Product full-process quality monitoring method or system based on big data

Abstract

The invention relates to the technical field of computers, discloses a product full-process quality monitoring method and system based on big data, and aims to solve the problem of insufficient quality management refinement caused by data islands, monitoring lag, tracing difficulty and single analysis dimension in the prior art. The method comprises the following steps: constructing a full-process data acquisition system covering design, materials, production, logistics and after-sale, integrating multi-source heterogeneous data, and performing cleaning and time alignment; using an LSTM and GRU mixed model to extract time sequence features of the process parameters, and generating a high-dimensional process feature vector; and constructing a material-process-quality association knowledge graph in combination with the graph attention network, and representing the internal association of the quality influence factors of each link. According to the scheme, full-process data fusion, active quality prediction, accurate traceability and intelligent optimization are realized, and the real-time performance, the accuracy and the self-adaptive capability of quality control are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically relating to a product end-to-end quality monitoring method or system based on big data. Background Technology

[0002] With the deep integration of information technology and modern manufacturing, product quality has become a key indicator for measuring a company's core competitiveness. Excellent quality management runs through the entire product lifecycle, from design and development, material procurement, production and manufacturing to market distribution and after-sales service, ensuring that quality is controllable at every stage of the process, which is crucial for enhancing brand reputation and customer satisfaction.

[0003] Among them, the product full-process quality monitoring aims to achieve real-time monitoring and dynamic management of the product's quality status at each stage of its life cycle through systematic data collection and analysis. Its core objective is to break down data barriers between different links, form a complete quality information chain, and provide a basis for quality improvement and decision support.

[0004] Existing technologies typically face numerous challenges in conducting end-to-end product quality monitoring. First, data sources are fragmented; data standards differ across stages, such as R&D, production, supply chain, and customer feedback, creating information silos and hindering effective correlation analysis. Second, monitoring methods are passive, relying heavily on post-event sampling and lacking real-time early warning and quality trend prediction capabilities based on process data, leading to delayed problem response. Third, quality traceability is difficult; when quality issues arise, incomplete data chains make it difficult to quickly and accurately pinpoint the root cause and specific batch. Finally, analytical dimensions are limited; there is a general lack of in-depth mining capabilities for massive amounts of multi-source heterogeneous data, failing to effectively reveal the complex nonlinear relationships between process parameters, environmental factors, and final product quality. These bottlenecks collectively constitute the technical challenges that urgently need to be addressed in the refined management of product quality. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a product end-to-end quality monitoring method or system based on big data, which can effectively solve the problems in the background technology. Traditional product quality monitoring methods mainly rely on sampling inspection and offline analysis, which suffer from problems such as monitoring lag, narrow coverage, and inability to trace the root cause of quality problems in real time. Specifically, firstly, the data dimensions in the production process are numerous, including equipment status, environmental parameters, material batches, process parameters, etc. Traditional methods are difficult to effectively integrate and correlate these heterogeneous and massive amounts of data, resulting in insufficient early warning and diagnosis capabilities for quality problems. Secondly, quality monitoring is often limited to the production stage and fails to cover the entire process from product design, material procurement, manufacturing, warehousing and logistics to after-sales service, forming data silos and making it extremely difficult to trace the root cause of quality problems. Finally, existing quality control models are mostly static models, which cannot adapt to the dynamic changes and complex disturbances in the production process, resulting in low model prediction accuracy and difficulty in achieving accurate online quality control and process parameter optimization.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In one aspect, a product end-to-end quality monitoring system based on big data, the system comprising the following components: The end-to-end data acquisition and preprocessing module is used to build an end-to-end data acquisition system covering product design, materials, production, logistics and after-sales. It acquires and integrates multi-source heterogeneous data in real time from multiple links. After preprocessing operations such as data cleaning, format conversion and timestamp alignment, it forms a time-series and structured end-to-end quality dataset. The process time-series feature extraction module is used to extract time-series features from key process parameters and equipment status data in the production process based on the full-process quality dataset and using a deep time-series network model. It deeply mines the time dependence and dynamic change patterns in the data and generates a high-dimensional process feature vector that can characterize the real-time status of the production process. The Material-Process-Quality Association Knowledge Graph Construction Module is used to combine the high-dimensional process feature vector with product design data, material batch data, and quality inspection result data, and use graph neural network technology to construct a material-process-quality association knowledge graph. This knowledge graph explicitly represents the intrinsic association and transmission path between various quality influencing factors in the entire product process in the form of entity nodes and relation edges. The intelligent prediction and early warning module for quality risks is used to establish a quality risk prediction model based on the material-process-quality association knowledge graph. By aggregating global information in the graph, it can predict potential quality defects in future product batches in real time and provide risk classification and early warning based on the prediction results. The automated root cause tracing module is used to execute a backpropagation tracing algorithm on the knowledge graph when a quality anomaly is detected or an early warning signal is received. Through attribution analysis, it automatically locates the key influencing factor nodes and core association paths that lead to quality problems. The closed-loop feedback and process optimization module is used to generate targeted dynamic optimization and adjustment strategies for process parameters based on the tracing results of the key influencing factor nodes and related paths, and to distribute the strategies to the production execution system to realize closed-loop feedback control and continuous improvement of quality issues.

