Intelligent scheduling optimization method for high-precision multi-layer PCBA full-automatic production line

Abstract

The invention relates to the technical field of comprehensive factory control, in particular to a high-precision multi-layer PCBA full-automatic production line intelligent scheduling optimization method which comprises the steps that production line physical equipment data streams are obtained in real time, feature extraction and alignment are carried out, and a production state data set is generated through a data fusion model; then, operating the data driving model to predict key production indexes and implicit states, and establishing an operation mechanism model to simulate physical behaviors of equipment at the same time; constructing a digital twinborn model, deeply fusing a prediction result of the data driving model and a simulation result of the mechanism model by adopting a physical information neural network, and calculating and updating a digital twinborn state of the production line; and finally, generating an optimized production scheduling scheme by applying a heuristic optimization algorithm, and generating a real-time executable control instruction in combination with a predefined self-adaptive control rule so as to adjust process parameters. According to the invention, the scheduling efficiency of the PCBA production line and the adaptive capacity to the dynamic environment can be improved.

Description

Technical Field

[0001] The present invention relates to the field of industrial control technology, and in particular to an intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line. Background Art

[0002] Industrial Control Systems (ICS) are the cornerstone of modern industrial automation and are widely used in critical infrastructure sectors such as manufacturing, energy, transportation, and water treatment. By integrating sensors, actuators, controllers (such as programmable logic controllers (PLCs), and communication networks, and combining the functions of distributed control systems and supervisory control and data acquisition (SCADA) systems, ICS enable real-time monitoring, data collection, and precise, programmed control of industrial processes, ensuring efficient, safe, and stable production operations.

[0003] Amid the increasing convergence of Industrial Internet and Internet of Things technologies, traditional industrial control systems are increasingly facing limitations in addressing the complex scheduling requirements of fully automated, high-precision, multi-layer PCBA production lines. These lines integrate a series of sophisticated and interdependent processes, including surface-mount mounting (SMT) placement, reflow soldering, automated inspection (AOI) inspection, plug-in assembly, and wave soldering. Each link places stringent demands on control accuracy, production cycle time, and final quality. However, the PCBA production environment is rife with dynamic events and uncertainties, such as unexpected equipment failures, material supply delays, urgent order insertions, fluctuations in actual processing times, and yield variations. These disturbances often render pre-programmed, static production plans or control programs based on deterministic assumptions rapidly ineffective. Consequently, these industrial control systems often lack sufficient dynamic adaptability, making it difficult to quickly and effectively adjust their pre-set control logic and scheduling sequences to adapt to new circumstances. This delayed response and insufficient adaptability directly impacts production continuity and stability, as well as the ability to rapidly respond to dynamic market demands.

[0004] To this end, an intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line is proposed. Summary of the Invention

[0005] The purpose of the present invention is to provide a high-precision multi-layer PCBA fully automatic production line intelligent scheduling optimization method to solve the problems of poor adaptability of existing PCBA production scheduling to dynamic disturbances and difficulty in collaborative optimization of multiple objectives. First, the data stream of the physical equipment of the production line is acquired in real time, and feature extraction and alignment are performed, and a production status data set is generated through a data fusion model; then, the data-driven model is run to predict key production indicators and implicit states, and an operation mechanism model is established to simulate the physical behavior of the equipment; a digital twin model is constructed, and the prediction results of the data-driven model and the simulation results of the mechanism model are deeply integrated using a physical information neural network to calculate and update the digital twin state of the production line; finally, a heuristic optimization algorithm is applied to generate an optimized production scheduling plan, and predefined adaptive control rules are combined to generate real-time executable control instructions to adjust process parameters. The present invention can improve the scheduling efficiency of the PCBA production line and its adaptability to dynamic environments.

[0006] To achieve the above object, the present invention provides the following technical solutions: An intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line, comprising: Obtain physical device data streams, including structured and unstructured data; Extract features from the physical device data stream to generate multi-source features; align features based on timestamps and work order identifiers, and fuse multi-source features of the same workpiece using a data fusion model to output a production status dataset; Training and running a data-driven model based on the production status dataset to predict comprehensive production status indicators and implicit production status of production line equipment; establishing and running a mechanism model to simulate the physical behavior of the production line equipment; and constructing a digital twin model to calculate and update the digital twin status of the production line by weighted fusion of the comprehensive production status indicators, the implicit production status, and the physical behavior; A heuristic optimization algorithm is used to generate a production scheduling plan based on the production status data set, the digital twin status and preset constraints; and a predefined adaptive control rule logic is applied to generate executable control instructions.

[0007] Furthermore, the sources of the structured data and the unstructured data include: The structured data includes equipment status and process parameter data from a programmable logic controller, physical parameter data from a sensor, and production plan information from a manufacturing execution system; the unstructured data includes quality inspection data from an automatic optical inspection device.

