Circuit board production whole-process material tracing and wrong material prevention intelligent management method

Abstract

The invention provides a circuit board production whole-process material tracing and wrong material prevention intelligent management method, which comprises the steps of comprehensively acquiring multi-dimensional input of production line materials, processes, equipment, personnel, environments and the like, inputting results of the production line materials, the processes, the equipment, the personnel, the environments and the like and environment characteristics into a lightweight dynamic arbiter model through rule engine quantitative evaluation and AI model multi-modal semantic judgment, self-adaptive weight distribution between the rule path and the AI path is realized, a comprehensive abnormal score is generated, and grading early warning is triggered; model parameters are continuously updated through online reinforcement learning, continuous optimization of an arbitration strategy is achieved in combination with manual feedback, the accuracy and timeliness of anomaly detection are improved, the adaptability and stability of a system to complex production line working conditions are enhanced, and digital and intelligent upgrading of a manufacturing link is assisted.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and industrial quality management technology, and in particular to a method for intelligent management of material traceability and error prevention throughout the entire circuit board production process. Background Technology

[0002] With the rapid development of intelligent manufacturing and industrial automation in recent years, intelligent management of material traceability and error prevention throughout the entire circuit board production process has gradually introduced expert systems centered on rule engines and anomaly detection solutions based on AI models. In mainstream technical solutions, enterprises often utilize rule engine mechanisms such as BOM (Bill of Materials) consistency verification, batch traceability, production process compliance verification, and operation log monitoring to analyze and send alarms for anomalies such as incorrect, missing, or mixed materials that may occur during production. Simultaneously, artificial intelligence technologies such as image recognition and natural language processing are also widely used in automatic material label recognition, work order text parsing, and operation process matching to uncover hidden defects that are difficult to cover by existing rules. Existing industrial intelligent quality control solutions often employ a collaborative approach, with static rules and pre-trained AI models working together. Some solutions support simple weighted fusion of the outputs from both, for example, determining the final warning result based on pre-set weights or priorities. While these methods can be effective in scenarios with stable processes and clear anomaly patterns, static fusion mechanisms struggle to cope with diverse and complex anomaly scenarios when faced with dynamic manufacturing environments characterized by significant fluctuations, complex and changing product categories, and varying equipment status and personnel experience. In the fields of intelligent manufacturing and industrial process monitoring, some technical literature has proposed integrating AI with business rules to improve the intelligence and flexibility of anomaly detection. However, mainstream solutions are still limited to the following typical scenarios: The rule engine can only handle clearly defined logical relationships. It is slow to react to noise disturbances, atypical defects or behavioral abnormalities, and cannot dynamically respond to fluctuations in the production line's operating status. While AI models excel in semantic understanding of data such as images and text, they have a high risk of misjudgment and missed judgment in extreme working conditions, data distribution drift, and scenarios where human visual observation is not possible. Furthermore, the flexibility of the model is limited once the parameters are fixed. When rule engines and AI models collaborate, they often adopt static weighting and fixed priority judgment, failing to automatically adjust the contribution weights of both parties according to the actual runtime context. This makes it difficult to achieve "on-demand fusion," which is not only difficult to improve the judgment accuracy, but also difficult to ensure the overall robustness of the system. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, this invention provides a method for intelligent management of material traceability and error prevention throughout the entire circuit board production process.

[0004] The technical solution of this invention is implemented as follows: a method for intelligent management of material traceability and error prevention throughout the entire circuit board production process, comprising: S1: Obtain the material status change event trigger signal during the circuit board production process, and collect the material identification code, work order number, process route parameters, equipment status data and operator identity information as the initial input dataset; S2: Perform structured feature extraction on the initial input dataset to generate a set of rule judgment parameters, including three quantitative indicators: BOM matching degree, batch shelf life remaining amount, and process route consistency deviation amount, and calculate the rule confidence score of each indicator; S3: Based on convolutional neural networks and attention mechanisms, an AI recognition model is built. The input is material images, work order texts and operation log sequence data. The output is an unstructured prediction result set containing material type mismatch probability, process parameter deviation, and label integrity index. S4: Collect production line operating context environment parameters, including equipment sensor stability index, shift personnel experience level, process cycle time fluctuation rate, and the most recent manual review time interval, construct a four-dimensional environment feature vector and perform normalization processing; S5: Construct a lightweight dynamic arbitrator model, taking the rule confidence score, AI prediction result set and environmental feature vector as input. After feature cross and normalization processing, a gating mechanism is used to generate rule path activation weights and AI path activation weights. The sum of the two weights equals 1. S6: Based on the rule path activation weight and AI path activation weight, the rule judgment parameter set and AI prediction result set are weighted and fused to generate a comprehensive anomaly score value. When the value exceeds the preset threshold, a graded early warning signal is triggered. S7: Record the complete judgment process log, including the original input data, the output results of each module, the weight allocation parameters and the manual confirmation feedback, and build a training dataset for online reinforcement learning updates of the dynamic arbitrator model; S8: Based on the updated dynamic arbitrator model parameters after reinforcement learning, adjust the weight allocation strategy for subsequent material state change events to continuously optimize the accuracy of anomaly warning.

[0005] The intelligent management method for material traceability and error prevention in the entire circuit board manufacturing process provided by this invention has the following beneficial effects: (1) This invention achieves context-aware adaptive fusion of rule engine and AI model by constructing a lightweight dynamic arbitrator, which significantly improves the accuracy and robustness of material state anomaly judgment. It introduces an environment vector encoded based on operating context environment parameters (such as equipment stability index, personnel experience level, manual review interval, etc.) and combines rule confidence and AI classification probability. The activation weights of rule path and AI path are generated in real time by a trainable gating network, realizing dynamic adjustment of the contribution of dual engines. In the environment of high-frequency noise or low-quality sensor data, the AI ​​path weight is automatically suppressed to avoid the risk of model misjudgment; while when there are implicit process deviations or unstructured semantic clues (such as fuzzy labels and ambiguous text descriptions), the influence of AI semantic understanding ability is enhanced, and the sensitivity of identifying potential material error risks is improved. This mechanism effectively overcomes the contradiction between the lack of flexibility of a single rule system and the poor interpretability of a pure AI model, and improves the comprehensive judgment ability in complex industrial scenarios while ensuring decision transparency. (2) This invention adopts an online reinforcement learning-driven gating parameter update mechanism to form a closed-loop iterative system of "judgment-feedback-optimization", which greatly enhances the method's adaptability and long-term stability. Unlike static fusion models that rely on offline training and cannot evolve with the environment, the dynamic arbitrator in this invention continuously receives real feedback signals from manual confirmation, which are then used as reward signals to input into the reinforcement learning strategy to adjust the internal parameters of the gating network. It can also continuously optimize the weight allocation logic based on the actual operation and maintenance effect. For example, in the stage where a certain type of equipment frequently generates false triggers, the system automatically identifies the pattern through feedback accumulation and reduces the AI ​​weight in the relevant period. When the probability of non-standard operation increases due to the arrival of new employees, it can proactively increase the reliance on rule consistency. This online evolution capability based on real production feedback enables the intelligent verification module to have good generalization and anti-drift characteristics. It can maintain stable performance under real disturbances such as production line upgrades, process changes, or personnel rotation, avoiding the problem of increased false negative rate caused by model degradation. At the same time, it reduces the reliance on frequent manual parameter tuning and expert experience intervention, effectively improving the system's autonomous operation and maintenance level and deployment efficiency. Attached Figure Description

