achieve preliminary binding of defect node parameters with actual process stages and spatial distribution.

[0203] Furthermore, through the automated retrieval module of process tag sequence (parameters: tag encoding rules, priority weight), the parameters of the located defect source are matched with the main process line tags and sub-process segment tags respectively, so as to realize the explicit labeling of parameters in the process stage sequence and obtain the process stage attribute statistical vector.

[0204] Furthermore, based on the spatial distribution clustering analysis algorithm (parameters: geographic coordinate vector, equipment area partitioning table, cluster radius threshold), high-precision clustering mapping is performed on the geographic location information of defect parameter nodes to generate spatial attribute identification factors, ensuring that each source parameter has a unique spatial field affiliation.

[0205] Furthermore, through a cross-association discrimination mechanism between process labels and spatial attributes (parameters: label spatial mapping relationship matrix, conflict discrimination rules), the consistency between process labels and geospatial information is verified, enabling automatic correction of anomalies such as label misalignment and equipment zoning ambiguity, and generating corresponding conflict markers for abnormal nodes.

[0206] The multidimensional parameter label generation module (parameters: process label, spatial attribute, conflict flag) integrates the process label sequence, geospatial markers and conflict discrimination results to output a multidimensional parameter label for the defect source with process identification and spatial markers, thereby realizing multidimensional and traceable attribute labeling of defect parameter nodes.

[0207] Through the linkage analysis mechanism of process labels and geographic location information, the defect source parameters extracted in the previous step are transformed into multi-dimensional parameter labels with process attributes and spatial indications, so as to realize the determination of defect root cause attribution at the fine-grained process stage and spatial domain.

[0208] For example, in the production of a batch of multilayer circuit boards, the defect source nodes located by the link recursive analysis algorithm are parameters P17 and P43 (board lamination temperature and AOI inspection location), with process labels of

Lamination - Secondary Curing

Inspection - AOI II

[0209] S7.4: Collect intermediate data containing multi-dimensional parameter labels of defect sources and causal chain inference results, and generate a source tracing feature vector that matches the current source tracing status based on the fault mode discrimination model, which serves as the input to the decision factor generation algorithm.

[0210] S7.5: Input the source traceability feature vector into the decision factor generation module, combine it with historical source traceability data and model evaluation function, calculate and output the source traceability decision factor for the current anomaly point, and provide a basis for subsequent visualization output and operation and maintenance response.

[0211] Step S8: Input the quality problem tracing decision factors and causal link structure data into the visualization analysis system to generate a graphical defect tracing report in real time, and automatically output quality warning signals based on the consistency or anomaly of the causal chain sequence. Specifically, this includes: S8.1: The input quality problem tracing decision factors and high-dimensional causal link structure data are preprocessed by data integration. A multi-dimensional data fusion algorithm is used to synchronously map the process parameter tracing decision factors and the causal link topology according to the node-edge attribute fields to obtain a multi-dimensional fused causal chain data structure that can support graphical analysis.

[0212] S8.2: Based on the obtained multidimensional fusion causal chain data structure, a visualization graph layout algorithm (such as force-directed method and hierarchical layout method) is applied to perform node clustering and hierarchical coordinate transformation on the causal link structure data to achieve hierarchical and scalable layout output of the high-dimensional process parameter causal chain network.

[0213] S8.3: The decision factor-driven visualization information encoding method is adopted to make the traceability decision factor marker nodes in the causal link structure significant. The defect traceability points are distinguished from ordinary causal link nodes by means of node color, shape or size, so as to form a graphical defect traceability result with decision guidance function.

[0214] The input includes process parameter traceability decision factors and causal link structure data processed by a multidimensional data fusion algorithm. The data has been synchronously mapped through node-edge attribute fields to form a multidimensional fused causal chain data structure that supports graphical analysis.

[0215] A decision factor-driven saliency labeling method (parameters: node decision factor weight vector, saliency threshold, node basic attribute set) is adopted to assign decision factor values ​​to each node in the causal link structure and set the labeling level accordingly to achieve priority highlighting of salient nodes.