[0008] Furthermore, generating the multi-source features includes: For the unstructured data, a lightweight convolutional neural network is used to identify and quantify welding defect features; For the structured data, the signal processing method is applied to the time-series physical parameters in the physical parameter data to extract the frequency domain features and the time-domain statistical features; the principal component analysis method is applied to the process parameter data and the non-time-series physical parameters in the physical parameter data to extract the principal component features; the categorical data in the equipment status and production plan information is encoded to generate categorical features; the welding defect features are combined with the frequency domain features, the time domain statistical features, the principal component features and the categorical features to form the multi-source features.

[0009] Furthermore, the process of outputting the production status dataset includes: The multi-source features are standardized, and the timestamp and the work order identifier are used as association keys to match and align the multi-source features of the same workpiece within a predetermined time window through a time window matching algorithm; Using a weighted fusion algorithm to perform weighted combination on the aligned multi-source features to generate a comprehensive production status indicator; using a probabilistic graphical model to infer the implicit production status based on the aligned multi-source features; The comprehensive production status indicator and the implicit production status are included in the output production status data set as a result.

[0010] Furthermore, establishing and running the mechanism model specifically includes: For the selected production line equipment, establishing an ordinary differential equation model based on simplified physical principles; Acquiring real-time parameters reflecting the current operating state of the equipment from the production status data set, and using the real-time parameters as initial conditions of the ordinary differential equation model; Loading preset physical parameters related to the production line equipment; using a numerical calculation library to numerically solve the ordinary differential equation model, simulate and output the physical behavior.

[0011] Furthermore, by weighted fusion of the comprehensive production status indicator, the implicit production status and the physical behavior, specifically including: Performing a preliminary weighted combination on the comprehensive production status indicator and the implicit production status to generate a first fusion indicator; Extracting features of the physical behavior and converting them into a second fusion index; The first fusion index, the second fusion index, and the third fusion index selected from the multi-source features are again weightedly combined to generate and update the digital twin state.

[0012] Furthermore, using the heuristic optimization algorithm to generate a production scheduling plan specifically includes: Using genetic algorithm as the heuristic optimization algorithm; Encode the production scheduling problem into individuals, where individuals represent candidate production scheduling solutions; Constructing a fitness function for evaluating the candidate production scheduling solutions; calculating the fitness function by weighted combination of scheduling objectives based on the multi-source features and the digital twin state, wherein the scheduling objectives include work order delivery delays, predicted production yield loss, and equipment energy consumption; Under the premise of satisfying preset constraints, the selection, crossover and mutation operations are iteratively performed to optimize the search and determine the optimal individual with the optimal fitness value from a population of N individuals, and the optimal individual is decoded into the production scheduling plan.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a comprehensive and detailed production status dataset through systematic real-time acquisition of multi-source heterogeneous data (structured and unstructured), deep feature engineering, and a primary data fusion model. This effectively overcomes the incomplete and low-quality information often associated with traditional single data sources or simple data aggregation. Through targeted feature extraction of raw data, precise alignment of time and work order dimensions, and preliminary fusion based on weighted algorithms and probabilistic graphical models, the quality and relevance of input data are significantly improved, generating rich intermediate-level information encompassing comprehensive production status indicators and implicit production status. This also lays a solid, reliable, and comprehensive data foundation for the subsequent construction of high-fidelity digital twins and driving intelligent decision-making.

[0014] 2. The present invention constructs and applies a digital twin state that includes a data-driven model, a simplified mechanism model, and is calculated and updated in real time through a specific fusion mechanism. This digital twin state is not a simple reproduction of the physical production line, but an organic combination of accurate prediction of complex production indicators and implicit states, and simulation of core physical behaviors. In particular, when segmented dynamic weighted fusion is adopted, it can ensure that the twin state not only fits the data rules, but also follows key physical constraints or dynamically adjusts the information focus according to the real-time context. This digital twin state that integrates insights from multiple parties provides the heuristic scheduling algorithm with unprecedented in-depth knowledge of current and future trends of the production line.

[0015] 3. Based on the production status data set and digital twin status, the present invention realizes the deep linkage and closed-loop operation of heuristic scheduling optimization and adaptive control rule logic. The heuristic optimization algorithm uses these rich real-time and predictive inputs to perform dynamic and accurate quantitative evaluation of its multi-objective fitness function, thereby generating a highly optimized production scheduling plan that can adapt to the current real working conditions. At the same time, the adaptive control rule logic generates executable process adjustment instructions in real time based on these same inputs. This close combination of scheduling and control and closed-loop feedback ultimately achieves comprehensive benefits in improving production efficiency, ensuring product quality, enhancing production line resilience and reducing operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention provides a flow chart of an intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line; Figure 2 Provides a schematic diagram of the structure of multi-source feature generation for the present invention; Figure 3 Schematic diagram of the structure of the weighted fusion process of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] See also Figures 1 to 3 The present invention provides an intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line. The technical solution is as follows: Example 1: In order to more specifically explain the technical details and application process of the intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line disclosed in the present invention, in an embodiment of the present invention, the automated production process of the core processing board of an automobile engine control unit is taken as an example to illustrate how this method can be applied to actual production scenarios to improve the continuity and stability of production and the ability to quickly respond to dynamic market demands.