[0007] On the other hand, a product end-to-end quality monitoring method based on big data includes the following specific steps: Step S110: Construct a product full-process quality data collection system, acquire and integrate multi-source heterogeneous data in real time from multiple links such as design, materials, production, logistics and after-sales, and form a time-series and structured full-process quality dataset. Step S120: Based on the full-process quality dataset, use a deep temporal network model to extract temporal features from key process parameters and equipment status data in the production process, and generate a high-dimensional process feature vector. Step S130: Combining the high-dimensional process feature vector with product design data, material batch data, and quality inspection result data, a knowledge graph of material-process-quality association is constructed using graph neural network technology. The knowledge graph represents the intrinsic association between quality influencing factors in the form of entity nodes and relation edges. Step S140: Based on the material-process-quality association knowledge graph, establish a quality risk prediction model to predict and classify potential quality defects of the product in real time. Step S150: When a quality anomaly is detected or an early warning signal is received, a backpropagation tracing algorithm is executed on the knowledge graph to automatically locate the key influencing factor nodes and related paths that lead to the quality problem. Step S160: Based on the key influencing factor nodes and associated paths, generate a dynamic optimization and adjustment strategy for process parameters, and send the strategy to the production execution system to achieve closed-loop feedback control of quality issues.

[0008] Preferably, the full-process data acquisition and preprocessing module achieves comprehensive data capture across the entire process through industrial Internet of Things (IIoT) sensors (such as temperature, pressure, and vibration sensors) deployed on the production line and data interfaces that interface with Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), Product Lifecycle Management (PLM), and Customer Relationship Management (CRM) systems. When the multi-source heterogeneous data enters the preprocessing process, it first undergoes data cleaning, including identifying and processing outliers during data acquisition using the 3-sigma principle or local anomaly factor algorithm, and handling missing values ​​using spline interpolation or mean imputation. Subsequently, format conversion is performed, uniformly converting unstructured or semi-structured data (such as text documents and images) from different sources into structured data tables. Finally, timestamp alignment is performed, precisely aligning all time-series data with a unified baseline timeline to ensure data consistency and comparability across time dimensions, providing a high-quality data foundation for subsequent time-series analysis.

[0009] Furthermore, the deep temporal network model employed by the process temporal feature extraction module is a hybrid parallel structure combining a Long Short-Term Memory (LSTM) network and a Gated Recurrent Unit (GRU). This structure feeds the input temporal data (such as continuous temperature curves) into the LSTM and GRU channels in parallel. The LSTM channel, through its internal input, forget, and output gates, effectively learns and captures long-term dependencies in the production process, such as the performance degradation of equipment over time and the slow, cumulative impact of seasonal environmental temperature and humidity changes on product quality. The GRU channel, through its update and reset gates, focuses on capturing short-term fluctuations in the production process with fewer parameters and higher computational efficiency, such as instantaneous operational differences among operators, minor batch-to-batch differences in materials, or parameter disturbances during equipment start-up and shutdown. Finally, the feature vectors output from the LSTM and GRU channels are concatenated or weighted and fused to form a more comprehensive and robust high-dimensional process feature vector that includes both long-term trends and short-term dynamics.

[0010] Preferably, the graph neural network technology used in the material-process-quality association knowledge graph construction module is specifically a graph attention network (GAT). When constructing the knowledge graph, material batches, equipment units, process parameter setpoints, actual process parameter values, operators, environmental indicators, intermediate quality inspection items, and final quality inspection items are defined as entity nodes. Physical connections between them (e.g., the output of equipment A is the input of equipment B), process flow relationships (e.g., process A is performed before process B), and causal relationships discovered through prior knowledge or data mining are defined as relationship edges. When propagating information, the graph attention network does not treat all neighboring nodes equally. Instead, it uses a self-attention mechanism to calculate an attention weight for each neighboring node. This weight reflects the importance of the neighboring node's information in the current central node's update task. Through this mechanism, the model can dynamically learn and identify which influencing factors are more strongly correlated in specific quality scenarios, thereby constructing a weighted, more accurate, and more physically interpretable association knowledge graph.

[0011] Furthermore, the quality risk intelligent prediction and early warning module employs a graph-level classification or regression model based on Graph Convolutional Networks (GCNs). This model takes the node feature matrix and adjacency matrix of the knowledge graph as input. Through multi-layer graph convolution operations, each node continuously aggregates information from its neighboring nodes, enabling each node's representation vector to capture its local structural information within the graph. After multi-layer propagation, the representation vector of the top-level node effectively integrates graph information from a wider range, even globally. The final representation vectors of all relevant nodes representing a specific product batch are then subjected to graph pooling (e.g., average pooling) to obtain the graph-level representation of that product batch. This representation ultimately outputs the probability of a specific quality defect occurring in that batch of products (classification task) or a specific predicted value for a key quality indicator (regression task) through one or more fully connected layers. The early warning classification sets multiple thresholds based on the predicted defect probability or the degree of deviation between the predicted value and the standard value. For example, a probability of 0-0.1 is set as "normal", 0.1-0.3 is set as "attention", 0.3-0.6 is set as "warning", and above 0.6 is set as "serious" to achieve differentiated risk management.

