Hierarchical stop rail control method and system based on SMT line equipment abnormal linkage

Abstract

The application discloses a kind of based on SMT line body equipment abnormal linkage's hierarchical stop rail control method and system, its method includes the following steps: real-time acquisition equipment abnormal original data set and extract each equipment abnormal feature vector, determine abnormal grade by dynamic weight mapping, generate the hierarchical abnormal identification of first early warning, second failure or third major shutdown;According to hierarchical abnormal identification triggers stop rail control instruction to pack linkage control data packet, send to MES system and execute fault repair;MES system automatically dispatches carrier to buffer track area and generates carrier shunt scheduling instruction to execute stop rail shunt scheduling;After repair, generate recovery application instruction and upload to MES system, apply recovery operation and check operator authority level, legal after track reset instruction is issued, to recover automatic dispatching transfer;MES system records all the way, forms traceable quality data chain.The application can greatly improve SMT whole line production control efficiency.

Description

Technical Field

[0001] This invention relates to the field of SMT equipment control technology, specifically, to a graded track stop control method and system based on abnormal linkage of SMT line equipment. Background Technology

[0002] Currently, most SMT production line conveyor systems use independent PLC control, operating independently from pick-and-place machines, SPI, AOI, reflow ovens, loading / unloading machines, and MES systems. The interconnectivity, timeliness, and controllability of these systems directly determine the overall production yield, uptime, and quality traceability capabilities. During continuous SMT production, various malfunctions frequently occur, including pick-and-place machine material rejection, workstation material shortages, track blockages, visual inspection failures, abnormal reflow oven temperatures, and hardware alarms from various process equipment. Traditional independent PLC control of these tracks cannot capture real-time anomalies in the front-end, mid-end, and back-end process equipment. Even after a malfunction occurs, the track continues to transport PCB carriers to the faulty workstation, leading to material accumulation, track jamming, and line blockages. This not only results in a large number of scrapped PCBs and material waste but also easily triggers secondary malfunctions such as equipment jamming and overload damage, significantly increasing production losses and reducing the overall quality control efficiency of the SMT production line. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a graded track stop control method based on abnormal linkage of SMT line equipment, comprising the following steps: Step S1: Collect multi-dimensional operating data of SMT line equipment in real time, including the number of parts thrown, material balance, PCB transfer status, SPI / AOI inspection yield, oven temperature data and equipment alarm codes, and generate raw dataset of equipment anomalies. Step S2: Extract the abnormal feature vector of each device based on the original dataset of device abnormalities, combine it with the preset abnormal threshold matrix and the historical fault mode library, determine the abnormal level through dynamic weight mapping, and generate the graded abnormality identifiers corresponding to Level 1 warning, Level 2 fault or Level 3 major shutdown. Step S3: Trigger the corresponding level of rail stop control command according to the graded anomaly identifier, and package the rail stop level, fault equipment ID, fault code, rail stop timestamp, and affected workstation range into a linkage control data package; Step S4: Send the linkage control data packet to the MES system, which will lock the upstream process and material preparation task, stop feeding materials to the faulty section, generate a work order track stop event record and perform fault repair. Step S5: During the fault repair process at the stop-rail, the MES system automatically dispatches the on-the-way vehicles to the buffer track area and generates vehicle diversion and dispatch instructions to avoid workstation blockage and track jamming. Step S6: After the equipment fault is repaired, the linkage control parsing layer receives and verifies the equipment status, generates a recovery request instruction, and uploads it to the MES system; Step S7: The MES system requests the resumption of operation based on the recovery request command and verifies the operator's permissions. After the permissions are valid, the system issues a track reset command to restore automatic scheduling and transfer, and records a complete traceable quality data chain.

[0004] Furthermore, this invention also provides a graded track stop control system based on SMT line equipment abnormal linkage, used to execute the graded track stop control method based on SMT line equipment abnormal linkage as described above. It is characterized by comprising a MES main control server, a signal acquisition module, a fault level determination module, a track locking control module, a personnel access control module, an automatic start / stop / restart module, and a connecting platform track solenoid valve. The MES main control server communicates with each module and executes control logic. The signal acquisition module collects multi-dimensional operating data of the SMT linear equipment in real time and generates a raw dataset of equipment abnormalities. The fault level determination module determines the abnormality level based on the raw dataset and outputs a graded abnormality identifier. The track locking control module triggers the corresponding level of track stop control command according to the graded abnormality identifier and packages a linkage control data packet to perform fault repair. The personnel access control module verifies the operator's permissions after fault repair. The automatic start / stop / restart module performs track reset and resumption scheduling after the permission verification is passed. The connecting platform track solenoid valve performs physical-level track locking or release actions.

[0005] The beneficial effects of this application are as follows: By collecting multi-dimensional operational data from SMT line equipment in real time, covering the quantity of ejected components, material balance, PCB transfer status, SPI / AOI inspection yield, oven temperature data, and equipment alarm codes, a raw dataset of equipment anomalies is formed. This step, by aggregating multi-dimensional production condition information, comprehensively covers various abnormal scenarios that induce board blockage, defective products, and equipment overload, making up for the shortcomings of traditional production lines' lack of full-domain condition perception capabilities. This provides comprehensive and accurate raw data support for subsequent anomaly classification and linkage shutdown control. Multi-dimensional anomaly feature vectors are extracted from the raw equipment anomaly data. Based on a preset anomaly threshold matrix and a historical fault mode library, a dynamic weight mapping mechanism is used to determine the level of operational anomalies, generating corresponding level-based anomaly identifiers. This step, by combining historical fault data and quantitative judgment rules, classifies various process anomalies into levels, distinguishing between daily warnings, routine faults, and major shutdown accidents, forming a differentiated anomaly classification system. This provides a reliable basis for subsequent matching of shutdown control strategies with corresponding strengths. Based on the graded anomaly identifier, corresponding level-based track stop control commands are triggered to achieve differentiated control measures such as new task interception, fault location locking, and full-line emergency stop. Simultaneously, key fault information is packaged and integrated to generate a linkage control data package. This step establishes a graded linkage blocking mechanism, matching corresponding control intensity according to the fault level, locking the fault's impact range, and synchronously uploading standardized fault data, thereby achieving real-time linkage between equipment anomalies and track control. The linkage control data package is uploaded to the MES system, and through system parsing, it automatically locks the upstream loading and material preparation processes, terminating material feeding operations in the faulty section and generating a work order event record. This step achieves deep linkage between process equipment, track system, and upper-level production management system. Fault triggering immediately cuts off new material feeding at the source, curbing the spread of the fault, while retaining work order operation records, providing basic data for subsequent quality traceability and process optimization. During the fault repair shutdown phase, the MES system automatically dispatches various on-the-go vehicles to preset buffer tracks by combining vehicle position timing data and buffer capacity thresholds. This completes the diversion scheduling under fault conditions. This step relies on dynamic capacity determination and adaptive timing position data to divert and temporarily store vehicles, effectively clearing materials stuck on the conveyor tracks, avoiding secondary accumulation problems during the fault, ensuring normal production order in non-faulty sections, and reducing production material loss. After the equipment fault repair is completed, the SMT line equipment outputs a normal operation signal, and the linkage control parsing layer completes a comprehensive verification of the equipment's operating condition. After confirming that there are no hidden dangers, a recovery request instruction is generated and uploaded to the MES system. This step establishes a closed loop of status verification after fault repair, ensuring that all equipment operating parameters return to normal before initiating a recovery request, avoiding production fluctuations caused by residual hidden dangers, and ensuring stable operation of the production line after restarting.The MES system verifies operator permissions. Once authorized, a track reset command is issued to restore automatic transfer scheduling across the entire line and prioritize the release of buffer vehicles. Simultaneously, all fault event data throughout the process is fully retained to form a traceability chain. This step restarts the operation process through access control standards, reduces production delays by prioritizing the scheduling of buffer vehicles, and binds work orders to retain full-dimensional data, thus building a complete production quality traceability system and improving the standardization level of production control across the entire line. Attached Figure Description

[0006] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating the steps of the graded track stop control method based on abnormal linkage of SMT line equipment in the embodiment. Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S2. Figure 3 This is a block diagram of the module architecture of the graded track stop control system based on abnormal linkage of SMT line equipment in the embodiment. Figure 4 This is a flowchart illustrating the SMT anomaly acquisition and graded track stop control in the embodiment. Detailed Implementation

[0007] The following drawings disclose several embodiments of the present invention. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the invention. That is, in some embodiments of the invention, these practical details are not essential. Furthermore, for the sake of simplicity, some conventional structures and components will be shown in the drawings in a simple schematic manner.

[0008] To further understand the invention's content, features, and effects, the following embodiments are provided, and detailed descriptions are given below in conjunction with the accompanying drawings: Reference Figure 1 , Figure 1 This is a flowchart illustrating the steps of the graded track stop control method based on SMT line equipment abnormality linkage in this embodiment. The graded track stop control method based on SMT line equipment abnormality linkage in this embodiment includes the following steps: Step S1: Collect multi-dimensional operating data of SMT line equipment in real time, including the number of parts thrown, material balance, PCB transfer status, SPI / AOI inspection yield, oven temperature data and equipment alarm codes, and generate raw dataset of equipment anomalies. In this embodiment, the operating status of each SMT line device is monitored in real time. The number of components ejected from the pick-and-place machine is collected; this data is reported at fixed time intervals by the ejection counter inside the pick-and-place machine, recording the number of components that failed to be successfully placed per unit time. Material balance is collected from the feeder; this data is continuously read by the balance detection device on the feeder, reflecting the percentage or absolute quantity of material remaining in the current tape. PCB transport status is collected from the track control system; this data represents the PCB board's position, blockage, or idle status at each station in binary signal form. Inspection yield is collected from the SPI and AOI inspection instruments; this data reports the percentage of PCB board with acceptable solder paste printing quality or component soldering quality. Oven temperature data is collected from the reflow oven's temperature control module; this data includes the actual temperature value and set temperature value for each temperature zone. Equipment alarm codes are collected from the alarm output interface of each device; each alarm code corresponds to a specific anomaly type. All collected data is aligned with a unified timestamp, and each record includes the device ID, timestamp, and values ​​from the six dimensions mentioned above, and written into a time-series database table to generate the original dataset of equipment anomalies.

[0009] Step S2: Extract the abnormal feature vector of each device based on the original dataset of device abnormalities, combine it with the preset abnormal threshold matrix and the historical fault mode library, determine the abnormal level through dynamic weight mapping, and generate the graded abnormality identifiers corresponding to Level 1 warning, Level 2 fault or Level 3 major shutdown. In this embodiment, abnormal feature vectors for each device are extracted from the original dataset of device anomalies. These vectors contain six dimensions: material throwing frequency slope, material balance decay rate, transmission blockage duration, yield deviation standard deviation, temperature fluctuation amplitude, and alarm code frequency. The abnormal feature vectors are compared dimension-by-dimensionally with a preset abnormal threshold matrix to determine if each dimension exceeds the threshold range. Simultaneously, a historical fault mode library is accessed to retrieve frequently occurring fault features of the device in past batches of production, generating a time-series background feature vector. The abnormal feature vectors and the time-series background feature vectors are fused using a multimodal approach to generate a composite anomaly index. This composite anomaly index is input into a dynamic weight mapping model based on a graph attention network. This model uses the device topology graph as a structural prior and outputs the global influence weights of each anomaly feature. The composite anomaly index is weighted and aggregated using these global influence weights to obtain a comprehensive anomaly score. This score is then compared against an adaptive threshold boundary matrix. If the score falls into a low range, a level one warning indicator is output; if it falls into a middle range, a level two fault indicator is output; and if it falls into a high range, a level three major shutdown indicator is output.