[0019] The core processing board is produced on a typical, high-precision, multi-layer PCBA fully automated production line. This line, along with the process flow, includes key processes such as solder paste printing (SPM), solder paste inspection (SPI), high-speed surface mount technology (SMT), multi-temperature reflow soldering (reflow), automated optical inspection (AOI), X-ray inspection (AXI), and in-circuit functional testing and program programming (FCT). Each key process device within the line is equipped with standard industrial communication interfaces (such as support for the OPC UA protocol), providing a foundation for real-time data collection and remote command issuance.

[0020] like Figure 1 As shown, a method for intelligent scheduling optimization of a high-precision multi-layer PCBA fully automatic production line includes: Capture physical device data streams, including structured and unstructured data.

[0021] Furthermore, the sources of the structured data and the unstructured data include: The structured data includes: In this embodiment, device status and process parameter data from programmable logic controllers (PLCs) is collected by a data acquisition software module deployed as an OPC UA client, connected to the PLC controllers of key equipment (such as placement machines and reflow ovens) via Industrial Ethernet. This software subscribes to predefined key data tags within the PLC in real time or at a configurable frequency (e.g., hundreds of milliseconds for high-frequency data like device status, and seconds for lower-frequency data like process parameters) to obtain device status and process parameter data, including machine operating mode codes, fault codes, actual temperature zones, and conveyor speeds. All collected data is accurately timestamped and associated with the currently processed work order ID and PCBA serial number.

[0022] Physical parameter data from sensors includes: Various industrial-grade sensors, such as vibration sensors, high-precision temperature sensors, ambient temperature and humidity sensors, and smart meters or power sensors for monitoring equipment energy consumption, have been installed in key equipment locations (such as SMT placement heads) and in the workshop environment (e.g., near reflow ovens and in specific areas of the workshop). These sensors are connected to an edge data acquisition gateway via a standard industrial bus (such as Modbus RTU / TCP) or digital / analog signal interface. The gateway reads the raw sensor data and performs preliminary processing (such as unit conversion and filtering). It then publishes the data, timestamped and source-identified, to the factory's internal messaging middleware using a lightweight messaging protocol (such as MQTT). The data acquisition software module, acting as an MQTT client, subscribes to the corresponding content to obtain real-time physical parameters. The data update frequency is dynamically set based on the specific sensors and monitoring requirements.

[0023] And production plan information from the manufacturing execution system, including: secure interaction with the manufacturing execution system (MES) through standard network service interfaces (such as RESTful APIs), triggered by production rhythm or specific events (such as work order start / end), or querying the MES for detailed information on current and subsequent work orders at preset intervals (such as every minute), including production instructions and plan data such as work order number, product model, planned quantity, priority code, planned start and end time, etc. The obtained structured data is parsed and used to enrich and associate production status data.

[0024] The unstructured data includes quality inspection data from automated optical inspection equipment, including: after the AOI equipment completes the inspection, in addition to generating a structured report containing defect details, it also saves an optical image of the defective area determined to contain defects according to the configuration. The image file name contains traceable information (such as PCBA serial number, inspection time, and defect coordinates, etc.) and is stored in a designated network file server or image database. While processing the structured inspection report, the data acquisition service records the storage location information of the associated image file.

[0025] By integrating data from PLCs, sensors, MES, and AOI, a comprehensive and multi-dimensional data foundation is built, providing more complete information for subsequent analysis and decision-making, avoiding the one-sidedness brought by a single data source.

[0026] Furthermore, feature extraction is performed on the physical device data stream to generate multi-source features; feature alignment is performed based on the timestamp and the work order identifier, and the multi-source features of the same workpiece are fused using a data fusion model to output a production status data set.

[0027] Among them, the data fusion model is a core computing logic set used to integrate aligned multi-source feature information to generate a production status dataset.

[0028] Furthermore, if Figure 2 As shown, using the data fusion model to generate the multi-source features includes: A lightweight convolutional neural network (CNN) is used to identify and quantify solder defect signatures in unstructured optical images of defective areas. During the acquisition process, a CNN model, pre-trained and fine-tuned on PCBA defect images, is used for processing. First, the input optical image of the defective area is preprocessed, resized to 224×224 pixels for size normalization, and pixel values ​​are normalized. After preprocessing, the image is fed into the CNN model. The model outputs a probability value for each predefined defect category, such as cold solder joint, solder bridging, tombstoning, insufficient solder, and component shift. These defect categories and their corresponding probabilities together constitute the solder defect signature.