[0012] Furthermore, the automated root cause analysis module employs a backpropagation-based graph-based attribution algorithm. When the quality prediction result of a product batch is determined to be abnormal (e.g., the defect prediction probability exceeds the warning threshold), the algorithm uses this abnormal prediction result as a loss function and calculates the gradient of this loss with respect to the features of each input entity node in the knowledge graph. According to the chain rule, the magnitude of this gradient value directly quantifies the contribution of small changes in each node feature to the final quality anomaly. The algorithm sorts the gradient values ​​of all nodes and identifies several nodes with the largest absolute gradient values ​​as potential root cause nodes. Further, by searching the knowledge graph for connected paths linking these high-gradient nodes, where all nodes on the path have high gradient values, the transmission chain of the quality problem can be clearly displayed, for example, "a certain batch of raw materials from supplier A deviates from a specific chemical composition index -> the temperature parameter of reactor No. 2 fluctuates abnormally -> leading to the final product's crystallinity index failing to meet standards," thereby achieving accurate and interpretable automated root cause analysis.

[0013] Preferably, the dynamic optimization and adjustment strategy for process parameters generated by the closed-loop feedback and process optimization module is based on a deep reinforcement learning surrogate model. The goal of this surrogate model is to maximize production efficiency or minimize energy consumption while ensuring product quality. Its environment consists of the entire production system; the state is defined as a multi-dimensional vector containing the real-time state feature vector of the current production line (derived from the output of step S120) and the current quality traceability result (derived from the output of step S150); the action space is discrete or continuous, defined as the specific adjustment amount to one or more process parameters (such as temperature, pressure, and feed rate) identified as key influencing factors; the reward function is carefully designed, with positive rewards derived from significant improvements in adjusted product quality indicators and increased production efficiency, and negative rewards derived from quality deterioration, increased energy consumption, or production interruption. By training with advanced reinforcement learning algorithms such as Proximal Policy Optimization (PPO), the agent can learn through continuous online or offline trial and error, explore and master which parameter adjustment actions can obtain the long-term optimal reward in different quality problem scenarios, thereby generating a precise, intelligent and adaptive closed-loop control strategy.

[0014] Compared with the prior art, the present invention has the following beneficial effects: Achieving full-process coverage and data integration: This invention breaks down the data silos of traditional quality monitoring by constructing a data collection system that runs through design, materials, production, logistics and after-sales service. It can analyze quality problems from a global perspective and improve the breadth and depth of data coverage.

[0015] This invention enhances the depth and accuracy of correlation analysis by constructing a knowledge graph of material-process-quality correlations using graph neural network technology. Compared with traditional statistical analysis methods, it can more profoundly explore the complex nonlinear causal relationships between multidimensional influencing factors, significantly improving the accuracy and interpretability of correlation analysis of quality issues.

[0016] This invention achieves a shift from passive detection to proactive prediction: through a knowledge graph-based quality risk prediction model, it can provide accurate early warnings of potential quality defects, transforming quality management from post-event remediation to pre-event prevention, effectively reducing defect rates and production costs.

[0017] It achieves automated traceability and intelligent closed-loop control: When quality problems occur, this invention can automatically and quickly locate the root cause of the problem, and generate dynamic optimization strategies based on reinforcement learning. It realizes intelligent closed-loop management from problem discovery, root cause analysis to solution execution, which greatly shortens the problem response time and improves the robustness and adaptability of the production process. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0020] Example 1 Please refer to Figure 1 This embodiment uses the manufacturing process of a multilayer printed circuit board (PCB) for a high-end smartphone motherboard as an example to elaborate in detail on the product end-to-end quality monitoring method and system based on big data proposed in this invention. This manufacturing process involves numerous materials, complex procedures, and narrow process parameter windows, which have a significant impact on the final product yield, making it a typical scenario for applying this invention. (Refer to...) Figure 1 The system architecture described in this invention is deployed and implemented in PCB manufacturing enterprises.

[0021] Step S110: The system performs full-process data acquisition and preprocessing. The full-process data acquisition and preprocessing module first constructs a data link covering the entire lifecycle of PCB design, material procurement, core manufacturing processes, warehousing and logistics, and after-sales repair. At the data acquisition level, industrial IoT sensors are deployed on key equipment on the production line to achieve real-time capture of underlying physical data. For example, temperature and pressure sensors are deployed on the laminating machine for inner layer circuit fabrication to acquire real-time temperature curves and pressure values ​​during the lamination process at a frequency of 10 Hz; light intensity sensors are deployed on the exposure machine for pattern transfer to monitor the stability of exposure energy; chemical analysis sensors are deployed on the electroplating line to monitor the concentration of copper ions, sulfuric acid, and additives in the copper plating solution in real time; and high-speed, high-precision automated optical inspection (AOI) equipment is deployed to acquire the quality status of intermediate products such as circuits and drill holes on each production board in image form. Simultaneously, the system deeply integrates with the enterprise's upper-level information systems through standardized data interface protocols, such as OPCUA or SECS / GEM. Specifically, the system interfaces with a Product Lifecycle Management (PLM) system to obtain design data such as original Gerber files, drilling files, laminated structure design parameters, and impedance requirements for PCB designs; it interfaces with an Enterprise Resource Planning (ERP) system to obtain supply chain data such as material purchase orders, supplier information, material batch numbers, and incoming inspection reports; it interfaces with a Manufacturing Execution System (MES) system to obtain production execution data such as production work orders, board transit records, equipment recipe call records, operator IDs, and equipment maintenance records; and it interfaces with a Customer Relationship Management (CRM) system to obtain after-sales quality data such as customer complaints and failure analysis reports.