[0010] Step S3: Trigger the corresponding level of rail stop control command according to the graded anomaly identifier, and package the rail stop level, fault equipment ID, fault code, rail stop timestamp, and affected workstation range into a linkage control data package; In this embodiment, a corresponding level of track stop control command is triggered based on the graded anomaly identifier. If the graded anomaly identifier is a Level 1 warning, the linkage controller sends a command to the track execution control module to prohibit the issuance of new transfer tasks. The track execution control module suspends the task scheduling interface of the faulty workstation and no longer assigns new PCB board transfer tasks to that workstation, while the remaining track areas maintain normal operation. If the graded anomaly identifier is a Level 2 fault, the linkage controller sends a command to lock the track points before and after the faulty workstation. The track execution control module drives the point locking devices on the track to lock the track sections at the entrance and exit of the faulty workstation, prohibiting any vehicle from entering or leaving the faulty area and isolating the logistics of the faulty workstation. If the graded anomaly identifier is a Level 3 major shutdown, the linkage controller sends a full-line track emergency stop command. The track execution control module immediately cuts off the drive power of all track sections, all vehicles brake urgently to stop moving, and all track transfer tasks are frozen. Simultaneously with triggering the command, the linkage controller packages five fields—track stop level, faulty equipment ID, fault code, track stop timestamp, and affected workstation range—into a linkage control data packet, ready to send it to the MES system.

[0011] Step S4: Send the linkage control data packet to the MES system, which will lock the upstream process and material preparation task, stop feeding materials to the faulty section, generate a work order track stop event record and perform fault repair. In this embodiment, the linkage control parsing layer sends the generated linkage control data packet to the linkage interface of the MES system via the industrial Ethernet protocol. After receiving the data packet, the MES system parses the affected workstation range field and calls the process topology mapping table to convert the range into the process node IDs of the upstream board loading process and the material preparation process. The process execution module of the MES system performs locking actions on these process nodes: it issues a pause command to the board loading process to stop the delivery of new PCB bare boards to the track; it issues a freeze command to the material preparation process to lock the inlet and outlet ports of the material queue and stop the delivery of material pallets to the faulty section. After the locking action is completed, the MES system merges the locking execution receipt with the linkage control data packet to generate a work order track stop event record, which includes the track stop level, faulty equipment ID, fault code, track stop timestamp, affected workstation range, locking success identifier, locking timestamp, and a snapshot of the process status after locking. This record is written to the MES event log library for persistent storage. On-site maintenance personnel locate the faulty equipment based on the fault code and equipment ID and begin to perform fault repair operations.

[0012] Step S5: During the fault repair process at the stop-rail, the MES system automatically dispatches the on-the-way vehicles to the buffer track area and generates vehicle diversion and dispatch instructions to avoid workstation blockage and track jamming. In this embodiment, during the track stop fault repair process, the MES system collects real-time position and time-series data of all vehicles on the track, including the current coordinates, moving speed vector, target station ID, dwell time, and vehicle type identifier for each vehicle. Simultaneously, it reads the current occupancy rate and remaining capacity of each buffer track area from the track PLC. Based on the distance and moving speed between the vehicle's current position and the faulty station, a Kalman filter is used to predict the future trajectory of each vehicle. Vehicles whose predicted trajectories conflict with the faulty station are marked as vehicles to be diverted. The future trajectories of the vehicles to be diverted are input into a conflict avoidance network along with the buffer capacity status. A graph search algorithm is used to plan one or more alternative paths to different buffer areas for each vehicle to be diverted. A multi-objective optimization model is used to simultaneously minimize the total travel time of all vehicles to be diverted and maximize the balanced utilization rate of the buffer areas, selecting the optimal diversion scheme from the Pareto front solution set. The optimal diversion scheme is converted into a vehicle diversion scheduling instruction executable by the track PLC, containing the ID of each vehicle to be diverted, target point coordinates, travel speed limit, and stop point number. The track execution control module controls the vehicle's steering, deceleration, and stopping actions according to instructions, guiding the vehicle to the buffer area to avoid workstation blockage and track jamming.

[0013] Step S6: After the equipment fault is repaired, the linkage control parsing layer receives and verifies the equipment status, generates a recovery request instruction, and uploads it to the MES system; In this embodiment, after the equipment fault is repaired, the SMT line equipment sends a normal operation signal to the linkage control parsing layer through the communication interface. The linkage control parsing layer extracts the equipment ID, equipment status code, self-test pass flag, and restart timestamp from this signal to generate an equipment recovery status vector. It retrieves the equipment's status baseline vector within the historical normal operation cycle from the historical database, calculates the Mahalanobis distance between the equipment recovery status vector and the status baseline vector, and obtains the status deviation index. The status deviation index is compared with a preset consistency threshold: if it is less than the threshold, the equipment status recovery is considered consistent, and a verification pass flag is generated; if it is greater than or equal to the threshold, the equipment status recovery is considered inconsistent, and a verification failure flag is generated, along with the name of the status dimension with the largest deviation and the direction of deviation. Based on the verification result, combined with the track stop level and fault code in the linkage control data packet, a recovery request instruction is generated. This instruction includes the equipment ID, recovery timestamp, verification result, and suggested recovery action. The recovery request instruction is uploaded to the MES system's recovery request service interface via the industrial Ethernet protocol and enters a waiting state, listening for permission verification and track reset instructions returned by the MES system.

[0014] Step S7: The MES system requests the resumption of operation based on the recovery request command and verifies the operator's permissions. After the permissions are valid, the system issues a track reset command to restore automatic scheduling and transfer, and records a complete traceable quality data chain.

[0015] In this embodiment, after receiving a recovery request command, the MES system parses the device ID and track stop level, and queries the minimum unlocking permission level corresponding to that track stop level. It retrieves the permission level of the currently requesting operator from the operator management database and compares it with the minimum unlocking permission level. If the operator's permission level is lower than the minimum unlocking permission level, the unlocking request is rejected, and an insufficient permission prompt is returned. If the operator's permission level reaches or exceeds the minimum unlocking permission level, the verification passes, and the MES system generates a track reset command, which is sent to the linkage control parsing layer via the industrial Ethernet protocol. The linkage control parsing layer forwards the track reset command to the track execution control module. The track execution control module releases the track point lock, restores the track drive power, and restarts the automatic scheduling and transfer process. During recovery operation, the track execution control module prioritizes scheduling vehicles parked in the buffer track area and re-enters them into the production line according to their original target workstation order. The MES system records the entire process, including the time of the fault occurrence, the time of the track stop trigger, the time of permission verification, the time of recovery, the fault type, the track stop duration, the recovery duration, and the operator ID. This data is then bound to the corresponding work order track stop event record, forming a complete and traceable quality data chain for subsequent quality analysis and auditing.

[0016] Furthermore, refer to Figure 2 , Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S2. Step S2 includes the following steps: Step S21: Extract the abnormal feature vectors corresponding to each SMT line equipment from the original dataset of equipment abnormalities, including the slope of the ejection frequency, the material balance decay rate, the duration of the transmission blockage, the standard deviation of the yield deviation, the temperature fluctuation amplitude, and the frequency of alarm codes. In this embodiment, by monitoring the operating status of each SMT line device in real time, a set of multi-dimensional abnormal feature vectors is extracted for each device from the original dataset of device anomalies. For the pick-and-place machine, the ratio of the number of ejections to the total number of placements per unit time is statistically analyzed. By calculating the linear regression slope of this curve within the most recent sliding time window, the ejection frequency slope is obtained, which reflects the speed at which the ejection problem worsens. For the feeder, the changes in the readings of the material balance sensor are monitored, and the rate at which the material balance decays over time is calculated, i.e., the material balance decay rate. If this rate deviates from the theoretical consumption value, it indicates a feeding anomaly. For the transport track, the dwell time of the PCB board at a specific station is recorded. When the dwell time exceeds a set benchmark value, the duration of the excess is accumulated to obtain the transport blockage duration. For the SPI and AOI inspection equipment, the inspection yield data of the most recent batch of products is collected, and the standard deviation of this batch of data is calculated to obtain the yield deviation standard deviation. The larger this value, the worse the process stability. For reflow ovens, the difference between the actual temperature and the set temperature in each temperature zone is collected, and the fluctuation amplitude of these differences within one cycle is calculated to obtain the temperature zone fluctuation amplitude. In addition, all alarm codes reported by each device within a unit of time are summarized, and the frequency of each code is counted to form the alarm code occurrence frequency. These six characteristic values ​​together constitute a six-dimensional anomaly feature vector, fully describing the abnormal state of the device at the current moment.

[0017] Step S22: Call the historical frequent fault feature set matched in the historical fault mode library to construct the time series background feature vector; fuse the abnormal feature vector with the time series background feature vector and standardize the process to generate a composite abnormal index; In this embodiment, the historical fault mode library is first accessed. Based on the unique device ID of the current SMT line equipment and the production batch number being executed, a set of historically frequent fault features matching the equipment and the batch of products is retrieved. This feature set records the fault modes that have repeatedly occurred in the same production batch and their corresponding feature sequences. These historical feature sequences are arranged in chronological order, and a temporal background feature vector characterizing the equipment degradation trend is constructed by calculating the changing trend of each feature on the time axis. This vector describes the baseline state of the equipment performance gradually deteriorating with the increase of production cycles. Next, the extracted real-time anomaly feature vector is fused with this temporal background feature vector in a multimodal manner. The fusion operation involves weighting and superimposing the two vectors in corresponding dimensions, where the weight of the real-time feature reflects the sudden intensity of the current anomaly, and the weight of the historical feature reflects the long-term accumulated degradation risk. After fusion, the resulting fused vector is standardized to eliminate the influence of different feature dimensions, and finally a scalar value, namely the composite anomaly index, is generated. This index integrates the current severity of the anomaly with the long-term accumulated health risk of the equipment, providing a unified quantitative basis for subsequent classification and judgment.