[0216] Furthermore, by using a node-based classification-based information encoding strategy (parameters: decision factor threshold grading rules, anomaly category code, node importance level), each node in the structure is grouped according to the magnitude of the source tracing decision factor, defect type, and importance, and a node saliency grouping table is established, with different visual attributes configured for each group.

[0217] Furthermore, a color mapping coding method (parameters: node category color palette, warning level color matching standard) is adopted to assign high-contrast warning colors (such as red and orange) to defect tracing point nodes, and neutral colors (such as gray and blue) to ordinary causal chain nodes. The warning nodes are then used to reflect the numerical intensity of decision factors through gradient color levels to achieve visual guidance.

[0218] Furthermore, by using node shape encoding and size adjustment methods (parameters: node size factor, abnormal shape template set), defect tracing point nodes are displayed as special shapes such as polygons, stars, or triangles, and the node radius / side length is set according to the weight of the decision factor. Ordinary nodes are displayed as standard circles or squares, thereby improving the recognizability of node categories.

[0219] Furthermore, based on the node attribute mapping results, the node saliency attribute set (including color, shape, and size triples) is injected into the causal chain graph data structure to generate a graphical node feature vector that integrates multi-dimensional annotation information.

[0220] Through the above algorithm processing, the defect tracing points with significant weights in the process causal chain network are encoded in a way that enhances the distinction between them and ordinary causal nodes. The output is a graphical defect tracing result that has both decision guidance and anomaly warning functions, providing a highly recognizable visual data foundation for subsequent consistency analysis and quality early warning.

[0221] For example, in a multi-layer causal chain analysis of a batch of circuit board manufacturing, the input parameters include a set of traceability decision factors {D17=0.92, D43=0.65, D51=0.15}, with corresponding node classification thresholds set to 0.8 (high risk) and 0.5 (medium risk). Node color coding is high-risk red (#FF3333), medium-risk orange (#FFA500), and ordinary nodes blue (#3399FF). Node shapes are assigned as follows: high-risk nodes are hexagonal stars, medium-risk nodes are pentagonal, and ordinary nodes are circular. The node radius depends on the decision factor (maximum radius 30px, minimum 10px). In actual processing, node D17 is prominently marked with a hexagonal star and bright red with a radius of 28px; node D43 is marked as an orange pentagon with a radius of 20px; and D51 and below are marked as blue dots with a radius of 12px. The output causal chain map clearly distinguishes two traceability defect points, and nodes with high decision factors are prominently displayed on both the platform and in the report. Performance tests show that the accuracy of source identification nodes reaches 99%, and the average user interpretation time is reduced to less than 3 seconds, effectively supporting the decision-making of quality anomaly response and graded disposal.

[0222] S8.4: Input the visualized causal link structure data with significant annotations into the dynamic temporal consistency analysis module. Based on the node temporal attributes and the global connectivity index of the causal chain, automatically identify the temporal consistency or anomaly of the causal chain network and generate a temporal consistency analysis report that is synchronized with the actual process status changes.

[0223] S8.5: Based on the output results of the time-series consistency analysis report and the quality problem tracing decision factors, the abnormal state-driven early warning rule engine is applied to automatically trigger real-time quality early warning signals for causal chain nodes that are judged to be abnormal and links that cannot meet the time-series consistency conditions. The early warning signal is then attached to the corresponding node in the graphical tracing report for visualization.

[0224] S8.6: Performs multi-format output configuration processing on the generated graphical defect tracing report and embedded quality early warning signal, supports Web visualization dashboard, PDF report and third-party MES system interface integration, so as to achieve efficient tracing across multiple terminals and platforms and collaborative distribution of upstream high-dimensional causal chain data.

[0225] Step S9: Monitor the dynamic changes of subsequent process parameters. If a process switch or change in environmental conditions is detected, automatically adjust the timing error estimation vector and causal chain inference model parameters to achieve adaptive traceability optimization for cross-process and cross-batch manufacturing scenarios. Specifically, this includes: S9.1: For the high-dimensional process parameter time series dataset reconstructed by the causal chain inference modeling network, call the dynamic monitoring module to continuously obtain the real-time dynamic change information of process parameters at each node in the process flow, so as to generate a dynamic sequence of process parameters.