[0022] The acquired multi-source heterogeneous data then enters the preprocessing process. The first step is data cleaning. For continuous time-series data acquired by sensors, a local anomaly factor algorithm is used. This algorithm identifies outliers by calculating the density difference between each data point and its neighbors, effectively removing invalid data caused by sensor jitter or electromagnetic interference. For example, for physically impossible instantaneous spikes in the temperature data of the laminating machine, the system automatically identifies and marks them as anomalies. For identified outliers and missing values ​​due to network packet loss, the system uses cubic spline interpolation to fill in the missing values. This method ensures that the filled data curve has second-order derivative continuity at the connection points, maximizing the reproduction of the actual process parameter variation trend. The second step is format conversion. The system processes the unstructured image data captured by the AOI equipment. Through a built-in image recognition algorithm, it automatically extracts information such as defect type (e.g., open circuit, short circuit, residual copper), defect coordinates, and defect size, converting it into structured data records and storing them in a relational database. For semi-structured design data such as Gerber files obtained from the PLM system, the system parses and extracts key design parameters, such as minimum line width, minimum spacing, and pad size, and associates them with production work orders. Finally, timestamp alignment is performed. Since the time bases for data generation may differ slightly between different systems, the system uses the transit timestamps from the MES system as a unified reference timeline. All collected sensor data, equipment status data, and quality inspection data are precisely time-aligned to ensure that for each PCB at any point in any process, all corresponding process parameters, material information, and environmental data have a unique and consistent time identifier. This forms a high-dimensional, structured, end-to-end quality dataset with the unique identifier (Panel ID) of each PCB as the primary key and time as the index, laying a solid data foundation for subsequent in-depth time-series analysis and correlation mining.

[0023] In step S120, the system performs process temporal feature extraction based on the preprocessed full-process quality dataset. The process temporal feature extraction module calls a hybrid parallel structure deep temporal network model combining Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRUs) to extract deep temporal features from key process parameters and equipment status data during production. Specifically, taking the PCB lamination process as an example, the input temporal data is a multivariate time series containing multiple process parameters such as temperature and pressure collected every second within a lamination cycle (e.g., 180 minutes). This data sequence is fed in parallel into the model's LSTM and GRU channels. The LSTM channel contains three precise control units: an input gate, a forget gate, and an output gate. It can learn and capture long-term, slowly changing dependencies during the lamination process. For example, during the heating and holding stage, the curing degree of the resin material is a slow, cumulative process, closely related to long-term factors such as the integral effect of the entire heating curve and the length of the holding time. The LSTM, through its memory units, can effectively capture this long-term cumulative effect. The GRU channel has a simpler structure, containing only update and reset gates. It focuses on capturing short-term fluctuations in the production process with higher computational efficiency. For example, during the pressurization stage, the instantaneous characteristics of the pressure curve's rise rate and the presence or absence of overshoot during the rapid transition from initial to high pressure are crucial for the removal of air bubbles within the sheet material. The GRU can accurately capture these short-term dynamic information. Each of the two channels outputs a feature vector: the LSTM output vector encodes long-term trend information, while the GRU output vector encodes short-term dynamic information. Finally, these two feature vectors are merged into a higher-dimensional process feature vector through a concatenation operation. This vector not only contains the static information of the original time-series data but, more importantly, encapsulates the dynamic changes and time dependencies of the production process, providing a more comprehensive and profound characterization of the current lamination process than traditional statistical features (such as mean and variance).

[0024] In step S130, the system constructs a knowledge graph relating materials, processes, and quality. The knowledge graph construction module combines the high-dimensional process feature vector generated in step S120 with product design data, material batch data, and quality inspection result data from PLM, ERP, and MES systems, utilizing Graph Attention Network (GAT) technology to construct a dynamic, weighted knowledge graph. First, the graph's schéma is defined, defining the core elements of the entire production process as different types of entity nodes. For example, node types include: material batches (e.g., FR-4 substrates, copper foil, PP sheets from a supplier), equipment units (e.g., press No. 3, drilling rig No. 5), process parameters (e.g., lamination temperature, drilling speed), operators, environmental indicators (e.g., workshop temperature and humidity), intermediate products (e.g., inner core board), and final quality inspection items (e.g., characteristic impedance, withstand voltage test). Subsequently, the relationship edges between entity nodes are defined based on the physical process, process logic and prior knowledge. For example, (material batch A) - [:used as] -> (intermediate product B), (intermediate product B) - [:processed by equipment C] -> (intermediate product D), (equipment C) - [:set process parameters E] -> (processing), (processing) - [:influence] -> (final quality inspection item F).