[0018] Step S23: Based on the composite anomaly index, the anomaly threshold matrix is ​​corrected in real time to generate an adaptive threshold boundary matrix; In this embodiment, an online learning autoencoder network is deployed, specifically designed to learn the dynamic distribution patterns of composite anomaly indicators. The autoencoder consists of an encoder and a decoder, and its training objective is to reconstruct the input as accurately as possible from the output. Under normal operating conditions, the autoencoder can reconstruct the input composite anomaly indicators well with relatively small reconstruction errors. When a new anomaly pattern appears in the device, the reconstruction error increases significantly. Utilizing this characteristic, the latest composite anomaly indicators are continuously received as input to drive the autoencoder for incremental online learning. Whenever a new indicator arrives, the current reconstruction error is calculated, and the network weights of the autoencoder are fine-tuned based on backpropagation of this error to adapt to the new data distribution. Simultaneously, the reconstruction error curve of the autoencoder is monitored. When the error curve experiences a sudden change or a sustained increase, it is determined that the current operating condition has drifted. Based on this determination, the anomaly threshold matrix is ​​dynamically adjusted. The initial values ​​of the threshold matrix are derived from historical experience, and each threshold element in the matrix is ​​adjusted by addition or subtraction based on the current reconstruction error level. For example, if the reconstruction error generally increases, the overall threshold is raised to avoid false alarms; conversely, the threshold is lowered to improve sensitivity. After this real-time correction process, an adaptive threshold boundary matrix that can dynamically change with the production line conditions is finally output.

[0019] Step S24: Perform causal correlation analysis on the abnormal feature vectors based on the adaptive threshold boundary matrix to identify secondary anomalies and primary anomalies, and calculate the independent contribution of each abnormal feature accordingly. In this embodiment, an in-depth causal correlation analysis is performed on the extracted abnormal feature vectors based on an adaptive threshold boundary matrix. The analysis process first traverses each feature in the abnormal feature vector, checking whether its value exceeds the upper or lower limits of the corresponding dimension in the adaptive threshold boundary matrix. For features exceeding the threshold, the causal relationship between them and other features is further analyzed. A Granger causality test based on time-series offset is used to determine the causal direction between two features by comparing whether the historical value of one feature can significantly predict the future value of another feature. For example, if it is found that the increase in the slope of the material discarding frequency always precedes the deterioration of the standard deviation of the yield deviation, then the former is determined to be the cause and the latter to be the effect. Based on this, all abnormal features are classified into two categories: one is secondary abnormalities directly caused by insufficient material supply, nozzle blockage, etc., whose characteristics are unidirectional driving effects on other features; the other is primary abnormalities caused by mechanical wear, aging, etc., whose characteristics are source-level effects on other features. After classification, the independent contribution of each abnormal feature is calculated. The independent contribution is calculated based on the ratio of the out-degree to the in-degree of the feature in the initial causal graph. The out-degree represents the number of times the feature influences other features as a cause, and the in-degree represents the number of times the feature is influenced by other features as a result. The out-degree of a primary anomalous feature is much greater than its in-degree, and its independent contribution equals its out-degree divided by the sum of all out-degrees. The in-degree of a secondary anomalous feature is greater than its out-degree, and its independent contribution equals its in-degree divided by the sum of all in-degrees. The contribution of a primary anomalous feature is usually higher than that of a secondary anomalous feature. The contribution is represented by a value between 0 and 1, and the sum of the contributions of all features is 1.

[0020] Step S25: Input the independent contribution of each abnormal feature into the dynamic weight mapping model, where the equipment topology diagram is used as the structural prior and the fault propagation path in the historical fault mode library is used as the training sample, and output the global influence weight of each abnormal feature in the current equipment interconnection state. In this embodiment, a dynamic weight mapping model based on a graph attention network is constructed. The input to this model is the calculated independent contribution vector corresponding to each anomalous feature. The prior structure of the model is a device topology graph, where nodes represent each SMT device and edges represent the physical connections between devices, such as the conveyor track connection between a pick-and-place machine and a reflow oven. The model uses fault propagation paths from a historical fault pattern library as training samples for supervised learning. Each training sample contains a known fault initiation device, the fault propagation process along the topology graph, and the final affected device range. During training, the mean squared error loss function is used to calculate the difference between the global impact weights output by the model and the actual impact range of historical faults. The parameters of the linear transformation layer and the attention mechanism in the graph attention network are updated through backpropagation. Through training with a large number of such samples, the graph attention network learns how to dynamically allocate attention weights to different devices and features based on the current anomalous state and topological connections of the devices. During actual inference, the current independent contribution vector and the real-time updated device topology are used to perform the inference. Figure 1 The input model is identical. The model's internal attention mechanism calculates the influence coefficient of each node on its neighboring nodes and aggregates neighbor information through multiple layers of network, ultimately outputting a global influence weight vector with the same dimension as the input features. Each element in this weight vector represents the global influence of the corresponding abnormal feature in the current device interconnection state. The higher the weight, the greater the damage to the entire production line if the feature goes out of control.

[0021] Step S26: Use global influence weights to perform weighted aggregation of composite anomaly indicators to obtain a comprehensive anomaly score, and compare it with the adaptive threshold boundary matrix to automatically match and output the graded anomaly identifiers corresponding to Level 1 warning, Level 2 fault, or Level 3 major shutdown.

[0022] In this embodiment, the generated global impact weight vector and the composite anomaly index are weighted and aggregated. The aggregation operation involves multiplying each element of the global impact weight vector by the composite anomaly index to obtain a weighted anomaly intensity value. Subsequently, all weighted values ​​are summed to obtain a single comprehensive anomaly score. This score not only considers the severity of the anomaly but also incorporates the anomaly's propagation risk and destructive potential throughout the entire production line network. Next, this comprehensive anomaly score is compared with an adaptive threshold boundary matrix. The threshold matrix defines three consecutive numerical intervals, corresponding to Level 1 warning, Level 2 fault, and Level 3 major shutdown, respectively. The interval into which the comprehensive anomaly score falls is determined, and the corresponding graded anomaly identifier is automatically matched and output. The dynamic adjustment process of the adaptive threshold boundary matrix and the generation process of the global impact weights are not performed in isolation but form a closed loop of mutual iterative optimization. After each determination, the determination result is compared with the actual fault consequences, and the error signal generated by the comparison is simultaneously fed back. On the one hand, the online learning parameters of the autoencoder are corrected to adjust the threshold matrix; on the other hand, the attention weights of the graph attention network are updated to optimize the global impact weights. This process continues to iterate until the hierarchical decision boundary reaches its optimum under the current steady state of the production line, that is, the false alarm rate and the false negative rate are reduced to the lowest level.

[0023] Furthermore, the process of fusing and standardizing the abnormal feature vector with the temporal background feature vector in step S22 includes the following steps: Extract the current-time anomaly intensity vector from the anomaly feature vector, and extract the historical cumulative risk vector from the temporal background feature vector; In this embodiment, for the previously extracted six-dimensional anomaly feature vector, in-depth analysis of the statistical distribution characteristics of each dimension is performed. For the dimension of the material throwing frequency slope, all observations of this slope within the most recent time window are collected, and a probability density function for this dimension's data is fitted. Based on this, the third-order central moment of this probability density function, i.e., skewness, is extracted to measure the asymmetry of the data distribution; the fourth-order central moment, i.e., kurtosis, is extracted to measure the sharpness or flatness of the data distribution. Simultaneously, the mean and variance of this probability density function are calculated, together forming the higher-order moment feature set for this dimension. The same operation is applied sequentially to the five dimensions: material balance decay rate, transmission blockage duration, yield deviation standard deviation, temperature fluctuation amplitude, and alarm code occurrence frequency. Each dimension ultimately yields a sub-vector composed of four values: skewness, kurtosis, mean, and variance. These six sub-vectors are concatenated sequentially to form a twenty-four-dimensional anomaly intensity vector for the current moment. This vector no longer merely describes the magnitude of abnormal values, but delves into the position and shape of each abnormal feature within its own statistical distribution, providing a richer and more robust anomaly representation for subsequent fusion. Simultaneously, three key historical features closely related to the current production batch number and equipment ID are extracted from the temporal background feature vector. The first feature is the historical failure frequency period feature. Specifically, this involves retrieving the equipment's historical production history for this batch, identifying the time period with the highest failure frequency, such as a specific shift or production hour, and encoding this time period number as a discrete feature into the vector. The second feature is the equipment degradation trend slope. This value is obtained by performing a linear regression on a composite anomaly index sequence from multiple past production runs of the equipment. The slope of the regression line is the degradation trend slope; a positive value indicates continuous performance decline, while a negative value indicates improvement. The third feature is the average recovery time for similar historical anomalies. All failure events of the same type as the current anomaly in the equipment's history are selected, and the duration from occurrence to recovery for each failure is extracted, calculating the arithmetic mean of these durations. These three feature values ​​are combined into a three-dimensional historical cumulative risk vector. This vector encapsulates the historical behavior patterns of the equipment during this batch of production, providing a long-term perspective for assessing current risks.

[0024] Based on the cross-modal attention mechanism, feature-level cross-enhancement is performed on the current anomaly intensity vector and the historical cumulative risk vector to generate a fused enhanced feature tensor; In this embodiment, a multimodal fusion network is constructed, which implements a cross-modal attention mechanism. The 24-dimensional anomaly intensity vector at the current moment is set as the query vector, and the 3-dimensional historical cumulative risk vector is set as the key-value vector. The calculation process of the attention mechanism is as follows: First, a linear transformation layer maps the query vector to a query space, and another linear transformation layer maps the key-value vector to a key space, maintaining the same dimensions in both spaces. Then, the dot product similarity between each element in the query space and each element in the key space is calculated, resulting in an attention score matrix. The number of rows in this matrix equals the dimension of the query vector, and the number of columns equals the dimension of the key-value vector. Next, this attention score matrix is ​​normalized along the column direction, so that the sum of the scores in each row is 1, resulting in an attention weight matrix. Using this attention weight matrix, the key-value vectors are weighted and summed to obtain a set of converged attention results. Simultaneously, the query vectors are similarly weighted and summed using the transpose of the attention weight matrix. Finally, the original query vector, the original key-value vector, and the two sets of attention convergence results are concatenated to form a high-dimensional augmented feature tensor. This tensor achieves deep cross-enhancement between the two modalities.

[0025] Centering and scaling corrections are performed based on the enhanced feature tensor to eliminate dimensional differences and distribution shifts between different modal features, generating a standardized fused feature vector. In this embodiment, the output augmented feature tensor is fed into a learnable normalization transformation layer. The core of this transformation layer is a set of trainable parameters, including a scaling factor vector and a shift factor vector. During processing, the mean and standard deviation vectors of the augmented feature tensor on the current batch of data are first calculated. Then, for each element in the augmented feature tensor, the mean of its corresponding dimension is subtracted, and then divided by its corresponding standard deviation to complete centering and scaling correction, eliminating inherent dimensional differences and distribution shifts between different modal features. This step unifies features from different sources and with different numerical ranges into a standard normal distribution space with a mean of 0 and a standard deviation of 1. Subsequently, the normalized data is scaled element-wise using the learnable scaling factor vector, and then translated element-wise using the learnable shift factor vector. These two learnable parameters allow the network to autonomously adjust the scale and center position of the feature distribution after normalization to adapt to the specific needs of subsequent tasks. After this series of operations, the augmented feature tensor is transformed into a well-distributed, dimensionally uniform, normalized fused feature vector.

[0026] The standardized fusion feature vector is input into the nonlinear activation mapping layer, and the composite anomaly index is output.