[0029] For the structured data: Signal processing methods are applied to extract frequency-domain and time-domain statistical features from the time-series physical parameter data. For example, a time series segment of the raw vibration signal collected from the patch head vibration sensor, synchronized with the PCBA assembly process, is used to extract time-domain statistical features. NumPy and the SciPy.stats library are used to calculate key statistics such as the mean, standard deviation, root mean square (RMS), peak-to-peak value, kurtosis, skewness, and crest factor of the vibration signal within a specific time window. For frequency-domain feature extraction, SciPy.fft.fft is used to perform a fast Fourier transform (FFT) to obtain spectral information. The main frequency components, such as the three highest-energy frequency points and their amplitudes, are then identified and extracted from the spectrum. Spectral kurtosis, spectral entropy, and the energy proportion of each preset frequency band are also calculated. Ultimately, this series of operations generates a set of numerical features, laying the foundation for subsequent analysis.

[0030] Principal component analysis (PCA) is applied to process parameter data and non-time-series physical parameters within the physical parameter data to extract principal component features. Process parameter data such as temperature in each temperature zone and conveyor speed from the PLC, as well as historical data from sensors for non-time-series physical parameters such as average ambient temperature, average humidity, and average equipment energy consumption, are collected and integrated into a single dataset. After standardizing this dataset, PCA is applied to select the top k principal components that retain 95% of the variance and save them to form a PCA model. In real-time, the currently acquired set of parameter values ​​undergoes the same standardization process and is projected into the principal component space using the saved PCA model. Ultimately, k linearly independent numerical principal component features are output, achieving data dimensionality reduction.

[0031] Categorical data in equipment status and production plan information is encoded to generate categorical features. Equipment status includes status codes provided by the PLC, such as "Running," "Idle," and "Fault Code A." Categorical data in production plan information includes order priorities transmitted by the MES, such as "High," "Medium," and "Low." One-hot encoding is applied to this categorical data, converting it into a set of binary vectors. For example, an order priority of "High" would correspond to the code [0, 0, 1].

[0032] Finally, the welding defect feature is combined with the frequency domain feature, the time domain statistical feature, the principal component feature, and the category feature to form the multi-source feature that includes features from different data sources and processing methods.

[0033] By applying appropriate feature extraction methods to different modal information such as images, sensor signals, equipment status and production plans, and converting them into numerical or uniformly coded category features, multimodal data fusion is achieved, laying the foundation for a comprehensive and accurate understanding of the production status.

[0034] Furthermore, the process of outputting the production status dataset includes: First, the multi-source features are standardized, and the timestamp and the work order identifier are used as association keys to match and align the multi-source features of the same workpiece within a predetermined time window through a time window matching algorithm; The time window matching algorithm uses aggregating feature records that are close in time (for example, within a preset short time threshold or belonging to the same logical processing unit, such as reflow soldering) and have the same work order and serial number. Features generated multiple times within the time window (such as high-frequency sensor features) can be aggregated (e.g., taking the average, maximum, or most recent value) or retained in sequence.

[0035] Then, a weighted fusion algorithm is used to weightedly combine the aligned multi-source features to generate the comprehensive production status indicator. This process involves selecting a subset of relevant features for each indicator and assigning them preset weights based on expert experience. This generates comprehensive production status indicators such as a comprehensive health score for key equipment and a quality risk index for the current process. For example, the "SMT placement quality risk index" is used to assess the overall placement quality of current PCBAs in the SMT process. The selected feature subset primarily derives from the aligned and standardized multi-source features closely related to the SMT process extracted in the previous step. These features may include: statistical features reflecting solder paste printing quality (e.g., solder paste volume consistency metrics derived from analyzing structured data); quantitative features reflecting the placement accuracy of key components (e.g., average X, Y, and rotational offsets of high-precision components derived from analyzing placement machine feedback data or AOI pre-inspection image features); vibration features reflecting placement machine operational stability (e.g., the time-domain RMS value or specific frequency band energy of the placement head vibration signal); and categorical features reflecting the placement machine's current status (e.g., a coded value indicating whether the equipment is operating without faults). After normalizing the above-selected eigenvalues, they are assigned corresponding weights based on expert experience and weighted summation is performed.