[0006] Figure 1 The flowchart shows the intelligent management method for material traceability and error prevention in the entire circuit board production process of the present invention. Figure 2 This is a sub-flowchart of the intelligent management method for material traceability and error prevention in the entire circuit board production process of the present invention; Figure 3 This is another sub-flowchart of the intelligent management method for material traceability and error prevention in the entire circuit board production process of the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides a method for intelligent management of material traceability and error prevention throughout the entire circuit board manufacturing process, specifically including: S1: Obtain the material status change event trigger signal during the circuit board production process, and collect the material identification code, work order number, process route parameters, equipment status data and operator identity information as the initial input dataset; S2: Perform structured feature extraction on the initial input dataset to generate a set of rule judgment parameters, including three quantitative indicators: BOM matching degree, batch shelf life remaining amount, and process route consistency deviation amount, and calculate the rule confidence score of each indicator; S3: Based on convolutional neural networks and attention mechanisms, an AI recognition model is built. The input is material images, work order texts and operation log sequence data. The output is an unstructured prediction result set containing material type mismatch probability, process parameter deviation, and label integrity index. S4: Collect production line operating context environment parameters, including equipment sensor stability index, shift personnel experience level, process cycle time fluctuation rate, and the most recent manual review time interval, construct a four-dimensional environment feature vector and perform normalization processing; S5: Construct a lightweight dynamic arbitrator model, taking the rule confidence score, AI prediction result set and environmental feature vector as input. After feature cross and normalization processing, a gating mechanism is used to generate rule path activation weights and AI path activation weights. The sum of the two weights equals 1. S6: Based on the rule path activation weight and AI path activation weight, the rule judgment parameter set and AI prediction result set are weighted and fused to generate a comprehensive anomaly score value. When the value exceeds the preset threshold, a graded early warning signal is triggered. S7: Record the complete judgment process log, including the original input data, the output results of each module, the weight allocation parameters and the manual confirmation feedback, and build a training dataset for online reinforcement learning updates of the dynamic arbitrator model; S8: Based on the updated dynamic arbitrator model parameters after reinforcement learning, adjust the weight allocation strategy for subsequent material state change events to continuously optimize the accuracy of anomaly warning.

[0010] Step S1: Obtain the trigger signal for material status change events during circuit board production, and collect material identification codes, work order numbers, process route parameters, equipment status data, and operator identity information as the initial input dataset. Specifically, this includes: S1.1: Based on the real-time event release mechanism in the industrial production line control system, listen to and obtain the material status change event trigger signals caused by the mounting, soldering, testing or handling processes in the circuit board production process. The trigger signals include the timestamp of material loading, process switching, batch replacement or manual intervention operation, which serve as the input conditions for starting the abnormal early warning judgment process to ensure the timeliness and event-driven characteristics of the system response. Based on the real-time event publishing mechanism of the industrial field production line control system, the event subscription mode (parameters: event source interface address, subscription topic list, QoS level) is adopted to realize the continuous monitoring function of trigger signals of placement, welding, inspection and handling processes. A multi-threaded event queue processing algorithm (parameters: number of concurrent threads, queue depth threshold) is adopted to achieve parallel reception of real-time material loading, process switching, batch replacement and manual intervention operation signal streams, and ensure the integrity of the time sequence during signal processing; Furthermore, through the event feature parsing module (parameters: signal type code table, timestamp precision level), the function of identifying the type of received signal and extracting the timestamp is realized, and the timestamp is quantized into a unified format at the nanosecond level to ensure that the time synchronization accuracy meets the requirements of subsequent correlation analysis. Furthermore, an event validity determination algorithm (parameters: noise signal threshold, valid type set) is adopted to filter the trigger signals, eliminate non-production-related noise events, and obtain a valid event set that conforms to the production status change definition; Furthermore, by utilizing the event buffering and synchronization mechanism (parameters: buffer time window size, synchronization triggering conditions), the aggregation of multi-source event streams within the time window and cross-system time synchronization processing are realized, and standardized event data packets containing event type labels and precise timestamps are output. Through the above-mentioned event subscription, parsing, judgment and synchronization processing methods, the original trigger signals in the real-time event publishing mechanism are transformed into structured material status change event data, so as to accurately capture the start conditions of the abnormal early warning judgment process, thereby ensuring the timeliness of the system response and the event-driven characteristics. For example, in a circuit board production line with a monthly output of 20,000 pieces, the event subscription mode is configured with the interface address tcp: / / 192.168.10.20:1883, the subscription topic list is ["SMT / Load","SMT / Switch","SMT / Batch","SMT / Manual"], and the QoS level is 2; the multi-threaded event queue processing algorithm parameters are set to 8 threads concurrently and 1024 queue depth thresholds; the signal type code table in the event feature parsing module includes four categories: placement (0x01), soldering (0x02), detection (0x03), and handling (0x04), and the timestamp precision level is set to nanoseconds, obtained through a built-in high-precision clock. Second-resolution timestamps; the noise threshold in the event validity judgment algorithm is a signal strength below 0.05V, filtering out false triggers caused by equipment debugging or idleness; the buffer time window size is set to 500ms, and the synchronization trigger condition is a time difference not exceeding [a certain value]. Seconds. Under this configuration, the system captures an average of about 4,000 valid material status change events per day. The event type labels and timestamps conform to the unified format specifications, providing high-quality, low-latency input data for the context association processing of S1.2. Field verification shows that the abnormal warning response latency has been significantly reduced to less than 5ms, and the event-driven reliability of the production site has been greatly improved. S1.2: Perform event context association processing on the material status change event trigger signal, and use message queue middleware to synchronously pull the material identification code bound to the current event from the Manufacturing Execution System (MES). The material identification code is a one-dimensional barcode or two-dimensional code based on GS1 standard encoding, which is obtained by scanning with an industrial barcode reader and serves as the input basis for identifying the unique identity of the material actually used, so as to establish a mapping relationship between physical materials and digital systems. The event context association processing is performed on the material status change event trigger signal. An asynchronous retrieval method based on message queue middleware (parameters: unique event ID, timeout, number of retries) is adopted to synchronously obtain the material identification code data stream bound to the current event from the Manufacturing Execution System (MES). Furthermore, by using an encoding and parsing algorithm based on the GS1 standard (parameters: encoding rule version, check bit algorithm type), the validity verification and formatting of the retrieved raw barcode data are achieved, and a set of one-dimensional barcode or two-dimensional barcode information conforming to the standard is obtained. Furthermore, by utilizing the industrial barcode reader hardware interface driver layer (parameters: serial port baud rate, scanning mode, trigger threshold), real-time scanning and acquisition of physical materials on the current production line is achieved, and a barcode original value data packet that matches the record in the MES system is generated. Furthermore, a two-way hash mapping method between barcodes and event IDs (parameters: hash function type, conflict resolution strategy) is adopted to achieve two-way binding between the material identification code obtained by scanning and the event trigger signal, and to establish a unique mapping relationship between physical materials and the digital system; By using unique constraint detection based on relational databases, the binding results of the previous step are transformed into traceable material identification records, achieving consistency and verifiability between production line material data and event information, and ensuring the accuracy of subsequent work order traceability and process compliance assessment. For example, in a single circuit board production batch, the unique ID of the event trigger signal is EVT20240315001, the maximum timeout of the message queue middleware is set to 3 seconds, the number of retries is 2, and the corresponding material identification code recorded in the MES is "6923450654321". The GS1 standard version is set to version 2021, the check digit algorithm type is modulo 10 weighted method, and the parsed barcode structure includes the prefix 690, manufacturer code 23450, product code 654321, and check digit 7. The industrial barcode reader is configured with a serial port baud rate of 115200bps, a continuous trigger scanning mode, a trigger threshold of 20ms, and the QR code data obtained from the on-site scanning is completely consistent with the MES record. The bidirectional hash mapping uses the SHA-256 algorithm and the collision resolution strategy is the chaining method. After generating the hash index, a uniqueness constraint check is performed in the database, which ultimately forms a one-to-one correspondence record between the event ID and the material identification code. This mapping relationship can be directly called in the subsequent work order number tracing steps to achieve accurate association and stable traceability of material identity throughout the entire process. S1.3: Based on the material identification code, perform work order traceability query, call the work order management module of the ERP system through the Enterprise Service Bus (ESB) interface to obtain the work order number associated with the material. The work order number is a structured string containing product model, production batch and planned quantity, which serves as the basic input data for subsequent BOM matching verification and process path comparison, and is used to limit the technical constraint boundary of the current production task. S1.4: Extract the corresponding process route parameters from the MES system according to the work order number. The process route parameters are an ordered set containing process sequence, standard operation time, allowed equipment type and quality control point, which are encapsulated and transmitted in XML format. By parsing the parameters, the standard process path that the current work order should follow is obtained as an input benchmark for evaluating the compliance of the actual process, so as to support the subsequent calculation of process consistency deviation. S1.5: Obtain equipment status data related to the current material operation through a distributed data acquisition gateway. The equipment status data includes the pick-and-place machine operating mode, reflow soldering temperature profile, AOI detection results, and PLC control signals. Use the OPC UA protocol to achieve standardized access to data from multiple brands of equipment. Perform noise reduction filtering and timestamp alignment processing on the equipment status data to generate a snapshot of equipment operation with spatiotemporal consistency, which serves as an input feature reflecting the stability of the current production environment. S1.6: Based on the identity authentication log of the on-site operation terminal, collect the identity information of the operator who performed this material operation. The identity information of the operator includes the employee ID, job permission level and the code of the work group. The validity of the information is verified by LDAP directory service. Combined with the training history database in the human resources system, a quantifiable personnel experience initial evaluation vector is generated as one of the context control factors that affect the subsequent dynamic arbitration weight allocation. S1.7: The material identification code, work order number, process route parameters, equipment status data and operator identity information obtained above are normalized and encoded according to the predefined JSON Schema structure. Null value filling and outlier marking operations are performed to generate an initial input dataset in a unified format. Each field is attached with a source system label and collection timestamp, which serve as the joint input for subsequent structured feature extraction and AI semantic understanding modules to ensure data traceability and processing consistency.