[0027] In this embodiment, the output standardized fused feature vector is input into a nonlinear activation mapping layer. This mapping layer consists of a three-layer multilayer perceptron. The first layer is a fully connected layer that maps the dimension of the standardized fused feature vector to a higher hidden dimension, for example, from high dimension to 128 dimensions, followed by a ReLU nonlinear activation function to introduce nonlinear transformation capability. The second layer is also a fully connected layer that maps the 128-dimensional hidden features to a lower dimension, for example, 64 dimensions, followed by another ReLU activation function to further extract higher-order interaction features. The third layer is an output layer that maps the 64-dimensional features to a single scalar value. This scalar value is the composite anomaly index. All weights and bias parameters of the multilayer perceptron are obtained through end-to-end backpropagation training. The entire mapping process essentially allows the network to automatically learn the complex nonlinear combination relationships between the various dimensions in the standardized fused feature vector, thereby capturing the deep coupling effect between the current anomaly intensity and historical accumulated risk. The final output composite anomaly index is a highly condensed value that combines the intensity of real-time sudden anomalies with the long-term cumulative risk of degradation, providing a reliable decision-making basis for subsequent threshold comparison and classification.

[0028] Furthermore, the causal correlation analysis of the abnormal feature vectors based on the adaptive threshold boundary matrix described in step S24 includes the following steps: Candidate abnormal nodes are selected from abnormal feature vectors based on an adaptive threshold boundary matrix. In this embodiment, for the extracted six-dimensional abnormal feature vector, the value of each dimension is compared with the upper and lower limits of the corresponding dimension in the generated adaptive threshold boundary matrix. When the value of a dimension exceeds the upper or lower limit set by the boundary matrix, the abnormal feature corresponding to that dimension is marked as a candidate abnormal node. For example, if the slope of the material throwing frequency exceeds the upper limit of the threshold, the feature is marked as a candidate node. All marked candidate abnormal nodes constitute a candidate set. At the same time, two time-series curves directly related to material flow are extracted separately from the multi-dimensional operating data. The first curve is the material balance decrease rate curve, which is calculated by the difference in readings of the material balance sensor on the feeder at fixed time intervals, depicting the trajectory of material consumption rate changing over time. The second curve is the sequence of abrupt change points in the quantity of material thrown, which is located by detecting the jump amplitude of the throwing counter per unit time. When the jump amplitude exceeds a preset multiple, the moment is recorded as the abrupt change point. These two time-series curves are aligned along the time axis and plotted on a graph to construct a material flow time-series map. This graph visually demonstrates the temporal correlation between material consumption behavior and sudden material disposal events, providing an important spatiotemporal reference for subsequent causal analysis.

[0029] Candidate abnormal nodes are input into the causal discovery network. The physical topology of the device is used as a priori constraint, and the known causal chains in the historical fault mode library are used as supervision signals. The directed causal relationship between each abnormal feature is automatically learned through iterative pruning, and the initial causal graph is output. In this embodiment, the set of candidate anomaly nodes and the material flow timeline graph are fed into a causal discovery network. This network uses the physical topology graph of the SMT production line as a priori constraint. The physical topology graph clarifies the upstream and downstream connections between devices; for example, the pick-and-place machine is located upstream of the reflow oven, and this physical connection determines the possible direction of anomaly propagation. The causal discovery network employs an iterative pruning algorithm based on conditional independence testing. The algorithm starts with a fully connected directed graph where each candidate anomaly node is connected to a bidirectional potential causal edge. Then, the network uses known causal chains recorded in the historical failure mode library as monitoring signals, gradually removing edges that contradict known causal chains or fail the conditional independence test. In each iteration, the network calculates the conditional mutual information between nodes in the current graph structure. If the conditional mutual information of an edge is below a threshold, the edge is pruned. After multiple rounds of iterative pruning, the remaining directed edges form a sparse directed acyclic graph, i.e., the initial causal graph. Each directed edge in the graph represents a potential causal relationship from a cause node to an effect node, such as from "material balance decay rate" to "discharge frequency slope".

[0030] Counterfactual intervention is performed on each candidate anomalous node based on the initial causal graph to determine whether it is a root cause node or a transitive node; In this embodiment, based on the output initial causal graph, counterfactual intervention is performed sequentially on each candidate anomalous node. Taking one candidate anomalous node as an example, the current observation values ​​of all other candidate anomalous nodes in the initial causal graph are kept unchanged, and only the current observation value of this node is replaced with the expected value of this node under normal operating conditions. The expected value under normal operating conditions is taken from the average statistical value of this feature during the fault-free history of the device. After the replacement, the forward inference capability of the causal discovery network is used to propagate downstream from this node along the directed edges in the initial causal graph, simulating the response changes of all downstream nodes under this counterfactual condition. If, after inference, the simulated value of a downstream node that was originally in an anomalous state falls back to the normal threshold range, that is, its anomalous state disappears, then the currently intervened node is determined to be the root cause of the anomalous state of the downstream node. Conversely, if the anomalous state of all downstream nodes does not disappear, then the current node is determined to be only a transitive node, and its own anomalousness is caused by a more upstream cause. The above counterfactual intervention and inference process is repeated for all candidate anomalous nodes, and finally all nodes are classified into two categories: root cause nodes and transitive nodes.

[0031] All candidate abnormal nodes identified as root cause nodes are mapped back to their corresponding physical parameter domains to establish a spatiotemporal alignment relationship between root cause nodes and physical phenomena. Based on the spatiotemporal alignment relationship, root cause nodes that have a strong causal relationship with the decrease in material balance are classified as secondary abnormalities caused by insufficient material supply, while root cause nodes that have no direct causal relationship with material balance but trigger sudden changes in material ejection are classified as primary abnormalities caused by mechanical wear of equipment.

[0032] In this embodiment, all candidate abnormal nodes identified as root cause nodes are mapped from the abstract feature space back to their corresponding physical parameter domains. For example, the root cause node "discharge frequency slope" is mapped back to the discharging action parameters of the pick-and-place machine, and "material balance decay rate" is mapped back to the feeding physical process of the feeder. After mapping, a spatiotemporal alignment relationship between each root cause node and a specific physical phenomenon is established by combining the constructed material flow time sequence map. Specifically, the abnormal occurrence time of each root cause node is compared and aligned on the time axis with the material balance depletion time point and the bursting start point of the discharging quantity in the material flow time sequence map. If the abnormal occurrence time of a root cause node and the material balance depletion time point highly coincide on the time axis, and the waveform changes of the two show a strong correlation, that is, the abnormality of the node intensifies as the material balance decrease rate accelerates, then it is determined that there is a strong causal relationship between the root cause node and the material balance decrease, and it is classified as a secondary abnormality caused by insufficient material supply. Conversely, if the occurrence of an anomaly at a root node is not significantly related to changes in the material balance, but can directly trigger a sudden surge in the amount of material thrown, then the root node is determined to have no direct causal relationship with the material balance and is classified as an anomaly caused by mechanical wear of equipment.

[0033] Furthermore, establishing the spatiotemporal alignment relationship between the root node and the physical phenomenon includes the following steps: Extract all candidate abnormal nodes identified as root cause nodes and their corresponding abnormal dimension identifiers. At the same time, retrieve the original physical parameter time series curves corresponding to the abnormal dimension identifiers from the original dataset of equipment abnormalities, including the material balance decay curve, the cumulative growth curve of the number of discarded materials, the PCB transmission blockage flag sequence, the detection yield deviation curve, the furnace temperature deviation curve, and the equipment alarm frequency histogram. In this embodiment, all candidate abnormal nodes identified as root cause nodes are extracted, and the abnormal dimension identifier corresponding to each node is recorded, such as the throwing frequency slope and the material balance decay rate. For each abnormal dimension identifier, the corresponding original physical parameter time-series curve is retrieved from the original dataset of equipment abnormalities. For the dimension of material balance decay rate, the decay curve of material balance changing over time is retrieved, which records the continuous change trajectory of the remaining material in the feeder. For the dimension of throwing frequency slope, the cumulative growth curve of the number of thrown materials changing over time is retrieved, which depicts the monotonically increasing process of the total number of thrown materials. For the dimension of transmission blockage duration, the blockage flag bit sequence of PCB transmission status changing over time is retrieved, which is a binary signal, where 1 indicates blockage and 0 indicates unobstructed. For the dimension of yield deviation standard deviation, the deviation curve of SPI / AOI detection yield changing over time is retrieved, which shows the degree of deviation of the yield from the target value. For the dimension of temperature fluctuation amplitude, the temperature deviation curve of furnace temperature data changing over time is retrieved, which records the difference between the actual furnace temperature and the set temperature. For the dimension of alarm code occurrence frequency, an alarm frequency histogram of the device alarm codes over time is retrieved. This histogram counts the number of alarms occurring at fixed time intervals. These six original physical parameter time-series curves are then bundled with the corresponding root cause node identifiers to form a complete root cause node physical mapping set.

[0034] Analyze the time-series curves of the original physical parameters and extract the timestamps of key events, including candidate time points for material depletion, candidate time points for material ejection, plate blockage start time points, yield degradation start points, temperature instability start points, and alarm density start points. In this embodiment, based on each time-series curve in the root node physical mapping set, specific signal processing operations are performed to extract the timestamps of key events. A first-order difference operation is performed on the material balance decay curve to calculate the change in material balance at adjacent time points, resulting in a slope sequence. The abrupt change point with the largest absolute value is found in this slope sequence; this point corresponds to the moment when material consumption is fastest and is marked as a candidate time point for material depletion. A second-order difference operation is performed on the cumulative growth curve of the number of ejected materials, and a further difference of its first-order difference is calculated to obtain a curvature sequence. The inflection point where the sign reverses is found in this curvature sequence; this point corresponds to the turning point of material ejection acceleration or deceleration and is marked as a candidate time point for material ejection outbreak. Rising edge detection is performed on the blockage flag sequence to find the first moment when the signal jumps from 0 to 1, marked as the blockage initiation time point. Zero-crossing detection is performed on the yield deviation curve to find the moment when the deviation value changes from negative to positive and first crosses a positive deviation threshold, marked as the yield degradation initiation point. A sliding window variance analysis was performed on the temperature deviation curve. A fixed-length sliding window was set, and the variance of the temperature deviation within the window was calculated. When the variance value jumped and exceeded a baseline level, the end time of the window was recorded and marked as the onset point of temperature instability. A cumulative sum detection was performed on the alarm frequency histogram. The cumulative sum sequence of alarm frequencies was calculated, and when the cumulative sum first exceeded a multiple of the baseline noise level, that moment was recorded and marked as the onset point of alarm density. All extracted timestamps were compiled into a multi-physical key event timestamp set.