[0036] Simultaneously, a probabilistic graphical model (PGM) is used to infer the implicit production state based on the aligned multi-source features. One or more probabilistic graphical models (PGMs), such as Bayesian networks, are constructed to infer implicit production states that are difficult to measure directly. The PGM's network structure and conditional probability table are trained and learned offline based on domain knowledge and extensive historical data. The multi-source features are input into the trained PGM as observational evidence, and the PGM's inference algorithm is used to calculate the posterior probability distribution of the target implicit state node. For example, the "current microcrack risk level within a weld" or the "probability range of unplanned equipment downtime in the next hour" can be inferred. The output implicit production state is a probabilistic description of the implicit state, for example, "microcrack risk": {"high": 0.1, "medium": 0.3, "low": 0.6}, "downtime probability (1 hour)": 0.05.

[0037] Finally, the comprehensive production status indicator and the implicit production status are included as the output production status dataset. Each record is associated with a unique artifact identifier (timestamp and work order number). This dataset can be published as a JSON object stream to a message queue (such as Kafka) for subsequent intelligent production line scheduling.

[0038] Aligned multi-source features from multiple sensors, equipment, and processes are converted into a few comprehensive indicators with clear business meaning through weighted combination. This conversion process enables production managers and subsequent intelligent algorithms to more easily understand and utilize the precise status of the current workpiece. Compared with scattered original features or simple pass / fail judgments, these indicators can more accurately reflect the quality of the status.

[0039] Based on the production status data set, a data-driven model is trained and run to predict the comprehensive production status indicators and implicit production status of the production line equipment; a mechanism model is established and run to simulate the physical behavior of the production line equipment; and a digital twin model is constructed to fuse the comprehensive production status indicators, the implicit production status and the physical behavior through a physical information neural network to calculate and update the digital twin status of the production line.

[0040] Furthermore, establishing and running the mechanism model specifically includes: For the selected production line equipment, establishing an ordinary differential equation model based on simplified physical principles; Acquiring real-time parameters reflecting the current operating state of the equipment from the production status data set, and using the real-time parameters as initial conditions of the ordinary differential equation model; Loading preset physical parameters related to the production line equipment; using a numerical calculation library to numerically solve the ordinary differential equation model, simulate and output the physical behavior.

[0041] Specifically, in this embodiment, a reflow oven is selected for modeling to simulate the temporal trend of the average temperature of the PCBA as it passes through a specific temperature zone of the reflow oven. It is assumed that heat transfer within this temperature zone is primarily composed of forced convection heating of the PCBA by the heating element and heat dissipation from the PCBA to the environment. Complex radiation heat transfer and temperature gradients at different locations within the PCBA are ignored, and the PCBA is considered as a lumped parameter system. Based on a simplified energy balance, the following ordinary differential equation is established to describe the average temperature of the PCBA: Rate of change: ; in, Indicates the average temperature of the PCBA The derivative with respect to time t, is the equivalent thermal capacity of the PCBA, and are the equivalent convection heat transfer coefficient and effective heat transfer area between the heated air and PCBA, The effective temperature of the heated air in the current temperature zone, and They are the equivalent heat dissipation coefficient and effective heat dissipation area between the PCBA and the non-heating environment in the temperature zone, is the reference temperature of the environment in the temperature zone, is the heat generation of the PCBA itself, which can be ignored and is zero.

[0042] When the target PCBA is about to enter or has just entered the modeled temperature zone, the latest data related to the PCBA and the temperature zone is queried from the production status dataset to obtain the initial temperature of the PCBA when entering the temperature zone. ,Will As the initial value for solving the ordinary differential equation, a time series is returned through numerical calculation, indicating the average temperature of the PCBA within the simulation time interval. For example, {"Timestamp": [0, 0.5, 1.0], "Simulated PCBA Average Temperature": [60.5, 68.2, 75.3]}. This result is the simulated physical behavior.

[0043] The mechanism model is built on physical principles, enabling it to clearly articulate the underlying physical reasons for analyzing equipment behavior, thereby enhancing the interpretability of the digital twin model. In practical applications, using real-time parameters from the production status dataset, such as current temperature and speed, as initial conditions or time-varying inputs for ordinary differential equations allows the mechanism model's simulation to be based on the actual current state of the physical production line. This results in short-term simulations that are closer to reality, providing a more accurate physical behavior benchmark for subsequent integration of physical information into neural networks.

[0044] Furthermore, if Figure 3 As shown, a schematic diagram of the weighted fusion process of the present invention is shown. By weighted fusion of the comprehensive production status indicator, the implicit production status and the physical behavior, specifically including: Performing a preliminary weighted combination on the comprehensive production status indicator and the implicit production status to generate a first fusion indicator; Extracting features of the physical behavior and converting them into a second fusion index; The first fusion index, the second fusion index, and the third fusion index selected from the multi-source features are again weightedly combined to generate and update the digital twin state.