[0226] S9.2: Based on the dynamic sequence of process parameters, apply the process flow switching identification algorithm and the environmental condition change detection strategy to detect the process flow state transition and environmental variable dynamic changes in real time, so as to output the process flow switching signal and environmental condition change event.

[0227] S9.3: Taking process flow switching signals and environmental condition change events as inputs, and based on preset adaptive rules, calling the model parameter adjustment module to perform adaptive correction processing on the current timing error estimation vector to obtain the timing error estimation vector after real-time correction.

[0228] S9.4: Based on the time-series error estimation vector that has been corrected in real time, the adaptive model parameter update algorithm is used to dynamically adjust the parameters of the causal chain inference model, so that the parameters of the causal chain inference model are adapted to the current process stage and environmental conditions, and the output of causal inference robustness is optimized.

[0229] S9.5: Apply the optimized causal chain inference model parameters to the continuous inference task of the high-dimensional process parameter time series dataset to generate a multi-level process causal chain after adaptive tracing optimization, which serves as the basis for the final tracing decision in cross-process and cross-batch manufacturing scenarios.

[0230] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0231] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0232] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A circuit board quality traceability method based on big data, specifically including: S1: Collect time series data of multi-source process parameters during the circuit board manufacturing process, and record the geographical location information and process labels of the acquisition equipment to form a multi-dimensional time series raw dataset; S2: Normalize and denoise the timestamps of each process node in the multidimensional time series raw dataset to generate a standardized process parameter data stream; S3: Based on standardized process parameter data streams, construct a multi-source error model, perform statistical evaluation, and generate adaptive timing error estimation vectors for each node; S4: Input the timing error estimation vector and the standardized process parameter data stream into the adaptive timing error compensation algorithm to obtain the high-dimensional process parameter timing dataset after error compensation; S5: For the high-dimensional process parameter time series dataset, cluster adjacent abnormal parameters within a preset event-driven dynamic window, and automatically identify the time-series causal relationship features between the parameters; S6: Input the causal relationship features of the parameters in the dynamic window into the causal reasoning modeling network, reconstruct the multi-level process causal links between the parameters, and mark the causal link breakpoints and time sequence anomaly nodes. S7: Based on the reconstructed multi-level process causal chain, automatically determine whether the current process anomaly point meets the preset defect tracing conditions, and generate quality problem tracing decision factors; S8: Input the quality problem tracing decision factors and causal link structure data into the visualization analysis system, and automatically output quality early warning signals for the consistency or anomaly of the causal link time sequence; S9: Monitors the dynamic changes of subsequent process parameters and automatically adjusts the timing error estimation vector and causal chain inference model parameters.

2. The circuit board quality traceability method based on big data according to claim 1, characterized in that, Step S1 specifically includes: Multiple types of sensor acquisition devices are deployed at all process nodes in the manufacturing process to obtain raw signal data of process parameters through industrial fieldbus protocol, and to obtain raw signals of multi-source process parameters including temperature, humidity, current and voltage. Based on the original signals of the process parameters, the local time of each sensor acquisition device is synchronized with the network time reference in real time to unify the timing reference of the original signals of the process parameters. Acquire the raw process parameter signals with unified timing reference, and attach the geographical location information of the acquisition device to each data packet; For data streams with accompanying geolocation, process tag information is automatically extracted and recorded to form tagged process parameter data with node identity and process stage; Based on the tagged data of process parameters, the data stream is structured and integrated according to the process node sequence and acquisition time sequence to construct a multi-dimensional time-series raw dataset.

3. The circuit board quality traceability method based on big data according to claim 1, characterized in that, Step S2 specifically includes: Based on multi-source process parameter time series data, the original timestamp information of each process node is extracted and batch-level timestamp aggregation is performed. The extracted raw timestamp sequence is processed by removing outliers and acquisition jitter noise to obtain cleaned timestamp data. The purification timestamp data is compared with the network time base, and the clock offset between devices is synchronized and adjusted to generate an aligned unified standard timestamp sequence. Using a unified standard timestamp sequence, a normalized mapping is performed on the time matching relationship of the original process parameter data stream to form a normalized process parameter matrix; Based on the normalized process parameter matrix, abnormal time records caused by acquisition synchronization delay are corrected and eliminated to obtain a preliminary standardized process parameter data stream. The initial standardized process parameter data stream is input into the end-to-end consistency verification module, which performs multi-dimensional verification of the timing logic of each process node and outputs the final standardized process parameter data stream.