[0035] The activation time window of the root node in the initial causal graph and the key event timestamps are input into the spatiotemporal alignment network. By adopting a dynamic time warping algorithm, the activation time window is used as the starting point of the warping path, and each timestamp in the key event timestamps is used as a candidate endpoint of the warping path. The optimal time offset between each root node and its corresponding physical phenomenon is calculated by minimizing the cumulative path distance. At the same time, the ratio of the cumulative path distance to the path length corresponding to the optimal time offset is calculated as the normalized alignment cost, generating a spatiotemporal alignment offset vector and a normalized alignment cost matrix. In this embodiment, the activation time window of each root node in the initial causal graph, along with the set of timestamps for multi-physical key events, is input into a spatiotemporal alignment network. This network employs a dynamic time warping algorithm for temporal alignment. The activation time window of the root node serves as the starting sequence of the warped path, and each timestamp in the set of timestamps for multi-physical key events serves as a candidate endpoint for the warped path. The dynamic time warping algorithm constructs a cumulative distance matrix, where rows correspond to the time points of the starting sequence, and columns correspond to the time points of the candidate endpoints. Starting from the top left corner of the matrix, the algorithm fills the matrix from left to right and top to bottom. The value of each cell equals the Euclidean distance between the current two time points plus the minimum value among the three adjacent cells to the left, above, and top left. After filling, a path with the minimum cumulative distance is traced back from the bottom right corner of the matrix; this path is the optimal warped path. The total cumulative distance of this path divided by the path length yields a normalized alignment cost. It is important to note that when the activation time window length of the root node is inconsistent with the number of timestamps of the candidate endpoints, the dynamic time warping algorithm can automatically perform nonlinear scaling and alignment on the two sequences without requiring them to be of equal length. The above process is repeated for each combination of the root node and each candidate endpoint to obtain the optimal time offset and normalized alignment cost between that node and each physical event. The results for all nodes are then integrated to generate a spatiotemporal alignment offset vector and a normalized alignment cost matrix.

[0036] The activation time window of the root cause node is shifted and corrected based on the spatiotemporal alignment offset vector and the normalized alignment cost matrix. This ensures that the corrected activation time window has maximum overlap with at least one of the following key event time points: material depletion candidate time point, material ejection outbreak candidate time point, plate blockage start time point, yield degradation start point, temperature instability start point, and alarm density start point. This establishes a spatiotemporal alignment relationship between the root cause node and the physical phenomenon.

[0037] In this embodiment, the activation time window of each root node is shifted and corrected based on the generated spatiotemporal alignment offset vector and the normalized alignment cost matrix. Specifically, for a given root node, the normalized alignment cost between the node and all candidate physical event timestamps is extracted from the normalized alignment cost matrix, and the candidate event with the lowest cost is found. Using the corresponding offset in the spatiotemporal alignment offset vector, the activation time window of the root node is shifted forward or backward on the time axis, so that the shifted activation time window overlaps with the timestamp of the candidate event to the greatest extent. For example, if the minimum alignment cost of the root node "discharge frequency slope" corresponds to the "discharge outbreak candidate time point", then the activation time window of "discharge frequency slope" is shifted so that its center point coincides with the "discharge outbreak candidate time point". After translation correction, it is necessary to verify whether the activation time window of the root node partially overlaps with the time point of at least one other physical key event. If there is an overlap, the reliability of the alignment relationship needs to be further confirmed. If there is no overlap, the alignment relationship is marked as weak and requires manual review. After translation correction and reliability verification, the root node establishes a clear spatiotemporal alignment relationship with the corresponding physical phenomenon. This operation is repeated to find the best-matching physical key event for each root node, thereby establishing a complete spatiotemporal alignment mapping table between root nodes and physical phenomena.

[0038] Furthermore, step S25, which involves inputting the independent contribution of each anomalous feature into the dynamic weight mapping model, includes the following steps: Extract the independent contribution of each abnormal feature to generate a labeled node contribution feature set; In this embodiment, the independent contribution values ​​corresponding to each calculated anomaly feature are extracted. These contribution values ​​have been classified into secondary anomalies and primary anomalies. The contribution value values ​​corresponding to secondary anomalies are organized into a secondary anomaly contribution value vector, and the contribution value values ​​corresponding to primary anomalies are organized into a primary anomaly contribution value vector. For each contribution value, its corresponding device node ID is labeled, such as pick-and-place machine A, reflow oven B, SPI detector C, etc., and its anomaly type label is also labeled, such as "insufficient material supply type secondary anomaly" or "mechanical wear type primary anomaly". Each contribution value thus becomes a triple: contribution value, device node ID, and anomaly type label. All such triples are grouped and merged according to device node ID, and all contribution values ​​under the same device node ID are merged into a sub-vector. Finally, a labeled node contribution feature set is formed, with device node ID as the index and contribution sub-vectors and anomaly type label set as values. This feature set clearly and structurally organizes the anomalies occurring on each device node and their severity and nature, preparing for subsequent propagation calculations on the graph structure.

[0039] Based on the physical layout and material flow path of SMT linear equipment, construct the equipment topology diagram; In this embodiment, a directed weighted graph is constructed based on the actual physical layout and material flow path of the SMT line. Each node in the graph corresponds to a specific SMT line device, such as a board mounter, printer, pick-and-place machine A, pick-and-place machine B, reflow oven, SPI inspection instrument, AOI inspection instrument, unmounter, etc. Each directed edge in the graph represents a material transfer link or signal interaction link between two devices. The direction of the material transfer link is consistent with the physical flow direction of the PCB board, for example, from pick-and-place machine A to the reflow oven. The signal interaction link represents the exchange of control signals and data packets between devices via industrial Ethernet. For each edge, a set of link attribute parameters are assigned. The transmission delay parameter represents the average time required for material or signal to travel from the upstream device to the downstream device, in milliseconds. The signal attenuation parameter represents the degree of decrease in signal-to-noise ratio during transmission, dimensionless. These attribute parameters are obtained through on-site measurements and statistical analysis of historical data and are embedded in the edge data structure. The completed equipment topology diagram fully depicts the physical connections and communication dependencies between production line equipment, providing a structural framework for subsequent graph neural network message passing.

[0040] The node contribution feature set and the device topology graph are input into the dynamic weight mapping model. The node contribution feature is used as the initial embedding vector, the adjacency matrix of the device topology graph is used as the message passing skeleton, and the influence coefficient of each node on its neighboring nodes is calculated through the attention mechanism to generate a dynamically updated node embedding representation. In this embodiment, the generated labeled node contribution feature set is used as the initial embedding vector and assigned to each corresponding node in the device topology graph. The adjacency matrix of the device topology graph is extracted and used as the skeleton for message passing in the graph attention network. The graph attention network employs a multi-head attention mechanism to calculate the influence coefficient of each node on its neighbors. Specifically, for each node, its initial embedding vector is mapped to a query vector, key vector, and value vector through three different linear transformation layers. Then, the dot product similarity between the node's query vector and the key vectors of all neighboring nodes is calculated to obtain an attention score. This score is normalized so that the sum of the attention scores of all neighboring nodes is 1, resulting in an attention coefficient. The multi-head attention mechanism executes multiple sets of such calculations in parallel, each using different linear transformation parameters, resulting in multiple sets of attention coefficients. These multiple sets of attention coefficients are weighted and summed on the value vectors of neighboring nodes, and the sums are concatenated and subjected to a linear transformation layer for dimensionality reduction to obtain the updated embedding representation of the node. All nodes synchronously perform the above operations to generate a dynamically updated node embedding representation.

[0041] Based on node embedding representation, the dynamic weight mapping model propagates layer by layer along the edges of the device topology graph. During each layer of propagation, the contribution features of neighboring nodes are weighted and aggregated using the influence coefficient. At the same time, residual connections are introduced to retain the node state of the previous layer, generating node features after propagation. In this embodiment, based on the generated node embedding representation, the message passing layer of the startup graph attention network propagates layer by layer. The operation of each message passing layer consists of three steps. First, using the attention coefficients of the current layer, the contribution features of all neighboring nodes of each node are weighted and aggregated. During aggregation, the attention coefficients determine the importance of each neighboring node's information; neighbors with higher weights contribute more information. Second, the aggregated neighbor information is concatenated with the current node's own embedding representation, then passed through a fully connected layer and a non-linear activation function to obtain the node's temporary features at this layer. The non-linear activation function is specifically a rectified linear unit (RCU), which sets all input values ​​less than zero to zero and keeps the values ​​greater than zero unchanged. After processing by the RCU, negative input values ​​are truncated to zero, while positive input values ​​pass through linearly, thus introducing sparsity and non-linear expressive power into the network, effectively alleviating the gradient vanishing problem and accelerating training convergence. Third, a residual connection is introduced, adding the node state obtained from the previous layer's propagation to the temporary features of this layer, resulting in the final node feature output of this layer. The introduction of residual connections ensures that gradients in deep networks can be smoothly backpropagated, avoiding the vanishing gradient problem. Through multiple layers of such message passing, the feature representation of each node not only contains its own information but also incorporates information from its multi-hop neighbor nodes, thereby capturing the propagation path and spread range of anomalies in the device topology graph. After all layers have propagated, the final propagated node features are obtained.

[0042] The propagated node features are aggregated through a global pooling operation, and the global impact weight of each abnormal feature in the current device interconnection state is output.

[0043] In this embodiment, the output propagated node features are input into the readout layer of the graph attention network. The readout layer performs a global pooling operation, aggregating the propagated features of all nodes. The global pooling operation uses a combination of average pooling and max pooling: first, the average pooling vector is calculated for each feature dimension across all nodes; simultaneously, the maximum value for each feature dimension across all nodes is obtained, resulting in a max pooling vector; then, the average pooling vector and the max pooling vector are concatenated to form a single global feature vector. This global feature vector encapsulates the abnormal states and their propagation impact of all nodes in the entire device topology graph. Finally, the global feature vector is input into a fully connected output layer, the number of neurons in which is equal to the total number of dimensions of the abnormal features. The fully connected layer maps the global feature vector back to the original feature space, outputting a weight vector that corresponds one-to-one with the dimensions of the abnormal features, i.e., the global impact weight. Each element in this weight vector represents the global influence of the corresponding abnormal feature in the current device interconnection state; the higher the weight, the greater the destructive impact on the entire production line should the abnormality get out of control.

[0044] Furthermore, the step S4 of sending the linkage control data packet to the MES system includes the following steps: Extract the track stop level, faulty equipment ID, fault code, track stop timestamp, and affected workstation range from the linkage control data packet to generate a structured parsed field set; In this embodiment, after receiving the linkage control data packet uploaded from the track execution control layer or the equipment signal acquisition layer, the linkage control parsing layer performs field-by-field parsing and extraction on the data packet. The track stop level field is read from a fixed offset position in the data packet header. This field's value is the enumerated code corresponding to Level 1 warning, Level 2 fault, or Level 3 major shutdown. Next, the fault equipment ID field is read from the data packet payload area. This field is a unique equipment identifier, such as the equipment number of pick-and-place machine A. The fault code field is then parsed, recording the specific anomaly type code reported by the equipment, such as a material throwing over-limit code or a temperature zone over-temperature code. The track stop timestamp is extracted from the data packet timestamp field. This timestamp, accurate to the millisecond, records the absolute moment the track stop event was triggered. Finally, the affected workstation range is parsed from the workstation range field at the end of the data packet. This range is given in the form of an interval between the start and end workstation numbers, such as workstation 05 to workstation 08. The values ​​of these five fields are reorganized according to a fixed data structure to generate a structured parsing field set. This field set converts the original binary data packet into a key-value pair format that can be directly referenced by the upper-layer business logic, providing a clear input basis for subsequent MES linkage operations.