[0045] Specifically, for example, a data-driven model (such as an LSTM model) can be used to predict comprehensive production status indicators, including the expected yield of the current batch and the current health score of key equipment A. This data-driven model also assesses the implicit production status and the risk level of microcracks in workpiece A. This information is then integrated to form the first fusion indicator.

[0046] The mechanism model, on the other hand, simulates the peak temperature and maximum thermal stress of BGA chips during the critical reflow soldering phase. For peak temperature, a process specification window is set, such as a target range of 240-245°C. The simulated peak temperature's conformance to or deviation from this specification window is then calculated. For maximum thermal stress, a safety threshold, such as 100 MPa, is set, and a safety margin is calculated. These converted values ​​are combined to form a second fusion indicator.

[0047] Furthermore, the real-time features most relevant to the current assessment objective are selected from multiple sources, such as the number of minor defects detected by upstream AOI and the normalized value of the current production line's real-time throughput. Ultimately, these features are weighted and combined to generate a comprehensive indicator reflecting the digital twin's state. The digital twin's state is a comprehensive, real-time, updated virtual information set that accurately describes the current operating status, inherent risks, and future trends of the physical production line, formed by integrating key indicators and implicit states predicted by data-driven models with physical behaviors simulated by mechanism models. This information forms the basis for subsequent intelligent scheduling and adaptive control decisions.

[0048] In addition, before executing the final weighted combination, the dynamic weights required for this round of fusion can be dynamically updated based on the latest contextual factors (such as the complexity level of the current processed product model) according to expert experience.

[0049] Based on real-time changes in production status datasets and contextual factors, the fusion weights of data-driven models, mechanism models, and key real-time multi-source features are dynamically adjusted to accurately map the actual production scenario under specific current working conditions. This not only comprehensively and accurately summarizes the real-time status and future trends of the production line, but also provides high-quality input for subsequent heuristic scheduling optimization algorithms and adaptive control rule logic, ensuring production continuity and stability, as well as the ability to quickly respond to dynamic market demands.

[0050] A heuristic optimization algorithm is used to generate a production scheduling plan based on the production status data set, the digital twin status and preset constraints; and a predefined adaptive control rule logic is applied to generate executable control instructions.

[0051] Furthermore, using the heuristic optimization algorithm to generate a production scheduling plan specifically includes: Using genetic algorithm as the heuristic optimization algorithm; The production scheduling problem is encoded as individuals, where individuals represent candidate production scheduling solutions; the production scheduling problem here is to find an optimal or near-optimal task-resource matching and sequencing solution.

[0052] Constructing a fitness function for evaluating the candidate production scheduling solutions; calculating the fitness function by weighted combination of scheduling objectives based on the multi-source features and the digital twin state, wherein the scheduling objectives include work order delivery delays, predicted production yield loss, and equipment energy consumption; Under the premise of satisfying preset constraints, the selection, crossover and mutation operations are iteratively performed to optimize the search and determine the optimal individual with the optimal fitness value from a population of N individuals, and the optimal individual is decoded into the production scheduling plan.

[0053] Specifically, an encoding method combining operation-based representation and machine allocation is used. The individual encoding methods include: The first part defines the global processing order of all operations in all work orders (for example, [Op1.1, Op2.1, Op1.2, Op3.1, Op2.2], where Opi.j represents the jth operation of work order i).

[0054] The second part specifies which specific machine each process is assigned to.

[0055] For each candidate scheduling solution represented by an individual, the calculation of its various scheduling objectives is as follows: Obtain the planned delivery date and priority of each work order j from multi-source features, calculate the delay time for each work order, and then assign corresponding weights based on the priority. The weighted sum is used to calculate the final work order delivery delay. Obtain the defect probability of processing work order j on equipment k assigned to each process Op in the scheduling plan from the digital twin state, directly deriving the yield loss for each process. This weighted sum is calculated based on the number of work orders to obtain the predicted production yield loss. Obtain the predicted energy consumption rate per unit time for each piece of equipment k in different operating states (e.g., processing, idle standby, and preheating) from the digital twin state. This predicted energy consumption rate is derived from a comprehensive assessment of the digital twin state based on the device model, current device health, and real-time load parameters from the multi-source features. Multiply the predicted energy consumption rate per unit time by the estimated duration of each piece of equipment k in each operating state to obtain the energy consumption of each piece of equipment. The energy consumption of each piece of equipment is then summed up to obtain the device energy consumption.

[0056] In addition, the preset constraints include: The first preset condition comes from the production status data set, including whether the equipment is currently available and whether the materials are complete.

[0057] The second preset condition comes from the digital twin status, whether the device health is higher than the safety threshold.

[0058] The fitness function is evaluated based on real-time production line data extracted from multiple sources, as well as high-level cognitive information deeply integrated into the digital twin state, including key indicators such as predicted processing time, yield rate, energy consumption rate, and equipment health. This allows scheduling decisions to be closely aligned with actual production conditions, accurately understand potential risks, and achieve dynamic optimization, thereby improving the flexibility and accuracy of the PCBA production process.