[0011] Step S2: Perform structured feature extraction on the initial input dataset to generate a rule judgment parameter set, including three quantitative indicators: BOM matching degree, batch shelf life remaining, and process route consistency deviation, and calculate the rule confidence score for each indicator. Specifically, this includes: S2.1: Based on the material identification code, work order number and process route parameters, obtain the corresponding standard bill of materials (BOM) from the enterprise resource planning system as the benchmark reference data to establish the expected set of materials to be used under the current work order and obtain the standard BOM dataset; Based on the material identification code, work order number, and process route parameters, the Enterprise Service Bus (ESB) interface is used to call the material management module in the Enterprise Resource Planning (ERP) system (parameter settings: request timeout 3000ms, retries 3, return format XML) to achieve accurate retrieval of the standard bill of materials (BOM) data configured for the current work order; Furthermore, by using the key-value mapping method (parameter: key is the work order number string, value is the BOM record object), an index lookup operation is performed in the ERP system data table to obtain the complete BOM record corresponding to the work order to which the current material identification code belongs, and the original BOM dataset containing material code, specifications, quantity range and quality grade is obtained. Furthermore, an XML parsing algorithm (parameters: DTD verification enabled, node cache limit 100KB) is used to decompose the structured fields of the original BOM dataset and align and convert the material code field with the GS1 standard code parsing rules to obtain a standardized set of BOM entries in a unified encoding format. Furthermore, by using a cross-validation method between process route parameters and BOM item set (parameter: process sequence comparison threshold set to 0.95), the function of verifying whether each material item in the BOM is allowed to be put into the current process node is realized, and a standard BOM subset that meets the process requirements is generated. By using field normalization processing (parameters: the normalization range of numerical fields is set to [0,1], and one-hot encoding is used for categorical fields), the standard BOM subset is transformed into a structured standard BOM dataset that can be directly called by the rule engine, thus achieving a unified input technology effect for this dataset in subsequent BOM matching degree calculation and rule confidence score generation. For example, in a certain type of circuit board production task, the material identification code is "6901234567890", the work order number is "PCBA-202404-A01", and the process route parameter node sequence is ["SMT", "Reflow Soldering", "AOI Inspection"]. By calling the ERP system through the ESB interface, the standard XML format BOM file for this work order is retrieved. The file size is 45KB and contains 24 material entries. Using key-value mapping indexing, the 8th material entry in the BOM record with material code "GS16901234567890" is located. After XML parsing and applying GS1 rules for unified encoding, the material specifications are obtained: quantity range [50, 60], quality grade Q2. Using cross-validation of the process route parameters, three material entries that do not require input for the "AOI Inspection" process are removed, resulting in a remaining subset of 21 BOM entries that satisfy the process route. The quantity and quality grade fields in this subset are mapped to the [0,1] interval using normalization parameters, and the category field is one-hot encoded to form a standard BOM feature vector of length 84. This feature vector significantly improves the comparison efficiency and the stability of the matching results when the Jaccard algorithm calculates the BOM matching degree. S2.2: Compare the standard BOM dataset with the actual collected material identification codes item by item, perform BOM matching degree calculation, use the Jaccard Similarity algorithm to process the matching results, generate a BOM matching metric value, and output the BOM matching degree index. S2.3: Based on the batch information associated with the material identification code, extract the production date, expiration date and current system timestamp of the batch of materials from the warehouse management system, calculate the remaining valid days using the time difference method, and normalize to the [0,1] interval to generate the batch validity period remaining index. Based on the batch information associated with the material identification code, the batch information query interface is called in the warehouse management system (WMS) to obtain the production date field and expiration date field of the batch of materials, and at the same time, the current system timestamp data is collected from the system clock to form an initial time dataset containing three time elements. The remaining valid days are calculated using the time difference method (parameters: production date Dp, expiration date De, current timestamp Dt). This method calculates the difference between the expiration date and the current timestamp on the timeline. The formula for extracting remaining time is as follows:

[0012] in, The UNIX timestamp of the expiration date. This is the current timestamp. It represents the number of seconds in a day, enabling the conversion between seconds and days. Furthermore, through the normalization function (parameters: ΔD, total number of days of validity), The remaining valid days are converted into a batch validity period balance index within the range of [0,1], using the following formula:

[0013] in, This refers to the total number of days from the production date to the expiry date for this batch of materials. This refers to the remaining quantity of the batch's expiration date. An outlier detection algorithm (parameters:) is used. (Threshold set), in the normalization result, batches with values ​​less than the set safety threshold are marked as high-risk batches, so as to realize the early identification of materials nearing their expiration date; Through the above normalization process, the original time difference result of the remaining valid days is transformed into a batch validity period indicator that can be directly used for rule engine calculation, thereby realizing the quantitative availability of material life cycle status in abnormal early warning decision-making. For example, on a component placement production line, the production date for a certain batch of materials is January 1, 2024, the expiration date is June 30, 2024, and the current system time is May 15, 2024. The corresponding UNIX timestamps for these three dates are as follows: , and According to the formula The remaining effective days ΔD are obtained as follows: Days. Total number of days valid. Calculated as ,Right now Heaven, normalized index ≈ When the low-risk threshold is set at 0.3, the batch is marked as being nearing its expiration date and is under warning status. In subsequent judgments, the rule engine incorporates the remaining shelf life contribution of this batch into the comprehensive risk assessment, thereby enabling early prevention and control of potential material quality changes due to approaching expiration. S2.4: Align the process route parameters corresponding to the current work order with the preset standard process route template, calculate the path deviation distance using the Dynamic Time Warping (DTW) algorithm, and then map it to the process route consistency deviation amount through reverse normalization, and output the process route consistency deviation amount index. S2.5: Set threshold ranges and grading functions for BOM matching degree index, batch shelf life remaining index, and process route consistency deviation index respectively. Calculate the rule confidence score corresponding to each index based on the multi-segment linear interpolation method to obtain three independent rule confidence score sub-items. S2.6: The three rule confidence score sub-items are integrated through a weighted fusion function. The weight coefficients are determined based on the statistical analysis of historical erroneous operation frequency. Finally, a unified rule confidence score is generated as the comprehensive credibility metric output by the rule engine. S2.7: Package the BOM matching degree index, batch validity period margin index, process route consistency deviation index and their corresponding rule confidence score sub-items into a structured rule judgment parameter set, and transmit it to the dynamic arbitrator module through the message middleware for subsequent weighted fusion.