[0045] Based on the structured parsed field set, the process topology mapping table in the MES system is called to determine the snapshot of the process status to be locked; In this embodiment, based on the affected workstation range field in the structured parsed field set, the process topology mapping table maintained in the MES system is invoked. This mapping table records the correspondence between each physical workstation number and the MES process node ID. For example, physical workstation 05 corresponds to board loading process node ID_UB01, and physical workstation 06 corresponds to material preparation process node ID_BL03. By looking up the table, the affected workstation range is converted into a list of process node IDs for the upstream board loading and material preparation processes. For each process node ID in the list, a status query request is sent to the process execution module of the MES system to obtain the current execution status of the process node, such as executing, idle waiting, or completed. At the same time, the material queue length of the process node is obtained, that is, the number of PCB boards or material pallets currently waiting to be processed by the process. The obtained execution status and material queue length are bound with the corresponding process node ID to generate a snapshot of the status of the process to be locked. This snapshot is a structured dataset that records the real-time operating status of all affected MES process nodes when the track stop event occurs, providing accurate on-site data for subsequent locking strategy decisions.

[0046] Input the snapshot of the status of the process to be locked and the track stop level into the locking strategy decision tree, and output the locking action instructions for each process to be locked, including prohibiting the issuance of new tasks, pausing the execution of the current task, and freezing the material queue allocation; In this embodiment, the snapshot of the process status to be locked and the parsed stoppage level are input into a locking strategy decision tree. This decision tree is a predefined rule engine, with the stoppage level as its root node. The first-level branches of the decision tree expand according to the stoppage level: if the stoppage level is Level 1 warning, it enters the warning branch; if it is Level 2 fault, it enters the fault branch; if it is Level 3 major shutdown, it enters the shutdown branch. Each branch further refines the granularity of the locking action by using the process execution status and material queue length in the snapshot of the process status to be locked as leaf node attributes. For example, under the Level 2 fault branch, if the execution status of a process is "in execution" and the material queue length is greater than zero, the "pause current task execution" instruction is output; if the execution status is "idle waiting", the "prohibit new task issuance" instruction is output. For the Level 3 major shutdown branch, regardless of the process status, the "freeze material queue allocation" instruction is uniformly output. After the decision tree is traversed, a list of locking action instructions for each process to be locked is generated, with each instruction explicitly specifying the target process node ID and the specific locking action type.

[0047] Send the locking action command to the process execution module of the MES system and obtain the locking execution receipt; In this embodiment, the output list of locking action instructions is sent to the process execution module of the MES system. Upon receiving the instructions, the process execution module performs the corresponding locking operation on the specified process node. For the "prohibit new task issuance" instruction, the process execution module suspends the task scheduling interface of that process, refusing to receive any new production task allocation requests. For the "pause current task execution" instruction, the process execution module sends a pause signal to the running equipment, interrupting the currently ongoing workflow. For the "freeze material queue allocation" instruction, the process execution module locks the inlet and outlet ports of the material queue, prohibiting any material pallets from entering or leaving the queue. After all locking operations are completed, the process execution module collects the execution results and generates a locking execution receipt. This receipt includes a locking success identifier to indicate whether the operation was successfully completed; a locking timestamp to record the precise time of operation completion; and a snapshot of the process status after locking, re-querying and recording the latest execution status and material queue length of each process after the locking operation. This receipt is encapsulated into a standard message and fed back to the source end of the linkage control data packet, i.e., the linkage control parsing layer, through the MES system's message bus.

[0048] The lock execution receipt and linkage control data packet are merged to generate a work order track stop event record and perform fault repair.

[0049] In this embodiment, the returned lock execution receipt is merged with the linkage control data packet. The merging operation involves appending the lock success identifier, lock timestamp, and post-lock process status snapshot from the receipt to the original stoppage level, faulty equipment ID, fault code, stoppage timestamp, and affected workstation range fields of the linkage control data packet, forming a complete work order stoppage event record. This record contains end-to-end information from the occurrence of the anomaly to the completion of the MES lock. Subsequently, this record is written as a new entry to the MES event log database. The MES event log database is a persistent database table, with each record arranged in order of stoppage timestamp, supporting fast retrieval by work order number, equipment ID, time range, and other dimensions. After writing is complete, the MES system automatically stops feeding materials to the faulty section, and the upstream board loading and material preparation processes no longer deliver new PCB boards and materials to the faulty workstation. Simultaneously, on-site maintenance personnel locate the faulty equipment based on the fault code and equipment ID and perform fault repair operations. Throughout the entire process, all operations are fully recorded, forming a traceable stoppage event archive.

[0050] Furthermore, the generation of vehicle diversion scheduling instructions in step S5 includes the following steps: Real-time acquisition of position and timing data of all vehicles on the track and capacity status of the buffer track area; generating spatiotemporal snapshots of vehicle status and buffer area capacity status. In this embodiment, real-time operational data of all vehicles on the track is retrieved from the MES system, and five key fields are extracted for each vehicle. The current coordinate field records the vehicle's precise position in the track coordinate system, represented by horizontal and vertical coordinate values. The moving speed vector field contains the vehicle's current speed and direction angle. The target workstation ID field indicates the destination workstation number assigned to the vehicle. The dwell time field records the cumulative time the vehicle has remained stationary at its current location. The vehicle type identifier field distinguishes whether the vehicle is a fully loaded vehicle carrying PCB boards or an empty vehicle returning empty. These five fields for all vehicles are aligned with timestamps to generate a spatiotemporal snapshot of the vehicle's status. Simultaneously, a buffer zone status query command is sent to the track PLC to read the current occupancy rate and remaining capacity of each buffer track area along the track. The occupancy rate indicates the proportion of the buffer area occupied by vehicles, and the remaining capacity indicates how many more vehicles can be accommodated. This buffer zone status data is also aligned with timestamps to generate a buffer zone capacity status snapshot. The two snapshots together form the complete situational awareness data foundation for the current moment of the rail logistics system.

[0051] Based on the spatiotemporal snapshot of the vehicle's state, predict the future trajectory of each vehicle and generate a set of vehicle future trajectory predictions. In this embodiment, trajectory prediction analysis is performed on each vehicle based on a spatiotemporal snapshot of its state. First, the straight-line distance between the vehicle's current position and the coordinates of the fault location is calculated to obtain the Euclidean distance. Then, along the permissible paths in the track topology, the shortest path length from the current position to the fault location is calculated to obtain the Manhattan distance. The estimated arrival time is obtained by dividing the Manhattan distance by the magnitude of the vehicle's current velocity vector. Subsequently, the vehicle's current position coordinates, velocity vector, and information such as the location of junctions and curvatures in the track topology are input into a Kalman filter. The Kalman filter first uses the vehicle's motion model to predict its state, estimating the vehicle's position and velocity in the next time step. Then, when the actual position measurement of the vehicle arrives, the filter calculates the error between the predicted and measured values ​​and corrects the predicted value using the Kalman gain to obtain the optimal estimate. This prediction-correction process is repeated, continuously pushing forward to predict the vehicle's position trajectory sequence over several future time steps. Simultaneously, when encountering track branch points, the filter splits into multiple trajectory branches, each corresponding to a different path direction. The predicted trajectories of all vehicles and their branches are aggregated to generate a set of predicted future vehicle trajectories.

[0052] Input the vehicle future trajectory prediction set and the buffer capacity status snapshot into the conflict avoidance network, and plan one or more alternative paths for each vehicle whose predicted trajectory is suspected of conflicting with the faulty workstation, generating a candidate diversion path set; In this embodiment, the predicted future trajectories of vehicles and the snapshot of buffer zone capacity status are input into the conflict avoidance network. This network uses the predicted future trajectories as its underlying prior, clearly identifying which vehicles may spatially conflict with faulty workstations within future time windows. The snapshot of buffer zone capacity status serves as a resource constraint, specifying how many vehicles each buffer zone can still accommodate. Internally, the conflict avoidance network runs a graph search process based on the A* algorithm. For each vehicle whose predicted trajectory conflicts with a faulty workstation, a path search is performed on the track topology graph, starting from the vehicle's current position and ending at the entrance point of each buffer track zone. During the search, the heuristic function uses the Manhattan distance from the current position to the entrance point, and the cost function comprehensively considers path length and travel time. The search algorithm finds one or more alternative paths to different buffer zones for each conflicting vehicle. For each alternative path, its path length and travel time are calculated, and it is compared pairwise with alternative paths of other conflicting vehicles. The proportion of overlapping path length to the total path length is calculated to obtain the overlap degree. All alternative paths and their attribute parameters for all conflicting vehicles are aggregated to generate a set of candidate diversion paths.

[0053] Based on the candidate diversion path set, the optimal diversion scheme is determined through a multi-objective optimization model, and the optimal allocation matrix is ​​generated. In this embodiment, a multi-objective optimization model is constructed based on the candidate diversion path set. The model has two optimization objectives: the first is to minimize the sum of the total travel times of all conflicting vehicles, and the second is to minimize the variance of the occupancy rate of each buffer zone, i.e., to maximize the balanced utilization of the buffer zones. The input variables of the model are the path length, travel time, overlap, and remaining capacity of the target buffer zone for each path in the candidate diversion path set. The model is solved using a multi-objective evolutionary optimization method based on a genetic algorithm. A population is initialized, where each individual represents a scheme for allocating conflicting vehicles to various buffer zones. Each individual is represented by an allocation matrix, where rows correspond to conflicting vehicles, columns correspond to buffer zones, and matrix elements are 0 or 1, indicating whether the vehicle is allocated to that buffer zone. The fitness of each individual is determined by the values ​​of both objective functions. The population is iteratively evolved through selection, crossover, and mutation operations, retaining non-dominated solutions in each generation. After multiple generations of evolution, a Pareto front solution set is obtained, where each solution is non-dominated, meaning that it cannot improve on any objective without harming the other objective. Select a compromise solution from the Pareto front solution set as the optimal diversion scheme and generate the corresponding optimal allocation matrix.

[0054] The optimal allocation matrix is ​​converted into vehicle diversion scheduling instructions and sent to the track execution control module for execution.

[0055] In this embodiment, the optimal allocation matrix is ​​converted into a set of vehicle diversion scheduling instructions executable by the track PLC. For each vehicle marked as needing diversion in the matrix, a scheduling instruction is generated. The instruction contains a unique ID number for the vehicle for precise addressing. It contains the target point coordinates, i.e., the precise coordinates of the entrance to the buffer track area that the vehicle should go to. It contains a speed limit, specifying the maximum speed of the vehicle during the diversion process to ensure safety. It contains a parking point number, specifying the specific parking space number that the vehicle should park at after entering the buffer area. All instructions are sorted by priority and sent to the track execution control module via the industrial Ethernet protocol. After receiving the instructions, the track execution control module parses the control parameters of each instruction, sends a steering control signal to the corresponding vehicle to drive the vehicle to change its direction of travel at the junction; sends a deceleration control signal to reduce the vehicle's speed when approaching the target point; and sends a parking control signal to stop the vehicle at the designated parking space. After all diversion actions are completed, the track execution control module collects the final docking position and execution status of each vehicle, generates a diversion execution status report, and feeds it back to the MES system through the message bus to complete the vehicle diversion scheduling closed loop.