[0059] By deeply integrating real-time data perception, a production status dataset constructed based on multi-source features and data fusion models, and a digital twin state updated through weighted fusion calculations based on data-driven model predictions (comprehensive production status indicators and implicit production status) and mechanism model simulations (physical behavior), combined with heuristic scheduling algorithms and adaptive control rule logic, this method ultimately achieves fully automated production optimization. Based on a comprehensive and accurate assessment of the current and future status of the production line, this invention enables dynamic, accurate, and globally optimized production scheduling, effectively responding to various disturbances. Furthermore, by anticipating quality risks and adjusting process parameters in a closed-loop, real-time manner, it further improves the scheduling efficiency of the PCBA production line and its adaptability to dynamic environments.

[0060] Example 2: Based on Example 1, to further verify the effectiveness of the present invention, this example will further describe a method for intelligent scheduling optimization of a high-precision multi-layer PCBA fully automatic production line, including: Obtain physical device data streams, including structured and unstructured data; Extract features from the physical device data stream to generate multi-source features; align features based on timestamps and work order identifiers, and fuse multi-source features of the same workpiece using a data fusion model to output a production status dataset; Training and running a data-driven model based on the production status dataset to predict comprehensive production status indicators and implicit production status of production line equipment; establishing and running a mechanism model to simulate the physical behavior of the production line equipment; and constructing a digital twin model to calculate and update the digital twin status of the production line by weighted fusion of the comprehensive production status indicators, the implicit production status, and the physical behavior; A heuristic optimization algorithm is used to generate a production scheduling plan based on the production status data set, the digital twin status and preset constraints; and a predefined adaptive control rule logic is applied to generate executable control instructions.

[0061] The adaptive control rule logic continuously monitors real-time information from the production status dataset, including current process parameter measurements, preliminary quality inspection feedback, and insights and comprehensive assessments provided by the digital twin state, such as the predicted comprehensive quality risk index of the current workpiece, the health score of key equipment, and the deviation of simulated physical behavior from the ideal state. This information is then continuously compared against control objectives, process windows, or risk thresholds in the rule base. Once the monitored status meets the trigger conditions of a predefined rule, such as a key process parameter consistently deviating from the set center, a quality risk indicator exceeding the warning level, or an equipment health score indicating potential instability, the corresponding rule is activated. Based on pre-set adjustment strategies, the activated rule automatically calculates the specific corrections or new setpoints for key process parameters, such as reflow oven zone temperature, SMT placement machine speed, and solder paste printer squeegee pressure. The adjustment decisions are then converted into a standard command format that can be understood and executed by the target production equipment controller. These commands are then distributed to the corresponding equipment via the industrial communication network interface, enabling real-time, closed-loop, and automated fine-tuning of the production process. This process enables the production line to ensure stable quality and efficient operation based on accurate judgment of current and future status, and to respond quickly to subtle changes and potential problems in the PCBA production line.

[0062] To verify the effectiveness of the method described in this invention, we simulated the production line for the core processing board of an automotive engine control unit described in Example 1. This production line encompasses key processes and equipment, including SPM, SPI, SMT, reflow, AOI, AXI, and FCT. The simulation ran for four consecutive weeks and introduced the following dynamic and uncertain factors: Regarding order arrival, we simulated a work order flow encompassing multiple engine control unit models with varying batch sizes, priorities (standard, expedited), and delivery deadlines. Regarding equipment failure, we randomly introduced equipment failure events based on the mean time between failures (MTBF) and mean time to repair (MTTR) of key equipment (e.g., SMT placement machines and reflow ovens). Regarding process parameter drift, we simulated minor drifts in process parameters (e.g., oven temperature and placement accuracy) caused by natural equipment aging or environmental factors.

[0063] For comparison, the following baseline scheduling and control strategies representing different technical levels are set: Baseline Model 1: A simple "first-in, first-out" rule is used in scheduling to determine the order in which work orders are processed in each process. In terms of control, all equipment process parameters (such as the reflow temperature curve and placement machine speed) use standard, offline fixed values ​​and are not adjusted in real time.

[0064] Baseline Model 2: A genetic algorithm similar to that of the present invention is used for optimization during scheduling. However, the targets such as processing time, yield, and energy consumption in its fitness function are calculated based on static historical average data or fixed estimates, and no real-time "production status dataset" or "digital twin status" is used for dynamic evaluation. The control method is the same as that of Baseline Model 1, using fixed process parameters.

[0065] Baseline Model 3: Scheduling is optimized based on the real-time "production status dataset" and "digital twin status" as described in the present invention. For control, fixed process parameters are used, and the genetic algorithm and adaptive control logic described in the present invention are not enabled.