[0014] like Figure 2 As shown, step S3 involves constructing an AI recognition model based on a convolutional neural network and attention mechanism. The model takes material images, work order text, and operation log sequence data as input and outputs an unstructured prediction result set containing material type mismatch probability, process parameter deviation, and label integrity index. Specifically, this includes: S3.1: Acquire multimodal raw input data related to the current material status change event during the circuit board production process, including material identification images captured by industrial cameras, work order structured text information extracted by the MES system, and operation log time sequence data recorded by PLC and operation terminal, as input data sources for AI recognition model to ensure that the model has comprehensive context awareness capabilities. S3.2: Perform preprocessing operations on the material identification image, including grayscale normalization, noise filtering and ROI region cropping, use bilinear interpolation algorithm to uniformly scale the image to 224×224 pixel size, and use histogram equalization to enhance local contrast, and generate a standardized image tensor to eliminate the interference caused by differences in imaging conditions to subsequent feature extraction. S3.3: An image feature extraction submodule is constructed based on the ResNet-34 residual convolutional neural network architecture. The standardized image tensor is input into the network, and through layer-by-layer convolution and pooling operations, a deep visual feature map with spatial hierarchy is extracted. After the last layer of global average pooling, a compact 512-dimensional image feature vector is output as the core representation of the material appearance. The input is the image tensor after standardization in step S3.2, with a fixed size of 224×224×C, where C is the number of channels and the pixel value range has been normalized to [0,1], which serves as the direct input to the image feature extraction submodule; A ResNet-34 residual convolutional neural network structure (parameters: number of layers = 34, number of residual units = 16, kernel size = 3×3, initial number of channels = 64) is adopted to achieve hierarchical capture of multi-scale spatial features; Furthermore, by combining convolutional downsampling with stride 2 and max pooling (parameters: pooling window size = 3×3, stride = 2), the signal is compressed stepwise in terms of spatial resolution, and a low-dimensional but semantically rich feature map is obtained, which is used to improve the receptive field of the subsequent feature abstraction layer. Furthermore, through the residual connection mechanism, shallow feature maps and deep convolution outputs are added element-wise in the channel dimension to diversify gradient propagation paths, alleviate the gradient vanishing phenomenon in deep networks, and enhance the model's ability to recognize local textures and overall structures. Furthermore, by performing batch normalization, the feature distribution of each batch of data is centered and standardized to suppress internal covariance shift, enhance training stability, and provide numerically scale-consistent input states for the residual module. Furthermore, by applying a nonlinear mapping to the convolutional output through the ReLU activation function, the expressive power of the model in the feature space is expanded, high-response regions are preserved and low-response noise is suppressed, which helps to capture nonlinear feature patterns such as characters, shapes, and printing defects on the material surface. Furthermore, the input two-dimensional feature map is aggregated by the spatial dimension mean through the global average pooling layer at the end of the network to generate a compact feature vector of length 512. This ensures that the feature vector has a fixed dimension while maintaining high-level semantics, and reduces the number of parameters in the fully connected layer, thereby reducing the risk of overfitting. By combining residual convolutional network structure with global average pooling, the standardized image tensor is transformed into a 512-dimensional core feature vector that reflects the appearance of the material, thereby achieving multi-scale and robust encoding of the material appearance. For example, under the input condition of RGB images acquired by an industrial camera on a patch panel, the first layer of ResNet-34 is set with a kernel size of 3×3, 64 channels, and a stride of 2. The batch normalization parameters are initialized with a mean of 0 and a variance of 1. In the first residual module, the input feature map size is 56×56×64. After two layers of convolution and residual concatenation, the output size remains unchanged, and the internal channel response values ​​are in the range of channel mean from 0.12 to 0.31, indicating that the feature distribution has been effectively compressed. In the fourth residual module, the input size is 14×14×256. After convolution downsampling with a stride of 2 and max pooling, the output size is 7×7×512, and the intra-channel variance is reduced to 0.045. In the global average pooling operation, the 7×7 spatial dimension is averaged to a single value per channel, forming a 512-dimensional vector. This vector significantly improves the Euclidean distance difference between different material categories, thereby enhancing the separability of materials in the appearance feature space. This vector is directly used as input to the cross-modal feature fusion unit, effectively improving the accuracy of abnormal mismatch detection and the sensitivity to complex label defects; S3.4: Semantic encoding processing is performed on the structured text information of the work order. First, keyword annotation is performed using the BOM dictionary and process terminology library. Then, the text sequence is embedded and represented by the pre-trained BERT-base model, outputting a 768-dimensional contextualized word vector sequence. Finally, a fixed-length semantic feature vector of the work order is generated using the max pooling strategy to capture the material requirement logic and process constraint relationship hidden in the work order. For the structured text information of work orders from the MES system, a keyword tagging method based on domain thesaurus matching (parameters: BOM dictionary and process terminology library) is adopted to identify and tag the core terms such as material model, process node, and quality control point involved in the text. Furthermore, by employing a position-based sequence enhancement strategy (parameters: word position information and domain label weights), the labeled text sequence is input into the pre-trained BERT-base model. The deep encoding of contextual information is achieved using a multi-layer bidirectional Transformer structure, resulting in a 768-dimensional contextualized word vector output matrix corresponding to each token. Furthermore, a masked attention mechanism (parameters: label weight matrix and word order masking strategy) is adopted to highlight the attention scores of domain keywords in the above word vector matrix, so as to achieve the focus and feature amplification of key information in the semantic space, and generate a weighted word vector matrix for subsequent pooling processing; Furthermore, by performing pooling operations on the weighted word vector matrix through the max pooling strategy (parameter: pooling window covers the entire sequence length), the feature components with the maximum activation value in all token dimensions are extracted to generate a fixed-length 768-dimensional work order semantic feature vector, so as to retain the global semantic representation that is most relevant to the material demand logic and process constraints. By using semantic encoding and pooling, the original structured text of the work order is transformed into a semantic feature vector of the work order, thereby achieving a quantitative expression of the material matching rules and process requirements implicit in the text. For example, when processing the production work order text for a certain type of circuit board, the work order field contains "Product Model: XJ-210, Process Path: Mounting → Reflow Soldering → AOI Inspection, BOM Required Materials: R100Ω Resistor, C10μF Capacitor". The system first matches "R100Ω" and "C10μF" as standard material entries in the BOM dictionary, and matches "mounting", "reflow soldering", and "AOI inspection" as standard process terms in the process terminology library, and then labels and assigns weights to them respectively. The labeled sequence is input into the BERT-base model. In the word vector matrix output by the encoding layer, each row is a 768-dimensional vector, and the weights in the attention weight matrix corresponding to the labeled terms are increased to twice that of regular words. After masked attention calculation, the vector corresponding to "AOI inspection" has the largest activation value in the 325th dimension. Through max pooling, this maximum value is selected from the 325-dimensional channels of all tokens to form the 325th dimension of the work order feature vector. The other dimensions are generated in the same way, finally obtaining a complete 768-dimensional work order semantic feature vector. In this scenario, after the feature vector is fused with image features and operation log features, it significantly improves the AI ​​recognition model's correct understanding of material demand logic and significantly enhances the detection sensitivity of process deviation anomalies in the subsequent arbitration process. S3.5: Perform sequence modeling on the time-series data of the operation log, align it by timestamp and convert it into an event encoding sequence, input it into a bidirectional long short-term memory network Bi-LSTM for context-aware time-series feature learning, capture the time dependence and abnormal patterns of operation behavior, and finally output a 256-dimensional hidden state feature vector of operation behavior to reflect the compliance level of human intervention process. S3.6: Construct a cross-modal feature fusion unit that receives image feature vectors, work order semantic feature vectors, and operation behavior latent state feature vectors as inputs. The three are mapped to a unified 640-dimensional joint semantic space through a learnable linear projection matrix. A multi-head attention mechanism is introduced to calculate the relevant weights between each modality and generate a weighted fusion comprehensive feature representation to achieve adaptive alignment and complementary enhancement of multi-source information. S3.7: Based on the fused comprehensive feature representation, three parallel output branches of fully connected networks are constructed to predict the probability of material type mismatch, deviation of process parameters, and label integrity index, respectively. Each branch contains two hidden layers (ReLU activation) and Dropout regularization. The final output is processed by Sigmoid or Tanh normalization to form a structured set of unstructured prediction results for subsequent dynamic arbitrator calls. S3.8: Perform confidence calibration processing on the unstructured prediction result set, and adjust the probability estimates of each output branch by temperature factor based on the historical validation sample set using the Platt Scaling method to generate the final AI prediction result set after probability correction, thereby improving the reliability and interpretability of the model output and meeting the strict requirements for prediction stability in industrial scenarios. The Platt Scaling probability calibration method (parameters: historical validation sample set, initial prediction probability vector) is used on the unstructured prediction result set to achieve post-processing stabilization of the probability estimates of each output branch; Furthermore, by fitting a logistic regression function (parameters: original probability p of the output branch, corresponding label y), the mapping learning of the calibration model is achieved, and the parameter vector is obtained. ; Furthermore, through the temperature factor adjustment method (parameter: temperature coefficient) Original probability This process smooths the probability distribution and generates adjusted branch probability values. ; Probability calibration is performed using the following mapping formula:

[0015] in, For logistic regression weights, For bias terms, These are the original probability values. For calibration probability values; Furthermore, by performing multi-dimensional splicing processing on the probability vectors after calibration of each output branch, a global calibration probability matrix is ​​generated for subsequent fusion calculation of the arbitrator; By combining the Platt Scaling algorithm with temperature factor adjustment, the AI ​​prediction results from the previous step are transformed into a calibrated final prediction probability set, achieving high reliability and interpretability in industrial scenarios. For example, in a circuit board material mismatch detection scenario, the input sample consists of 500 tagged historical event records. The three output branches are the material type mismatch probability, process parameter deviation, and tag integrity index. Platt Scaling is used for fitting, with logistic regression weights. Bias term Temperature coefficient For the original probability of mismatched branches The calibration probability is calculated using the formula. ≈ After temperature adjustment, the probability of this branch is corrected to... This effectively suppressed the overfitting tendency of the original model under high-confidence conditions. During the testing phase, the fluctuation of the calibration probability of each branch at the decision threshold was significantly reduced, and the output results showed stable consistency in multiple batch validations, which helps to accurately execute the subsequent dynamic arbitrator weight allocation.

[0016] like Figure 3 As shown, step S4 involves collecting production line operating context parameters, including equipment sensor stability index, shift worker experience level, process cycle time fluctuation rate, and the most recent manual review time interval, constructing a four-dimensional environmental feature vector, and performing normalization processing. Specifically, this includes: S4.1: Obtain the device sensor stability index as input condition, and use the exponential weighted moving average algorithm to smooth the original sampled data based on the signal-to-noise ratio and data packet loss rate sequence within the sliding time window to eliminate the fluctuation effect caused by instantaneous interference, and output the intermediate value of device sensor stability. S4.2: Perform threshold mapping transformation on the intermediate value of the device sensor stability, quantify it into a standardized score in the [0,1] interval according to the preset three-level grading standard (high / medium / low), and generate the device sensor stability index as the first dimension input of the environmental feature vector; The intermediate values ​​of device sensor stability obtained after step S4.1 are processed using a graded threshold mapping method (parameter: graded boundary). This allows for the mapping of stability values ​​to corresponding interval levels. Furthermore, through the interval quantization algorithm (parameters: mapping mode is linear scaling, interval range [0,1]), the high, medium and low levels are mapped to standardized influence values ​​of 1.0, 0.5 and 0.0 respectively, and the initial standardized score matrix is ​​obtained; Furthermore, a piecewise interpolation compensation algorithm is used (parameter: interpolation nodes are...). and At both ends, the interpolation order is linear (first-order), which realizes continuous numerical compensation for the stability intermediate value at the interval boundary and generates a standardized score value with smooth transition; Furthermore, by using a discretization consistency correction algorithm (parameter: consistency coefficient λ, adjusted according to the mean and variance of the continuous rating distribution), the numerical distribution of the rating is morphologically corrected to eliminate the gradient breakage phenomenon that may occur during the rating determination process, and the final rating is generated. By combining hierarchical threshold mapping with interpolation compensation and consistency correction, the intermediate value of device sensor stability in the previous step is transformed into the device sensor stability index, thus achieving a quantitative representation effect as the first dimension input of the environmental feature vector. For example, in a circuit board production line scenario, the sliding time window length is set to 120 seconds, the signal-to-noise ratio after S4.1 smoothing is 22.4dB, the data packet loss rate is 0.8%, and the calculated median value of the device sensor stability is 0.86. (Preset) 0.8 0.5 The value is 0.2, and the hierarchical threshold mapping formula is used:

[0017] in, This is the intermediate value for stability. and These are the low and high cutoff values, respectively. Substituting S=0.86 into the formula, we obtain a standardized score of 0.96, which falls within the high-level range. Using a linear interpolation compensation method... A compensation value of 0.02 is added, and the score is corrected to 0.98. Then, a consistency correction algorithm is applied, setting λ=0.05, with a score distribution mean of 0.85 and variance of 0.03 for correction, resulting in a final device sensor stability index of 0.97. This index is directly used as the first dimension input to the environmental feature vector, significantly improving the sensitivity of the dynamic arbiter to characterize the stability of device operation. S4.3: Collect the identity information and job history data of the operators in the current shift. Based on the historical training records and error operation statistics in the human resources system, calculate the experience level score of the shift personnel using the experience level evaluation model, and compress it to the range of [0,1] using the Sigmoid function to generate the experience level index of the shift personnel, which serves as the second dimension input of the environmental feature vector. S4.4: Obtain the raw data stream of process cycle time volatility. Based on the actual cycle time series of the most recent N production cycles, calculate its relative dispersion using the coefficient of variation method to obtain the basic value of process cycle time volatility. Then, perform Z-score standardization on this value to convert it into a standard distribution value with a mean of 0 and a variance of 1. Finally, normalize it to the [0,1] interval using Min-Max to generate the process cycle time volatility index, which serves as the third-dimensional input of the environmental feature vector. S4.5: Read the timestamp record of the most recent manual review, calculate the most recent manual review time interval based on the time difference between the current system time and the timestamp; use the negative exponential decay function to perform nonlinear mapping on the time difference, generate a continuous variable in the interval [0,1] that reflects the freshness of supervision, and output the most recent manual review time interval index as the fourth dimension input of the environmental feature vector; Based on the most recent manual verification timestamp record obtained from the production process data acquisition terminal, a high-precision time difference method is used (parameter: current system time). Review timestamp This allows for the calculation of the review interval time Δt. Furthermore, using the time difference formula:

[0018] in, This is the time difference between the system time and the timestamp of the most recent manual review. The current system time. The timestamp of the most recent manual review enables precise quantification of the review interval; The nonlinear transformation from time interval to regulatory freshness index is achieved through a negative exponential decay function mapping method (parameter: decay coefficient λ), as shown in the formula:

[0019] in, For monitoring freshness indicators (value range [0,1]), The exponential decay parameter is set according to the production line review frequency; Furthermore, by normalizing the function... Boundary compression is performed to stabilize the output results within the closed interval [0,1] to ensure comparability at different time scales; An outlier detection algorithm (parameters: upper and lower thresholds) , This enables the filtering of anomalies in the mapping results and outputs a default substitution value when data out of bounds is detected, preventing extreme time differences from affecting the stability of downstream models. Furthermore, the validated regulatory freshness index is used as the fourth dimension input of the environmental feature vector, which together with the equipment sensor stability index, shift personnel experience level and process cycle time fluctuation rate to form a complete environmental state descriptor. By combining the above-mentioned time difference and negative exponential decay mapping, the timestamp difference of the previous step is transformed into quantitative data reflecting the timeliness of manual supervision, thereby achieving accurate construction of the "freshness of manual review" dimension in the input of dynamic arbitrator environment features. For example, on a certain SMT production line, the current system time is 08:00:00 on June 12, 2024, and the most recent manual verification time is 20:00:00 on June 11, 2024. High-precision time difference calculation yields Δt = 43200 seconds. Setting the attenuation coefficient λ = 0.00005, Δt is substituted into the negative exponential attenuation function for calculation:

[0020] The result is R≈0.114. After normalization and boundary compression, the output regulatory freshness index is 0.114, which is compared with the preset threshold. =0.05、 =0.95, which is within the valid range and is directly used as the fourth dimension input of the environmental feature vector. In the subsequent dynamic arbitrator weight allocation process, when this low value appears in combination with other three-dimensional environmental parameters, the system automatically increases the weight of the rule engine, significantly improving the anomaly detection rate in scenarios without long-term manual review. In this embodiment, because it promptly reflects the interval of manual supervision in the verification batch, the misjudgment rate of the comprehensive judgment result is significantly reduced, and the response accuracy is improved. S4.6: Combine the standardized indicators of the above four dimensions—equipment sensor stability index, shift personnel experience level, process cycle fluctuation rate, and most recent manual review time interval—in a predetermined order to form a four-dimensional environmental feature vector; perform L2 norm normalization on this vector to eliminate dimensional differences and ensure consistent vector magnitude, and output an environmental feature vector at a unified scale for subsequent use by the dynamic arbiter model.

[0021] Step S5: Construct a lightweight dynamic arbitrator model. Taking the rule confidence score, AI prediction result set, and environmental feature vector as input, and after feature cross-validation and normalization, a gating mechanism is used to generate rule path activation weights and AI path activation weights. The sum of the two weights equals 1. Specifically, this includes: S5.1: Based on the rule confidence score, AI prediction result set, and normalized four-dimensional environmental feature vector output from the previous steps, construct a multi-source heterogeneous input tensor for the dynamic arbitrator. The rule confidence score constitutes a structured logical judgment subspace, the AI ​​prediction result set constitutes an unstructured semantic understanding subspace, and the environmental feature vector constitutes a runtime context control subspace. The three together serve as the joint input conditions for the dynamic arbitrator, providing a data foundation for subsequent weight allocation. S5.2: Perform feature normalization processing on the multi-source heterogeneous input tensor, and use the Z-score normalization method to perform zero-mean unit variance transformation on the features of each dimension to eliminate the dimensional differences between indicators from different sources, and ensure that the rule confidence score, AI classification probability distribution and environmental parameters participate in the calculation under a unified numerical scale, thereby obtaining a standardized feature matrix and improving the stability and convergence speed of subsequent gating network training. S5.3: Perform high-order feature cross operations based on the standardized feature matrix, and construct second-order cross feature terms using the outer product coupling method to generate a composite cross feature set including rule-AI interaction strength, environment-rule adjustment coefficient, and environment-AI sensitivity gain, so as to capture the nonlinear collaborative relationship between the rule engine and the AI ​​model under specific production line conditions and obtain an enhanced feature representation with context awareness. S5.4: Construct a lightweight gated neural network as the core decision-making unit of the dynamic arbitrator. Its network structure consists of an input layer, a single hidden layer (containing 16 ReLU activated neurons), and a dual-output Softmax layer. The enhanced feature representation is input into the network, and the original weights logits are calculated through forward propagation. Then, after normalization by the Softmax function, the regular path activation weights and AI path activation weights are output to realize the differentiable dynamic allocation of the contribution of the dual judgment sources. S5.5: Based on the rule path activation weight and AI path activation weight output by the dynamic arbitrator, an interpretable feedback channel is constructed. The weight ratio is encoded as a quantitative indicator of the system's dependence on AI under the current operating conditions and written into the real-time monitoring log to support subsequent anomaly root cause tracing and model behavior auditing, ensuring the transparency and engineering controllability of the arbitration process.