[0025] When constructing the edge weights, i.e., the association strength, a graph attention network is used for learning. The core of GAT lies in its self-attention mechanism. For each node in the graph, when aggregating information from its neighbors, it doesn't simply take an average; instead, it dynamically calculates an attention weight for each neighbor. This weight represents the importance of the information contained in that neighbor in the current information propagation and node update task. The attention coefficient is calculated as follows:

[0026] in, Represents a node Information for nodes The importance of and It is the feature vector of the node. It is a shared linear transformation weight matrix. It is the weight vector of a single-layer feedforward neural network. It is a node The set of all neighboring nodes, , The node feature vector after linear transformation This means concatenating two vectors to form a longer vector, which is then used as input to the attention mechanism. Calculate a linear combination of attention scores; the result is a scalar. The activation function introduces non-linearity. LeakyReLU can avoid the "neuron death" problem. An exponential function is used to construct the numerator of the softmax function, ensuring that the attention weights are positive. For nodes All neighbors Normalization is performed to ensure that the sum of all attention weights is 1, and the superscript is set to 1. (like () represents the transpose of a vector or matrix. This represents a dummy variable used in summation. Through this mechanism, the model can automatically learn the strength of the correlation between different factors based on the data. For example, when analyzing characteristic impedance non-compliance issues, GAT might assign a higher attention weight to the neighbor node "etch uniformity of inner layer circuitry" than "workshop humidity," thus forming a stronger correlation edge in the graph. This allows the constructed knowledge graph to not only represent "who is related to whom" but also quantify the "tightness of the correlation," forming a more accurate and physically interpretable quality correlation network.

[0027] In step S140, the system performs intelligent prediction and early warning of quality risks based on the constructed knowledge graph. The intelligent prediction and early warning module for quality risks adopts a graph-level classification model based on Graph Convolutional Network (GCN). When a new production work order is issued, the system extracts a subgraph composed of all nodes and edges related to the work order from the complete knowledge graph based on information such as the material batches, equipment, and process formulas to be used in the work order. The node feature matrix (containing material attributes, high-dimensional process feature vectors, etc.) and adjacency matrix of this subgraph are used as inputs to the GCN model. The model realizes information propagation and aggregation between nodes through multi-layer graph convolution operations. In the first layer, each node gathers information from its direct neighbors to update its own feature representation; in the second layer, the information propagation range is expanded to second-order neighbors. After, for example, three layers of graph convolution, the final feature representation of each node contains structural information within its three-hop neighborhood in the subgraph. Subsequently, a graph pooling operation, such as average pooling, is performed on the final feature vectors of all nodes in the subgraph to average the vectors of all nodes, resulting in a fixed-dimensional graph-level representation vector that represents the overall state of the production order. This vector is fed into a classifier consisting of two fully connected layers and a Softmax activation function, ultimately outputting the probability value of a specific quality defect (such as delamination, open circuit, or impedance deviation) occurring in this batch of PCBs. The system internally sets multiple warning thresholds. For example, when the predicted probability of delamination is below 0.1, the risk level is "normal"; between 0.1 and 0.3, it is "attention," and the system will alert production line engineers; between 0.3 and 0.6, it is "warning," and the system will automatically mark the batch as a key monitoring target; when the probability exceeds 0.6, it is "serious," and the system will immediately issue a stop-line command and prevent the batch of products from flowing into the next process, thus realizing a shift from passive detection to proactive prevention.

[0028] In step S150, when a quality anomaly is detected or a severe warning signal is received, the system automatically performs root cause tracing. The automated root cause tracing module employs a gradient-based graph tracing algorithm. Assuming that in step S140, the model predicts a layering defect probability of 0.7 for a batch of PCBs, triggering a severe warning, the algorithm defines this anomaly prediction (i.e., a layering defect probability of 0.7) as the value of the loss function. Then, using backpropagation, it calculates the gradient of this loss function with respect to the initial features of each entity node in the input subgraph. According to the chain rule in calculus, the absolute value of this gradient directly quantifies the contribution of small changes in the features of each initial node to the final high layering probability. The algorithm traverses all nodes, calculating and storing the feature gradient of each node. Subsequently, the absolute values ​​of the gradients of all nodes are sorted, and the top 5 nodes with the highest gradient values ​​are identified as the most likely key influencing factor nodes. For example, the tracing results might highlight the "resin content" feature of the "PP sheet batch_P12345" node and the "vacuum degree" process feature of the "press 3" node. Furthermore, the algorithm searches the graph for paths connecting these high-gradient nodes, requiring that all intermediate nodes along the path also have relatively high gradient values. Ultimately, the system outputs a clear and visual attribution path, such as: "Supplier X's PP sheet batch_P12345 has a resin content below the standard lower limit -> resulting in insufficient resin flowability during lamination on press 3 -> causing incomplete filling in localized areas within the sheet -> ultimately leading to delamination defects in the finished product." This approach achieves accurate, rapid, and highly interpretable automated root cause tracing.