[0066] By running the above four strategies (including the method of the present invention) 10 times each in the set scenario, the average value was taken to obtain the comparison results shown in Table 1.

[0067] Table 1 Comparison results of main performance indicators Table 2 Comparison of robustness to dynamic disturbance responses As shown in Table 2, a performance comparison under dynamic and uncertain factors is presented. It can be seen that the present invention demonstrates optimal robustness in dealing with emergency orders and equipment failures, thereby minimizing additional delays and output losses, and improving the scheduling efficiency of the PCBA production line and its adaptability to dynamic environments.

[0068] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A high-precision multi-layer PCBA fully automatic production line intelligent scheduling optimization method, characterized in that: include: Obtain physical device data streams, including structured and unstructured data; Performing feature extraction on the physical device data stream to generate multi-source features; Perform feature alignment based on timestamps and work order identifiers, and use a data fusion model to fuse multi-source features of the same workpiece to output a production status dataset. Training and running a data-driven model based on the production status dataset to predict comprehensive production status indicators and implicit production status of production line equipment; Establish and run a mechanism model to simulate the physical behavior of the production line equipment; Constructing a digital twin model for calculating and updating the digital twin state of the production line by weighted fusion of the comprehensive production state indicator, the implicit production state, and the physical behavior; A heuristic optimization algorithm is used to generate a production scheduling plan based on the production status data set, the digital twin status and preset constraints; and a predefined adaptive control rule logic is applied to generate executable control instructions.

2. The intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line according to claim 1 is characterized in that: The sources of the structured data and the unstructured data include: The structured data includes equipment status and process parameter data from a programmable logic controller, physical parameter data from a sensor, and production plan information from a manufacturing execution system; the unstructured data includes quality inspection data from an automatic optical inspection device.

3. The intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line according to claim 1 is characterized in that: Generating the multi-source features includes: For the unstructured data, a lightweight convolutional neural network is used to identify and quantify welding defect features; For the structured data, the signal processing method is applied to the time-series physical parameters in the physical parameter data to extract the frequency domain features and the time-domain statistical features; the principal component analysis method is applied to the process parameter data and the non-time-series physical parameters in the physical parameter data to extract the principal component features; the categorical data in the equipment status and production plan information is encoded to generate categorical features; the welding defect features are combined with the frequency domain features, the time domain statistical features, the principal component features and the categorical features to form the multi-source features.

4. The intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line according to claim 1 is characterized in that: The process of outputting the production status dataset includes: The multi-source features are standardized, and the timestamp and the work order identifier are used as association keys to match and align the multi-source features of the same workpiece within a predetermined time window through a time window matching algorithm; Using a weighted fusion algorithm to perform weighted combination on the aligned multi-source features to generate a comprehensive production status indicator; using a probabilistic graphical model to infer the implicit production status based on the aligned multi-source features; The comprehensive production status indicator and the implicit production status are included in the output production status data set as a result.

5. The intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line according to claim 1 is characterized in that: Establishing and running the mechanism model specifically includes: For the selected production line equipment, establishing an ordinary differential equation model based on simplified physical principles; Acquiring real-time parameters reflecting the current operating state of the equipment from the production status data set, and using the real-time parameters as initial conditions of the ordinary differential equation model; Loading preset physical parameters related to the production line equipment; using a numerical calculation library to numerically solve the ordinary differential equation model, simulate and output the physical behavior.

6. The intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line according to claim 1 is characterized in that: By weighted fusion of the comprehensive production status indicator, the implicit production status and the physical behavior, specifically including: Performing a preliminary weighted combination on the comprehensive production status indicator and the implicit production status to generate a first fusion indicator; Extracting features of the physical behavior and converting them into a second fusion index; The first fusion index, the second fusion index, and the third fusion index selected from the multi-source features are again weightedly combined to generate and update the digital twin state.

7. The intelligent scheduling optimization method for a high-precision multi-layer PCBA fully automatic production line according to claim 1 is characterized in that: Generating a production scheduling plan using the heuristic optimization algorithm specifically includes: Using genetic algorithm as the heuristic optimization algorithm; Encode the production scheduling problem into individuals, where individuals represent candidate production scheduling solutions; Constructing a fitness function for evaluating the candidate production scheduling solutions; calculating the fitness function by weighted combination of scheduling objectives based on the multi-source features and the digital twin state, wherein the scheduling objectives include work order delivery delays, predicted production yield loss, and equipment energy consumption; Under the premise of satisfying the preset constraints, the selection, crossover and mutation operations are iteratively performed to optimize the search and determine the optimal individual with the optimal fitness value from a population of N individuals, and the optimal individual is decoded into the production scheduling plan.

Images