[0022] Step S6: Based on the rule path activation weight and the AI ​​path activation weight, the rule judgment parameter set and the AI ​​prediction result set are weighted and fused to generate a comprehensive anomaly score. When this score exceeds a preset threshold, a graded early warning signal is triggered. Specifically, this includes: S6.1: Obtain the rule path activation weight and AI path activation weight output by the lightweight dynamic arbitrator, where the sum of the two weights is 1, and the weight allocation is determined by the normalized environmental feature vector, rule confidence score and AI prediction probability distribution in the previous steps, and serves as the input control parameter for the weighted fusion operation. S6.2: Perform quantization mapping processing on the rule judgment parameter set generated in the previous step S2, and convert the three indicators of BOM matching degree, batch validity period remaining amount, and process route consistency deviation into standardized anomaly sub-scores R1, R2, and R3 respectively, and map their value range to the interval [0,1] to eliminate the difference in dimensions and obtain the structured rule anomaly contribution vector. S6.3: Perform semantic parsing and numerical normalization on the AI ​​prediction result set output in the previous step S3, convert the three unstructured prediction results of material type mismatch probability, process parameter deviation degree, and label integrity index into corresponding risk sub-scores A1, A2, and A3, and construct an unstructured AI anomaly contribution vector to ensure that it is comparable to the rule path score. Based on the AI ​​prediction result set output in the previous step S3, a semantic parsing method (parameters: unstructured prediction result set, model label dictionary, process terminology library) is used to realize the semantic structured mapping of "material type mismatch probability", "process parameter deviation" and "label integrity index" to obtain the corresponding standardized feature label set. Furthermore, through a numerical normalization algorithm (parameters: normalization interval [0,1], original prediction probability or deviation value), the numerical results corresponding to the semantic labels are mapped to intervals, and various unstructured prediction indicators are converted into risk measurement forms with unified dimensions. Furthermore, a direct normalization mapping is performed on the "probability of material type mismatch" to generate a risk sub-score A1; a threshold sensitivity adjustment factor is introduced for the "deviation of process parameters", and risk quantification is achieved through a linear scaling formula to generate a risk sub-score A2; and a reverse mapping strategy is adopted for the "label integrity index" to convert the degree of integrity loss into a risk quantification value to generate a risk sub-score A3. Furthermore, an eigenvector construction method (parameters: A1, A2, A3) is employed to achieve an ordered combination of the three risk sub-scores, forming an unstructured AI anomaly contribution vector. = [A1, A2, A3], ensuring comparability with the structured rule anomaly contribution vector in terms of numerical range and index semantics; The above processing method transforms the unstructured AI prediction results from the previous step into a structured and quantified risk contribution vector, which is a prerequisite for the unified integration of rule-based path scoring and AI path scoring. For example, in a circuit board production line anomaly detection, the AI ​​predicted a material type mismatch probability of 0.73, a process parameter deviation of 0.12, and a label integrity index of 0.85. The model label dictionary maps the "material type mismatch probability" to a risk sub-score A1, using an interval normalization formula. A1 = 0.73 was obtained; the deviation of process parameters was adjusted by a threshold sensitivity adjustment factor k = 1.5, and the calculation formula was obtained. A2 = 0.18; the label integrity index is calculated using the inverse risk mapping formula. A3 = 0.15. The three sub-scores are combined according to the eigenvector construction method as follows: = [0.73, 0.18, 0.15], which is comparable to the regular path anomaly contribution vector [R1=0.80, R2=0.22, R3=0.30] under the same scale. After subsequent weighted fusion calculation, the accuracy of the final anomaly score is significantly improved. S6.4: Based on the rule-based anomaly contribution vector and the AI ​​anomaly contribution vector, perform weighted fusion calculation, and use a linear combination method to fuse the sub-scores of each dimension to generate a fused multi-dimensional anomaly state vector, retaining the joint decision results of each dimension; Based on the structured rule-based anomaly contribution vector of the input With unstructured AI anomaly contribution vector A linear combination fusion method is adopted (parameter: regular path activation weight). AI path activation weight This enables the joint calculation of anomaly sub-scores across various dimensions; Anomaly scoring of the i-th rule is performed through dimension-wise weighting operations. AI anomaly scoring The weighted summation process is performed using the following formula: ,in The score after fusion of the i-th dimension of the abnormal factor. and To meet The real-valued weight coefficients of the constraints ensure the consistency of the total weight allocation; Furthermore, through vectorization, the weighted results of each dimension are used to construct a fusion state vector. The independent performance values ​​of BOM matching path, batch validity period path and process route consistency path are retained after fusion calculation; The numerical consistency verification algorithm (parameters: synthesized state vector norm, single-dimensional normalization range) is used to verify that the fusion results of each dimension are within the acceptable scoring range, preventing the fusion value from going out of bounds due to abnormal weight settings or input data mismatch. The matrix addition and scalar multiplication interfaces based on high-precision floating-point arithmetic are adopted to ensure the stability and low latency of fused computing in real-time industrial environments. By using a linear combination weighted fusion method, the rule anomaly contribution and AI anomaly contribution from the previous step are transformed into multidimensional state assessment data that can be directly used for comprehensive anomaly score aggregation, thereby achieving the quantitative integration effect of the collaborative judgment results of the rule engine and the AI ​​model. For example, in a single circuit board mounting process, the regular anomaly contribution vector Set to [0.60, 0.45, 0.30], AI anomaly contribution vector Set to [0.75, 0.55, 0.40], the output of the lightweight dynamic arbitrator... 0.4 The value is 0.6. According to the formula, the weighted result for the first dimension is... The weighted result for the second dimension is: The weighted result for the third dimension is: Fusion state vector The scores are [0.69, 0.51, 0.36]. After numerical consistency verification, all three dimensions of the score are within the range of [0,1], meeting the input requirements for subsequent comprehensive aggregation calculations. Under this condition, the values ​​of each dimension of the fused vector are significantly improved compared to the original rule score or AI score, which can effectively reflect the gain of the AI ​​model in judgment sensitivity under complex defect scenarios, and provide high-quality input for the final comprehensive score calculation of S6.5. S6.5: Perform aggregation operation on the fused multidimensional abnormal state vector, calculate the comprehensive abnormality score using weighted Euclidean norm or entropy weight method, and output a scalarized overall abnormality index; S6.6: Compare the comprehensive anomaly score with the preset multi-level judgment thresholds. If the comprehensive anomaly score ∈ (Level 1 warning threshold, Level 2 warning threshold), a Level 1 warning signal is triggered; if the comprehensive anomaly score ∈ (Level 2 warning threshold, Level 3 warning threshold), a Level 2 warning signal is triggered; if the comprehensive anomaly score > Level 3 warning threshold, a Level 3 emergency warning signal is triggered, thus realizing a graded response mechanism. Comprehensive abnormal score As the computational object for grading and determination, the input is a standardized overall anomaly index calculated using the weighted Euclidean norm or entropy weight method. An interval comparison algorithm is employed (parameter: Level 1 warning threshold). Level II warning threshold Level III Emergency Warning Signal This enables the determination of the interval assignment between abnormal score values ​​and multi-level judgment thresholds; Furthermore, by using a logical interval mapping method (parameters: judgment result, warning signal level mapping table), the following can be achieved: The numerical data points are categorized and the corresponding warning signal level codes are obtained. ; Furthermore, through a conditional triggering mechanism (parameter: (Signal trigger interface configuration) enables the call and control of the MES system's early warning interface or HMI terminal's warning module, and generates trigger signals containing early warning level identifiers and corresponding processing instruction codes; Furthermore, by using the event routing table lookup method (parameters: warning level, anomaly type category), dynamic matching of response paths is achieved to determine the target node set for information push, including on-site HMI terminals, MES system management terminals, and team leader mobile devices; By using a distributed push control algorithm driven by early warning levels, the interval judgment result of the previous step is transformed into an executable hierarchical response trigger signal, so as to achieve the expected technical effect of rapid and hierarchical handling of abnormal events at different levels. For example, in a certain circuit board production batch, the comprehensive anomaly score is calculated as follows: Preset threshold for , for , for Use interval comparison algorithm to determine satisfy Conditions. Obtained through the level mapping table. A value of 2 corresponds to a level 2 warning signal. The trigger mechanism invokes the intermediate warning channel on the on-site HMI and the team leader's mobile device, pushing a data packet containing the anomaly type "incomplete material label," the current work order number, and suggested actions. In another scenario, if... for The interval comparison results satisfy Then the mapping A score of 3 triggers a Level 3 emergency alert. In addition to on-site notifications, the alert also invokes the emergency shutdown command interface of the MES system, significantly improving the timeliness and effectiveness of preventing potential material errors. S6.7: Generate a complete early warning message package containing the early warning level, anomaly type, confidence score, material identification code involved, work order number, and suggested handling actions, and push it to the MES system, on-site HMI terminal, and team leader's mobile device through industrial communication protocol to complete the closed-loop transmission of information on abnormal events.

[0023] Step S7: Record a complete judgment process log, including the original input data, output results of each module, weight allocation parameters, and manual confirmation feedback, and construct a training dataset for online reinforcement learning updates of the dynamic arbitrator model. Specifically, this includes: S7.1: Based on the material status change event triggered during the circuit board production process, obtain the original data stream consisting of the initial input dataset. The original data stream includes material identification code, work order number, process route parameters, equipment status data and operator identity information, which serve as the starting input conditions for judging process log records, so as to ensure that subsequent audit traces have a complete contextual traceability basis. S7.2: Perform feature serialization processing on the structured output from the rule judgment module, extract three quantitative indicators: BOM matching degree, batch validity period remaining amount, and process route consistency deviation amount, as well as their corresponding rule confidence scores, generate a structured judgment evidence set, and bind it with the original input data stream by timestamp alignment to form intermediate log record items with causal correlation; S7.3: Connect to the unstructured prediction result set output by the AI ​​recognition model. The result set includes the probability of material type mismatch, deviation of process parameters, and label integrity index. Use a semantic encoder to convert it into a vector representation, generate a standardized AI prediction feature vector, and package and encapsulate the vector with the corresponding work order context information to complete the structured archiving of the unstructured decision basis. S7.4: Collect multi-dimensional environmental context parameters during production line operation, including equipment sensor stability index, shift personnel experience level, process cycle fluctuation rate, and the most recent manual review time interval. Perform numerical scaling transformation on the four-dimensional environmental feature vector based on the normalization function to obtain a unified dimension environmental state descriptor, and write it into the log entry as a context annotation field for the dynamic arbiter weight allocation behavior. S7.5: Obtain the rule path activation weight and AI path activation weight generated by the lightweight dynamic arbitrator. The output interface based on the weighted fusion mechanism captures this set of weight parameters in real time. Combined with the comprehensive anomaly score and the graded early warning signal type, a snapshot of decision behavior is generated and stored in the log database as the core metadata reflecting the dual-engine collaborative status. S7.6: After the early warning is pushed, receive a manual confirmation feedback signal from the on-site operator or quality management personnel. The feedback signal includes three types of tag information: anomaly confirmation, false alarm mark, and root cause correction suggestion. Use a tag parser to classify and decode them, and associate and bind the decoded feedback tag with the log record of the corresponding early warning event with a unique ID. S7.7: Based on the preset log aggregation strategy, perform spatiotemporal alignment and field integration operations on the above-mentioned original input data, structured judgment evidence set, AI prediction feature vector, environmental state descriptor, decision behavior snapshot and manual confirmation feedback annotation to generate a complete and semantically consistent judgment process log record, and persist it in the distributed audit database in JSON Schema format; S7.8: Periodically extract the latest batch of judgment process log records with manual confirmation feedback annotations from the audit database to construct a labeled training dataset. The training dataset uses environmental state descriptors, rule confidence scores, and AI prediction feature vectors as input features and manual confirmation feedback annotations as reward signal targets to drive the online reinforcement learning update process of the dynamic arbitrator model. S7.9: Input the completed training dataset into the online learning subsystem of the dynamic arbitrator, use the proximal policy optimization (PPO) algorithm to update the gradient of the trainable parameters of the gated network, adjust the mapping strategy of generating rule path activation weights and AI path activation weights under different environmental conditions, and improve the ability of the arbitration mechanism to recognize and adapt to the actual operation of the production line. S7.10: Verify the performance of the updated dynamic arbitrator model on the reserved test set, calculate the changing trends of its three key indicators: false positive rate, false negative rate, and response latency. If the preset model iteration admission criteria are met, the new version of the model will be deployed to the production environment to replace the old weight allocation strategy, completing the closed-loop control chain from log recording to intelligent optimization.