[0029] In step S160, the system performs closed-loop feedback and process optimization based on the traceability results. The closed-loop feedback and process optimization module incorporates a process parameter optimization surrogate model based on deep reinforcement learning. This surrogate model aims at both optimizing production quality and minimizing production costs. The surrogate model is activated upon receiving the traceability results output in step S150. Its operating environment consists of the MES system and the underlying equipment control system. The state is defined as a vector containing the current production line state (e.g., equipment load, work-in-process quantity), the real-time high-dimensional process feature vector obtained from step S120, and the key influencing factor node information output in step S150. The action space is defined as the specific values ​​to be adjusted for the key process parameters located in the traceability (in this example, the lamination temperature, pressure, and holding time of press No. 3). For example, the action could be "increase the lamination temperature by 3 degrees Celsius and increase the pressure by 0.1 MPa". The reward function is carefully designed, and its mathematical expression is:

[0030] in, In time step The reward value obtained. This refers to the increase in the yield rate of PCB batches produced after the adjustment, which is a positive incentive. It represents the change in production rhythm. If the adjustment leads to an extension of production time, it is a negative reward. It is the change in energy or material consumption, and it is also a negative reward. , , These are weighting coefficients used to balance the relationship between quality, efficiency, and cost. The system employs the Proximal Policy Optimization (PPO) algorithm to train the agent. Through extensive trial-and-error learning in a simulation environment or in controlled real production, the agent gradually learns which process parameter adjustment strategies should be adopted to obtain the maximum long-term cumulative reward when facing quality problems caused by fluctuations in different material batches and equipment state drift. After learning, when a similar problem occurs again, the agent can automatically generate an optimal dynamic optimization adjustment strategy for process parameters and issue this strategy to the MES system in the form of an instruction. The MES system then updates the production program of Press No. 3, thereby achieving rapid, intelligent, closed-loop feedback control and continuous process improvement for quality problems.

[0031] Example 2 This embodiment uses the continuous polymerization process of polypropylene (PP) resin for mass production as an example to illustrate the specific application of the technical solution of the present invention in the process industry. Unlike the discrete manufacturing in Embodiment 1, continuous chemical production is characterized by highly coupled process variables and complex dynamic responses. The core of quality control lies in maintaining the stability of key quality indicators (such as melt flow index, MFI).

[0032] Step S110: The full-process data acquisition and preprocessing module integrates data for the polymerization process. At the data acquisition level, the distributed control system (DCS) frequently collects continuous process variables within the polymerization reactor, such as temperature, pressure, stirring rate, catalyst feed rate, and monomer concentration. Online analytical instruments (e.g., near-infrared spectrometers) are used to acquire the physical properties of intermediate products in real time. Simultaneously, the system connects to the ERP system to obtain supplier, batch, and purity inspection data for upstream raw materials (ethylene, propylene monomers, and catalysts). It also connects to the Laboratory Information Management System (LIMS) to obtain offline test results for the final product, PP oliwa, including melt index, density, and tensile strength. In the data preprocessing stage, a moving average filtering algorithm is used to smooth out high-frequency noise from the high-frequency data acquired by the DCS. For the low-frequency offline quality data provided by the LIMS, a combination of interpolation and forward padding is used to align it with the high-frequency process data on the time axis, generating a full-process dataset with a unified time granularity.

[0033] Step S120, the process time-series feature extraction module also employs a deep time-series network model. However, considering the more significant inertia and long-term drift characteristics of process variables in continuous production, the model structure in this embodiment focuses on capturing dependencies over longer time scales. The model uses a stacked LSTM network structure, expanding the receptive field by increasing network depth, thereby learning, for example, the long-term impact of the slow decay of catalyst activity over time on the reaction process, or the lag effect of small fluctuations in upstream raw materials that only manifest in the final product quality several hours later. The high-dimensional process feature vector output by the model can characterize the macroscopic operating state of the entire polymerization reaction system in real time.

[0034] Step S130: The material-process-quality correlation knowledge graph construction module adopts a simplified technical route in this embodiment. Given that the mechanisms of chemical processes are relatively clear, and there is a large amount of prior expert knowledge regarding the causal relationships between variables, the initial topology of the knowledge graph is mainly constructed artificially based on the principles of chemical reaction engineering. For example, mechanism-based relationship edges are established such as (catalyst feed rate) - [:influence] -> (reaction rate), (reaction rate) - [:determines] -> (polymer molecular weight), (polymer molecular weight) - [:negatively correlated with] -> (melt index MFI). Based on this, this embodiment uses a standard graph convolutional network (GCN) instead of a graph attention network (GAT) to learn node representations and edge weights. Since the main correlation paths have already been determined by expert knowledge, using GCN for information aggregation can effectively verify and quantify the strength of these known correlations with lower computational complexity, making it suitable for process industry scenarios with relatively clear mechanisms and achieving cost-effectiveness optimization of the technical solution.