4. The circuit board quality traceability method based on big data according to claim 1, characterized in that, Step S3 specifically includes: A preliminary statistical analysis is performed on the timestamp data of all process nodes in the standardized process parameter data stream to form a set of timestamp distribution characteristic parameters. Based on the timestamp distribution feature parameter set, the time reference offset between detection devices corresponding to each process node is quantified, and the synchronization error feature matrix between devices is constructed. For standardized process parameter data streams, by comparing the event trigger time of each node under the same process flow with the network reference clock, the acquisition delay feature quantity is calculated in real time to form a node-level acquisition delay statistical feature vector. By integrating the synchronization error feature matrix between devices with the statistical feature vector of node-level acquisition delay, a statistical probability distribution model of timestamp drift and delay components of each node is output. Based on the multi-source error model, adaptive timing error estimation vectors for each process node are dynamically generated, and the numerical ranges of standardized process parameter data streams are correlated in real time to obtain the timing error characteristics and outputs under different time periods and different process conditions.

5. The circuit board quality traceability method based on big data according to claim 1, characterized in that, Step S4 specifically includes: Obtain the standardized process parameter data stream and the adaptive timing error estimation vector as the joint input of the adaptive timing error compensation algorithm; An adaptive timing error compensation algorithm is used to perform global timing correction on the timestamp information of each process node in the standardized process parameter data stream to obtain a high-dimensional process parameter timing stream after preliminary calibration. Based on the high-dimensional process parameter time flow after preliminary calibration, the timing deviation residuals between each process node are dynamically calculated, and the adaptive timing error estimation vector is recursively optimized. The recursive optimized adaptive timing error estimation vector is input again into the adaptive timing error compensation algorithm to perform multiple rounds of dynamic iterative adjustment on the high-dimensional process parameter timing flow, and finally obtain the high-dimensional process parameter timing dataset after error compensation. The integrity and consistency of the high-dimensional process parameter time series dataset after error compensation are checked, and a high-dimensional process parameter time series dataset with high time series accuracy and causal inference stability is output.

6. The circuit board quality traceability method based on big data according to claim 3, characterized in that: In step S2, the original timestamp information of each process node is processed by timestamp extraction, batch-level timestamp aggregation, time-series denoising and consistency verification, and the timestamps between devices are synchronized and normalized based on Network Time Protocol (NTP) or Precision Time Protocol (PTP).

7. The circuit board quality traceability method based on big data according to claim 4, characterized in that: In step S3, based on the extracted process node timestamp distribution parameters, the device synchronization error feature matrix and the node-level acquisition delay statistical feature vector are used. Through data fusion, weighted modeling, Gaussian mixture distribution modeling and confidence interval evaluation, the adaptive timing error estimation vector of each node is generated.

8. The circuit board quality traceability method based on big data according to claim 5, characterized in that: In step S4, by introducing an adaptive error feedback optimization mechanism, and combining dynamic calculation of node timing deviation residuals with recursive feedback correction, a high-dimensional process parameter timing dataset with high consistency and accurate sorting after error compensation is output.

9. The circuit board quality traceability method based on big data according to claim 1, characterized in that: In step S6, the causal relationship features of the parameters in the event-driven dynamic window are input into the time-series graph neural network after causal embedding, and a multi-level process causal link structure is automatically generated. The weakly correlated links are screened out by the causal consistency scoring index, and the causal chain breakpoints and abnormal nodes are automatically labeled.

10. The circuit board quality traceability method based on big data according to claim 1, characterized in that: In step S7, the defect source parameters are traced through the link recursive analysis algorithm, and the anomaly attribution is determined by combining process tags and geographical location information, generating a multi-dimensional parameter tag for the defect source with process identification and spatial marker.

Images