[0056] Furthermore, the generation of the recovery request instruction in step S6 includes the following steps: When the SMT line equipment feeds back a normal operation signal, extract the equipment recovery state vector; In this embodiment, after the SMT line equipment completes fault repair and restarts, the equipment sends a normal operation signal to the linkage control parsing layer through its communication interface. Upon receiving this signal, the linkage control parsing layer extracts four key pieces of information from fixed field positions in the signal frame. The first field is the device ID, a unique device identifier used to identify which device sent the recovery signal. The second field is the device status code, which contains the current operating status codes of each subsystem of the equipment, such as servo drive status, pneumatic system pressure value, and communication link connectivity status. The third field is the self-test pass flag, a Boolean value indicating whether all self-test procedures executed after power-on have passed; true if all pass, false if any self-test fails. The fourth field is the restart timestamp, recording the precise moment when the equipment completed restarting and entered a stable operating state. The values ​​of these four fields are assembled according to a fixed data structure to form a four-dimensional device recovery status vector. This vector fully describes the comprehensive status information of the equipment at the recovery time, providing input data for subsequent recovery consistency verification.

[0057] The state deviation index is calculated by comparing the device recovery state vector with the historical state baseline vector of the SMT linear device during the historical normal operation cycle. In this embodiment, the state baseline vector of the SMT line equipment during its historical normal operation cycle is retrieved from the historical database. This baseline vector is a reference vector obtained by statistically averaging data such as equipment status codes, self-test pass flags, and restart timestamps from multiple past normal operations, representing the typical state distribution of the equipment during healthy operation. The equipment recovery state vector is compared with the state baseline vector to calculate the Mahalanobis distance between them. The Mahalanobis distance calculation process is as follows: First, the difference vector between the equipment recovery state vector and the state baseline vector is calculated. Then, the covariance matrix of all state vectors during the historical normal operation cycle is obtained, which describes the correlation and dispersion between each state dimension. Next, the difference vector is transposed and multiplied by the inverse of the covariance matrix, then multiplied by the difference vector itself, and finally the square root of the product is taken to obtain the Mahalanobis distance value. This value is the state deviation index. The advantage of Mahalanobis distance is that it eliminates the interference of dimensional differences and correlations between state dimensions, and can comprehensively reflect the deviation between the equipment recovery state and the historical healthy state.

[0058] If the state deviation index is less than the preset threshold, the device state is determined to be consistent and a verification pass flag is generated; otherwise, a verification failure flag is generated. In this embodiment, a preset consistency threshold is set. This threshold is a fixed value representing the upper limit of acceptable deviation in the device's recovery state. The calculated state deviation index is compared with this preset consistency threshold. If the state deviation index is less than the preset consistency threshold, it is determined that the current recovery state of the device is not statistically significantly different from the historical health state, and the device has recovered to normal operating conditions. At this time, a verification pass flag is generated, which is a Boolean true value. If the state deviation index is greater than or equal to the preset consistency threshold, it is determined that the current recovery state of the device is significantly different from the historical health state, and the device has not fully recovered to normal operating conditions. At this time, a verification failure flag is generated, which is a Boolean false value. In the case of verification failure, the differences between the device recovery state vector and the state baseline vector in each dimension are further analyzed. The dimension with the largest absolute value of the difference is identified, and the name of this dimension is recorded, such as "pneumatic system pressure value". At the same time, the deviation direction of this dimension is recorded, i.e., whether the actual value is higher or lower than the baseline value. The verification failure flag, the name of the dimension with the largest deviation, and the deviation direction are packaged and output together.

[0059] Based on the verification pass or failure flag, combined with the track stop level and fault code in the linkage control data packet, a recovery request instruction is generated and uploaded to the MES system, and then the system waits for the MES system's permission verification and track reset instruction to return.

[0060] In this embodiment, a recovery request instruction is constructed based on the verification pass or failure identifier, combined with the track stop level and fault code from the linkage control data packet in the previous steps. This instruction contains five fields. The device ID field directly reuses the extracted device ID. The recovery timestamp field is filled with the restart timestamp. The verification result field is filled with the verification pass or failure identifier. The suggested recovery action field is filled according to the verification result: if the verification passes, fill in "Allow production to resume"; if the verification fails, fill in "Manual intervention required for review" and attach the name of the state dimension with the largest deviation and the direction of deviation. The track stop level and fault code fields are directly copied from the linkage control data packet. This recovery request instruction is encapsulated according to the message format of the industrial Ethernet protocol and uploaded to the recovery request service interface of the MES system via a TCP / IP socket connection. After the upload is completed, it enters a waiting state and continuously listens for the response message returned by the MES system. After receiving the recovery request instruction, the MES system will initiate the permission verification process to verify whether the permission level of the requesting operator meets the requirements for unlocking the track stop level. After the permission verification is passed, the MES system will generate a track reset instruction and send it back through the same communication channel. After receiving the track reset command, the linkage control parsing layer forwards it to the track execution control module to execute the track operation resumption action.

[0061] Furthermore, refer to Figure 3-4 , Figure 3 This is a block diagram of the module architecture of the graded track stop control system based on abnormal linkage of SMT line equipment in the embodiment. Figure 4 This is a flowchart illustrating the SMT anomaly acquisition and graded track stop control in the embodiment. The present invention also provides a graded track stop control system based on SMT line equipment anomaly linkage, used to execute the graded track stop control method based on SMT line equipment anomaly linkage as described above. It includes an MES main control server, a signal acquisition module, a fault level determination module, a track locking control module, a personnel access control module, an automatic start / stop / restart module, and a connecting platform track solenoid valve. The MES main control server communicates with each module and executes control logic. The signal acquisition module acquires multi-dimensional operating data of the SMT linear equipment in real time, generating a raw dataset of equipment anomalies. The fault level determination module determines the anomaly level based on the raw dataset and outputs a graded anomaly identifier. The track locking control module triggers the corresponding level of track stop control command according to the graded anomaly identifier and packages a linkage control data packet to perform fault repair. The personnel access control module verifies operator permissions after fault repair. The automatic start / stop / restart module performs track reset and resumption scheduling after permission verification. The connecting platform track solenoid valve performs physical-level track locking or release actions.

[0062] In this embodiment, the MES master control server serves as the core of the system. It communicates with the signal acquisition module, fault level determination module, track locking control module, personnel access control module, and automatic start-stop and restart module via the industrial Ethernet protocol to achieve real-time data acquisition, anomaly determination, track stop control, access verification, and restart scheduling. Specifically, the MES main control server receives the raw dataset of equipment anomalies uploaded by the signal acquisition module and forwards it to the fault level determination module for anomaly level determination; it receives the graded anomaly identifier output by the fault level determination module and forwards it to the track locking control module to trigger the track stop control command; it receives the permission verification result returned by the personnel permission management module and controls the automatic start-stop and restart module to perform the restart operation; at the same time, the MES main control server packages information such as track stop level, faulty equipment ID, fault code, track stop timestamp, and affected workstation range into a linkage control data packet and sends it to the linkage control parsing layer via the industrial Ethernet protocol; the signal acquisition module connects to the photoelectric sensors, cardboard sensors, and overload detection sensors of the SMT production line via the industrial Ethernet protocol, collects data from each sensor in real time, and uploads the collected data to the MES main control server; specifically, the signal acquisition module collects sensor data every 100ms, including the number of ejected materials, material balance, PCB transmission status, SPI / AOI detection yield, furnace temperature data, and equipment alarm codes, generates the raw dataset of equipment anomalies, and transmits it via TCP or IP. The socket connection uploads data to the MES main control server; the fault level determination module extracts abnormal feature vectors from the original dataset of device anomalies, matches them with feature sequences in the historical fault mode library, calculates the Mahalanobis distance between the feature vectors and the historical baseline, and dynamically maps the anomaly level by combining it with a preset anomaly threshold matrix, outputting the corresponding graded anomaly identifier; specifically, the fault level determination module first extracts abnormal feature vectors, then retrieves feature sequences from the historical fault mode library that match the device ID and historical operating cycle, calculates the Mahalanobis distance between the two, and obtains the state deviation index; if the state deviation index is less than the preset consistency threshold, it is determined as a level one warning; if it is greater than or equal to the threshold but less than the level two threshold, it is determined as a level two fault; if it is greater than or equal to the level two threshold, it is determined as a level two fault; if it is greater than or equal to the level two threshold, it is determined as a level two fault. If the value is not specified, it is determined to be a Level 3 major shutdown. The track locking control module issues a stop command to the track execution control module according to the level of anomaly. At the same time, it packages the data into a linkage control data packet and sends it to the linkage control parsing layer via the industrial Ethernet protocol. Specifically, the track locking control module generates the corresponding stop control command according to the level of anomaly: for a Level 1 warning, only the two track points before and after the faulty workstation are locked; for a Level 2 fault, the section where the faulty workstation is located is locked (such as the front section A or the rear section C); for a Level 3 major shutdown, the entire track (section A+B) is locked. At the same time, the stop level, faulty equipment ID, fault code, stop timestamp, and affected workstation range are packaged into a linkage control data packet and sent to the linkage control parsing layer via the industrial Ethernet protocol.The personnel access control module receives a recovery request command, retrieves the current operator's access level from the operator management database, compares it with the minimum unlocking access level corresponding to the track stop level, and generates an access pass flag if the access is valid. Specifically, after receiving the recovery request command, the personnel access control module parses the equipment ID and track stop level, queries the minimum unlocking access level corresponding to that track stop level (e.g., Level 1 warning requires operator access, Level 2 fault requires technician access, and Level 3 major shutdown requires engineer access); then it retrieves the current requesting operator's access level from the operator management database and compares it; if the access is insufficient, unlocking is rejected and a prompt is displayed; if the access is valid, an access pass flag is generated; after receiving the access pass flag, the automatic start / stop / restart module sends a track reset command to the track execution control module. When a vehicle diversion scheduling command is triggered, vehicles in the buffer track area are prioritized for scheduling to normal workstations. Specifically, after receiving the permission pass flag, the automatic start-stop-return module sends a track reset command to the track execution control module to release the track point lock. Simultaneously, it retrieves the vehicle diversion scheduling record before the fault from the MES event log library, prioritizing the scheduling of vehicles in the buffer track area to normal workstations and restoring automatic scheduling and transfer. When the connecting platform track solenoid valve receives a stop control command or a track reset command, it performs physical track locking or releasing actions. Specifically, when the connecting platform track solenoid valve receives a stop control command, it closes the track entrance / exit, prohibiting vehicle entry and exit; when it receives a track reset command, it opens the track entrance / exit, allowing normal vehicle passage. The control signal for the solenoid valve is sent by the track execution control module through the PLC I / O port, with a response time of less than 50ms.