[0035] In step S140, the task of the intelligent prediction and early warning module for quality risks in this scenario is to establish a regression prediction model for quality indicators. The model takes the current process subgraph as input, and through GCN and graph pooling operations, outputs a specific predicted value for the melt flow index (MFI) of the product one hour from now. The early warning mechanism is based on the degree of deviation of the predicted value from the target control range of MFI (e.g., 2.0 ± 0.2). When the predicted value deviates from the target center by more than 0.1, the system issues a "pay attention" signal; when the deviation exceeds 0.15, a "warning" signal is issued, prompting the operator to intervene; when the predicted value is about to exceed the upper or lower control limits, a "critical" early warning is triggered.

[0036] In step S150, the automated root cause tracing module also employs a gradient attribution-based graphical tracing method. When the MFI prediction value deviates significantly, for example, if the prediction value is 2.3, the algorithm reverse-calcules the contribution of this prediction deviation to all upstream process parameters and material properties. The tracing results may pinpoint "the activity index of catalyst batch A" and "the cooling water temperature of the reactor jacket" as the two main influencing factors leading to the increase in MFI.

[0037] In step S160, the closed-loop feedback and process optimization module in this embodiment employs a Model Predictive Control (MPC) strategy, a significant technical difference from the reinforcement learning approach used in Embodiment 1. MPC is an advanced process control technology, particularly suitable for chemical processes with large time delays and multivariate coupling characteristics. Upon receiving the traceability results, the MPC controller is activated. It utilizes a built-in mechanistic model or data-driven model capable of describing the dynamic response of the polymerization process, using "low catalyst activity" and "high cooling water temperature" as disturbance inputs. The controller's goal is to optimize the sequence of actions of manipulated variables (catalyst feed rate and cooling water valve opening in this example) over a future period (e.g., 2 hours) so that the predicted trajectory of the MFI can quickly and smoothly return to the target value of 2.0, while minimizing the amplitude and energy consumption of the control actions. MPC solves a constrained optimization problem, calculates the optimal control action at the current moment, and sends it to the DCS system for execution. In the next control cycle, the controller performs rolling optimization again based on the latest measurements. Compared to model-free exploration using reinforcement learning, this model-based optimization control method offers faster response and more stable control in scenarios with a clear process model. It represents another efficient and mature technological path for achieving intelligent closed-loop control in the process industry.

[0038] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A product end-to-end quality monitoring system based on big data, characterized in that, Includes the following parts: Full-process data acquisition and preprocessing module: used to build a full-process data acquisition system covering product design, materials, production, logistics and after-sales, to acquire and integrate multi-source heterogeneous data in real time, and to perform data cleaning, format conversion and timestamp alignment preprocessing operations on the multi-source heterogeneous data; Process time series feature extraction module: Based on the multi-source heterogeneous data, it uses a deep time series network model to extract time series features from key process parameters and equipment status data in the production process, and generates a high-dimensional process feature vector that represents the real-time status of the production process. Material-Process-Quality Association Knowledge Graph Construction Module: This module combines the high-dimensional process feature vectors with product design data, material batch data, and quality inspection result data, and constructs the knowledge graph using graph neural network technology. Intelligent prediction and early warning module for quality risks: used to establish a quality risk prediction model based on the material-process-quality correlation knowledge graph; Automated Root Cause Analysis Module: When a quality anomaly is detected or an early warning signal is received, it executes a gradient attribution-based backpropagation tracing algorithm on the knowledge graph to automatically locate the key influencing factor nodes and core associated paths that lead to the quality problem. Closed-loop feedback and process optimization module: Based on the key influencing factor nodes and core correlation paths, it generates targeted dynamic optimization and adjustment strategies for process parameters through a deep reinforcement learning agent model, and distributes the dynamic optimization and adjustment strategies for process parameters to the production execution system to achieve closed-loop feedback control and continuous improvement of quality issues.

2. The product end-to-end quality monitoring system based on big data according to claim 1, characterized in that, In the process time-series feature extraction module, the deep time-series network model is a hybrid parallel structure of a long short-term memory network and a gated recurrent unit. This structure feeds the input time-series data into the long short-term memory network channel and the gated recurrent unit channel in parallel, respectively learning and capturing long-term dependencies and short-term fluctuation features in the production process. The feature vectors output by the two channels are then concatenated or weighted and fused. The long short-term memory network channel learns and captures long-term dependencies in the production process through its internal input gate, forget gate, and output gate structure; the gated recurrent unit channel captures short-term fluctuation features in the production process through its update gate and reset gate.

3. The product end-to-end quality monitoring system based on big data according to claim 1, characterized in that, The material-process-quality association knowledge graph explicitly represents the intrinsic relationships and transmission paths between various quality-influencing factors in the entire product process in the form of entity nodes and relational edges. The graph neural network technology used in the material-process-quality association knowledge graph construction module is a graph attention network. This network calculates an attention weight for each neighbor node through a self-attention mechanism to dynamically learn and identify the association strength of different influencing factors.