[0063] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A graded track stop control method based on abnormal linkage of SMT line equipment, characterized in that, Includes the following steps: Step S1: Collect multi-dimensional operating data of SMT line equipment in real time, including the number of parts thrown, material balance, PCB transfer status, SPI / AOI inspection yield, oven temperature data and equipment alarm codes, and generate raw dataset of equipment anomalies. Step S2: Extract the abnormal feature vector of each device based on the original dataset of device abnormalities, combine it with the preset abnormal threshold matrix and the historical fault mode library, determine the abnormal level through dynamic weight mapping, and generate the graded abnormality identifiers corresponding to Level 1 warning, Level 2 fault or Level 3 major shutdown. Step S3: Trigger the corresponding level of rail stop control command according to the graded anomaly identifier, and package the rail stop level, fault equipment ID, fault code, rail stop timestamp, and affected workstation range into a linkage control data package; Step S4: Send the linkage control data packet to the MES system, which will lock the upstream process and material preparation task, stop feeding materials to the faulty section, generate a work order track stop event record and perform fault repair. Step S5: During the fault repair process at the stop-rail, the MES system automatically dispatches the on-the-way vehicles to the buffer track area and generates vehicle diversion and dispatch instructions to avoid workstation blockage and track jamming. Step S6: After the equipment fault is repaired, the linkage control parsing layer receives and verifies the equipment status, generates a recovery request instruction, and uploads it to the MES system; Step S7: The MES system requests the resumption of operation based on the recovery request command and verifies the operator's permissions. After the permissions are valid, the system issues a track reset command to restore automatic scheduling and transfer, and records a complete traceable quality data chain.

2. The graded track stop control method based on abnormal linkage of SMT line equipment according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Extract the abnormal feature vectors corresponding to each SMT line equipment from the original dataset of equipment abnormalities, including the slope of the ejection frequency, the material balance decay rate, the duration of the transmission blockage, the standard deviation of the yield deviation, the temperature fluctuation amplitude, and the frequency of alarm codes. Step S22: Call the historical frequent fault feature set matched in the historical fault mode library to construct the time series background feature vector; fuse the abnormal feature vector with the time series background feature vector and standardize the process to generate a composite abnormal index; Step S23: Based on the composite anomaly index, the anomaly threshold matrix is ​​corrected in real time to generate an adaptive threshold boundary matrix; Step S24: Perform causal correlation analysis on the abnormal feature vectors based on the adaptive threshold boundary matrix to identify secondary anomalies and primary anomalies, and calculate the independent contribution of each abnormal feature accordingly. Step S25: Input the independent contribution of each abnormal feature into the dynamic weight mapping model, where the equipment topology diagram is used as the structural prior and the fault propagation path in the historical fault mode library is used as the training sample, and output the global influence weight of each abnormal feature in the current equipment interconnection state. Step S26: Use global influence weights to perform weighted aggregation of composite anomaly indicators to obtain a comprehensive anomaly score, and compare it with the adaptive threshold boundary matrix to automatically match and output the graded anomaly identifiers corresponding to Level 1 warning, Level 2 fault, or Level 3 major shutdown.

3. The graded track stop control method based on abnormal linkage of SMT line equipment according to claim 2, characterized in that, The process of fusing and standardizing the abnormal feature vector with the temporal background feature vector in step S22 includes the following steps: Extract the current-time anomaly intensity vector from the anomaly feature vector, and extract the historical cumulative risk vector from the temporal background feature vector; Based on the cross-modal attention mechanism, feature-level cross-enhancement is performed on the current anomaly intensity vector and the historical cumulative risk vector to generate a fused enhanced feature tensor; Centering and scaling corrections are performed based on the enhanced feature tensor to eliminate dimensional differences and distribution shifts between different modal features, generating a standardized fused feature vector. The standardized fusion feature vector is input into the nonlinear activation mapping layer, and the composite anomaly index is output.

4. The graded track stop control method based on abnormal linkage of SMT line equipment according to claim 2, characterized in that, Step S24, which describes causal correlation analysis of abnormal feature vectors based on an adaptive threshold boundary matrix, includes the following steps: Candidate abnormal nodes are selected from abnormal feature vectors based on an adaptive threshold boundary matrix. Candidate abnormal nodes are input into the causal discovery network. The physical topology of the device is used as a priori constraint, and the known causal chains in the historical fault mode library are used as supervision signals. The directed causal relationship between each abnormal feature is automatically learned through iterative pruning, and the initial causal graph is output. Counterfactual intervention is performed on each candidate anomalous node based on the initial causal graph to determine whether it is a root cause node or a transitive node; All candidate abnormal nodes identified as root cause nodes are mapped back to their corresponding physical parameter domains to establish a spatiotemporal alignment relationship between root cause nodes and physical phenomena. Based on the spatiotemporal alignment relationship, root cause nodes that have a strong causal relationship with the decrease in material balance are classified as secondary abnormalities caused by insufficient material supply, while root cause nodes that have no direct causal relationship with material balance but trigger sudden changes in material ejection are classified as primary abnormalities caused by mechanical wear of equipment.

5. The graded track stop control method based on abnormal linkage of SMT line equipment according to claim 4, characterized in that, The process of establishing the spatiotemporal alignment between root nodes and physical phenomena includes the following steps: Extract all candidate abnormal nodes identified as root cause nodes and their corresponding abnormal dimension identifiers. At the same time, retrieve the original physical parameter time series curves corresponding to the abnormal dimension identifiers from the original dataset of equipment abnormalities, including the material balance decay curve, the cumulative growth curve of the number of discarded materials, the PCB transmission blockage flag sequence, the detection yield deviation curve, the furnace temperature deviation curve, and the equipment alarm frequency histogram. Analyze the time-series curves of the original physical parameters and extract the timestamps of key events, including candidate time points for material depletion, candidate time points for material ejection, plate blockage start time points, yield degradation start points, temperature instability start points, and alarm density start points. The activation time window of the root node in the initial causal graph and the key event timestamps are input into the spatiotemporal alignment network. By adopting a dynamic time warping algorithm, the activation time window is used as the starting point of the warping path, and each timestamp in the key event timestamps is used as a candidate endpoint of the warping path. The optimal time offset between each root node and its corresponding physical phenomenon is calculated by minimizing the cumulative path distance. At the same time, the ratio of the cumulative path distance to the path length corresponding to the optimal time offset is calculated as the normalized alignment cost, generating a spatiotemporal alignment offset vector and a normalized alignment cost matrix. The activation time window of the root cause node is shifted and corrected based on the spatiotemporal alignment offset vector and the normalized alignment cost matrix. This ensures that the corrected activation time window has maximum overlap with at least one of the following key event time points: material depletion candidate time point, material ejection outbreak candidate time point, plate blockage start time point, yield degradation start point, temperature instability start point, and alarm density start point. This establishes a spatiotemporal alignment relationship between the root cause node and the physical phenomenon.

6. The graded track stop control method based on abnormal linkage of SMT line equipment according to claim 4, characterized in that, Step S25, which involves inputting the independent contribution of each anomalous feature into the dynamic weight mapping model, includes the following steps: Extract the independent contribution of each abnormal feature to generate a labeled node contribution feature set; Based on the physical layout and material flow path of SMT linear equipment, construct the equipment topology diagram; The node contribution feature set and the device topology graph are input into the dynamic weight mapping model. The node contribution feature is used as the initial embedding vector, the adjacency matrix of the device topology graph is used as the message passing skeleton, and the influence coefficient of each node on its neighboring nodes is calculated through the attention mechanism to generate a dynamically updated node embedding representation. Based on node embedding representation, the dynamic weight mapping model propagates layer by layer along the edges of the device topology graph. During each layer of propagation, the contribution features of neighboring nodes are weighted and aggregated using the influence coefficient. At the same time, residual connections are introduced to retain the node state of the previous layer, generating node features after propagation. The propagated node features are aggregated through a global pooling operation, and the global impact weight of each abnormal feature in the current device interconnection state is output.

7. The graded track stop control method based on abnormal linkage of SMT line equipment according to claim 1, characterized in that, The step S4 of sending the linkage control data packet to the MES system includes the following steps: Extract the track stop level, faulty equipment ID, fault code, track stop timestamp, and affected workstation range from the linkage control data packet to generate a structured parsed field set; Based on the structured parsed field set, the process topology mapping table in the MES system is called to determine the snapshot of the process status to be locked; Input the snapshot of the status of the process to be locked and the track stop level into the locking strategy decision tree, and output the locking action instructions for each process to be locked, including prohibiting the issuance of new tasks, pausing the execution of the current task, and freezing the material queue allocation; Send the locking action command to the process execution module of the MES system and obtain the locking execution receipt; The lock execution receipt and linkage control data packet are merged to generate a work order track stop event record and perform fault repair.

8. The graded track stop control method based on abnormal linkage of SMT line equipment according to claim 1, characterized in that, The generation of vehicle diversion scheduling instructions in step S5 includes the following steps: Real-time acquisition of position and timing data of all vehicles on the track and capacity status of the buffer track area; generating spatiotemporal snapshots of vehicle status and buffer area capacity status. Based on the spatiotemporal snapshot of the vehicle's state, predict the future trajectory of each vehicle and generate a set of vehicle future trajectory predictions. Input the vehicle future trajectory prediction set and the buffer capacity status snapshot into the conflict avoidance network, and plan one or more alternative paths for each vehicle whose predicted trajectory is suspected of conflicting with the faulty workstation, generating a candidate diversion path set; Based on the candidate diversion path set, the optimal diversion scheme is determined through a multi-objective optimization model, and the optimal allocation matrix is ​​generated. The optimal allocation matrix is ​​converted into vehicle diversion scheduling instructions and sent to the track execution control module for execution.

9. The graded track stop control method based on abnormal linkage of SMT line equipment according to claim 1, characterized in that, The step S6 of generating the recovery request instruction includes the following steps: When the SMT line equipment feeds back a normal operation signal, extract the equipment recovery state vector; The state deviation index is calculated by comparing the device recovery state vector with the historical state baseline vector of the SMT linear device during the historical normal operation cycle. If the state deviation index is less than the preset threshold, the device state is determined to be consistent and a verification pass flag is generated; otherwise, a verification failure flag is generated. Based on the verification pass or failure flag, combined with the track stop level and fault code in the linkage control data packet, a recovery request instruction is generated and uploaded to the MES system, and then the system waits for the MES system's permission verification and track reset instruction to return.

10. A graded track stop control system based on abnormal linkage of SMT line equipment, used to execute the graded track stop control method based on abnormal linkage of SMT line equipment as described in any one of claims 1-9, characterized in that, The system includes an MES main control server, a signal acquisition module, a fault level determination module, a track locking control module, a personnel access control module, an automatic start-stop-restart module, and a connecting platform track solenoid valve. The MES main control server communicates with each module and executes control logic. The signal acquisition module collects multi-dimensional operating data from the SMT linear equipment in real time, generating a raw dataset of equipment anomalies. The fault level determination module determines the anomaly level based on the raw dataset and outputs a graded anomaly identifier. The track locking control module triggers corresponding level-based track stop control commands based on the graded anomaly identifier and packages a linkage control data packet to perform fault repair. The personnel access control module verifies operator permissions after fault repair. The automatic start-stop-restart module performs track reset and resumption scheduling after permission verification. The connecting platform track solenoid valve performs physical-level track locking or release actions.

Images