4. The product end-to-end quality monitoring system based on big data according to claim 1, characterized in that, The quality risk prediction model uses global information from the aggregated graph to predict potential quality defects in future product batches in real time, and provides risk classification and early warning based on the prediction results and multiple preset thresholds. The quality risk intelligent prediction and early warning module uses a graph-level classification or regression model based on graph convolutional networks. This model takes the node feature matrix and adjacency matrix of the knowledge graph as input, and uses multi-layer graph convolution operations to continuously aggregate the information of its neighboring nodes. The final representation vectors of all relevant nodes representing a specific product batch are then subjected to graph pooling to obtain the graph-level representation of the product batch. Finally, the quality prediction results of the batch are output through one or more fully connected layers.

5. The product end-to-end quality monitoring system based on big data according to claim 1, characterized in that, The closed-loop feedback and process optimization module uses a deep reinforcement learning agent model, whose state is defined as a multi-dimensional vector containing the real-time state feature vector of the current production line and the current quality traceability result; its action space is defined as the specific adjustment amount of one or more process parameters identified as key influencing factors; and its reward function is calculated based on the improvement amount of the adjusted product quality indicators, the change in production efficiency, and the change in energy consumption.

6. The product end-to-end quality monitoring system based on big data according to claim 5, characterized in that, The reward function is calculated as follows: in In time step The reward value obtained; It is the increase in the yield of the PCB batch produced after the adjustment; It is the change in production cycle time; It is the change in energy consumption or material consumption; , , It is the weighting coefficient.

7. A monitoring method applied to the big data-based product end-to-end quality monitoring system as described in any one of claims 1-6, characterized in that, Includes the following steps: A comprehensive data acquisition system covering product design, materials, production, logistics, and after-sales service is constructed. This system utilizes industrial IoT sensors deployed on the production line to interface with enterprise resource planning (ERP), manufacturing execution (MEP), product lifecycle management (PLM), and customer relationship management (CRM) systems. It acquires and integrates multi-source heterogeneous data in real time and performs preprocessing operations such as data cleaning, format conversion, and timestamp alignment. Data cleaning employs the 3-sigma principle or local anomaly factor algorithm to identify and handle outliers, and spline interpolation or mean imputation to handle missing values. Format conversion transforms unstructured or semi-structured data into structured data tables. Timestamp alignment precisely aligns all time-series data using a unified benchmark timeline, ultimately forming a time-series, structured, full-process quality dataset. Based on the full-process quality dataset, a deep temporal network model is used to extract time-series features from key process parameters and equipment status data in the production process. The deep temporal network model is a hybrid parallel structure of a long short-term memory network and a gated recurrent unit. This structure feeds the input time-series data into the long short-term memory network channel and the gated recurrent unit channel in parallel, respectively learning and capturing the long-term dependencies and short-term fluctuation features in the production process. The feature vectors output by the two channels are concatenated or weighted and fused to generate a high-dimensional process feature vector that can characterize the real-time status of the production process. By combining the high-dimensional process feature vector with product design data, material batch data, and quality inspection result data, a material-process-quality association knowledge graph is constructed using graph neural network technology. This material-process-quality association knowledge graph explicitly represents the intrinsic association and transmission path between various quality influencing factors in the entire product process in the form of entity nodes and relational edges. Based on the material-process-quality association knowledge graph, a quality risk prediction model is established. By aggregating global information in the graph, the potential quality defects of future product batches are predicted in real time, and risk classification and early warning are carried out based on the prediction results and multiple preset thresholds. When a quality anomaly is detected or an early warning signal is received, a gradient attribution-based backpropagation tracing algorithm is executed on the knowledge graph. By calculating the gradient of the anomaly prediction result with respect to the features of each input entity node in the knowledge graph, the contribution of each node feature to the quality anomaly is quantified, and the key influencing factor nodes and core association paths leading to the quality problem are automatically located. Based on the tracing results of the key influencing factor nodes and related paths, a targeted dynamic optimization and adjustment strategy for process parameters is generated through a deep reinforcement learning agent model. The strategy is then distributed to the production execution system to achieve closed-loop feedback control and continuous improvement of quality issues.

8. The product end-to-end quality monitoring method based on big data according to claim 7, characterized in that, In the step of constructing the knowledge graph of material-process-quality association, material batches, equipment units, process parameter settings, actual process parameter values, operators, environmental indicators, intermediate quality inspection items, and final quality inspection items are defined as entity nodes; physical connections, process flow relationships, and causal influence relationships are defined as relationship edges.

9. The product end-to-end quality monitoring method based on big data according to claim 7, characterized in that, After executing the backpropagation tracing algorithm, the method further includes searching the knowledge graph for connected paths that link the key influencing factor nodes, where all nodes on the path have high gradient values, to demonstrate the transmission chain of the quality problem.

10. The product end-to-end quality monitoring method based on big data according to claim 7, characterized in that, The quality risk prediction step also includes, when a new production work order is issued, extracting a subgraph composed of all nodes and edges related to the work order from the complete knowledge graph based on the work order's planning information, and using this subgraph as the input to the quality risk prediction model.

Images