[0024] Step S8: Based on the updated dynamic arbitrator model parameters after reinforcement learning, adjust the weight allocation strategy for subsequent material state change events to continuously optimize the accuracy of anomaly warnings. Specifically, this includes: S8.1: Based on the training dataset built in S7, extract a multimodal historical decision sample sequence containing the original input data, rule confidence scores, AI prediction result set, environmental feature vector, weight allocation parameters and manual confirmation feedback signals, as the state input space of the reinforcement learning agent, so as to form a complete representation of the current system decision context. S8.2: Define the action space in the reinforcement learning framework as a combination adjustment vector of the rule path activation weights and AI path activation weights output by the dynamic arbitrator gating network, where each action represents a set of executable weight allocation strategies to ensure the consistency and interpretability of the arbitration logic; S8.3: Design a reward function based on the weighted sum of the anomaly detection accuracy gain and the false alarm rate reduction. Use the human confirmation feedback signal as a sparse reward source. When the judgment result of the comprehensive anomaly score is consistent with the human confirmation feedback signal, a positive reward is given, and otherwise a negative penalty is applied, thereby guiding the agent to learn the optimal weight allocation strategy. S8.4: The Deep Deterministic Policy Gradient (DDPG) algorithm is used to construct a reinforcement learning agent in a continuous action space. Its policy network receives the state vector generated by S8.1 as input and outputs the optimal weight allocation action. The value network is used to evaluate the long-term reward expectation of the current state-action pair, achieving stable convergence in a high-dimensional continuous control space. S8.5: Update the latest policy network parameters trained in S8.4 to the gating mechanism of the dynamic arbitrator, replace the original gating weight configuration, complete the online deployment of model parameters, and enable the weight allocation strategy for subsequent material state change events to have adaptive adjustment capabilities based on historical experience optimization. S8.6: Set up a performance monitoring module within the sliding time window to statistically analyze the anomaly warning accuracy, false negative rate, and response latency indicators of the system after the update in real time. If no significant performance improvement is observed in N consecutive batches, the incremental retraining process is triggered to collect new period data to expand the training set and maintain the model's evolution momentum. S8.7: Write each parameter update process and its corresponding performance change log to the audit database, and mark the version number and effective timestamp to establish a traceable model iteration genealogy, support subsequent compliance review and fault attribution analysis, and ensure the reliable operation of the system in industrial scenarios.

[0025] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0026] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent management of material traceability and error prevention throughout the entire circuit board manufacturing process, characterized in that: Includes the following steps: S1: Obtain the material status change event trigger signal during the circuit board production process, and collect the material identification code, work order number, process route parameters, equipment status data and operator identity information as the initial input dataset; S2: Perform structured feature extraction on the initial input dataset to generate a set of rule judgment parameters, including three quantitative indicators: BOM matching degree, batch shelf life remaining amount, and process route consistency deviation amount, and calculate the rule confidence score of each indicator; S3: Based on convolutional neural networks and attention mechanisms, an AI recognition model is built. The input includes material images, work order texts, and operation log sequence data. The output includes a set of AI prediction results containing material type mismatch probability, process parameter deviation, and label integrity index. S4: Collect production line operating context environment parameters and construct a four-dimensional environment feature vector; S5: Construct a dynamic arbitrator model, taking the rule confidence score, the AI ​​prediction result set and the four-dimensional environmental feature vector as inputs, and after feature cross-processing and normalization, use a gating mechanism to generate rule path activation weights and AI path activation weights. S6: Based on the rule path activation weight and the AI ​​path activation weight, the rule judgment parameter set and the AI ​​prediction result set are weighted and fused to generate a comprehensive anomaly score value. When the comprehensive anomaly score value exceeds a preset threshold, a graded early warning signal is triggered.

2. The intelligent management method for material traceability and error prevention throughout the entire circuit board production process according to claim 1, characterized in that, Following step S6, the following is also included: S7: Record the complete judgment process log, including the original input data, the output results of each module, the weight allocation parameters and the manual confirmation feedback, and build a training dataset for online reinforcement learning updates of the dynamic arbitrator model; S8: Adjust the weight allocation strategy for subsequent material state change events based on the updated dynamic arbitrator model parameters after reinforcement learning.

3. The intelligent management method for material traceability and error prevention throughout the entire circuit board production process according to claim 1, characterized in that, Step S1 specifically includes: Based on the real-time event release mechanism in the industrial production line control system, the system monitors and acquires material status change event trigger signals caused by mounting, soldering, testing or handling processes in the circuit board production process. The event context association processing is performed on the material status change event trigger signal, and the material identification code bound to the current event is synchronously pulled from the manufacturing execution system using the message queue middleware; Based on the material identification code, a work order traceability query is performed. The work order management module of the ERP system is called through the enterprise service bus interface to obtain the work order number associated with the material corresponding to the material identification code. Based on the work order number, the corresponding process route parameters are extracted from the MES system, encapsulated and transmitted in XML format, and the standard process path that the current work order should follow is obtained by parsing the process route parameters. The equipment status data related to the current material operation is obtained through a distributed data acquisition gateway. The equipment status data is then processed by noise reduction filtering and timestamp alignment to generate a snapshot of equipment operation. Based on the identity authentication logs of the on-site operation terminal, the identity information of the operator who performed this material operation is collected, its validity is verified, and combined with the training history database in the human resources system, a preliminary assessment vector of the personnel's experience is generated. The material identification code, the work order number, the process route parameters, the equipment operation snapshot, and the personnel experience initial evaluation vector are normalized and encoded according to a predefined structure. Null value filling and outlier marking operations are performed to generate an initial input dataset in a unified format.

4. The intelligent management method for material traceability and error prevention throughout the entire circuit board production process according to claim 3, characterized in that, The material identification code is a one-dimensional barcode or two-dimensional barcode information based on the GS1 standard encoding, which is obtained by scanning with an industrial barcode reader and serves as the input basis for identifying the unique identity of the material actually used.

5. The intelligent management method for material traceability and error prevention throughout the entire circuit board production process according to claim 1, characterized in that, Step S2 specifically includes: Based on the material identification code, work order number and process route parameters, the corresponding standard bill of materials is obtained from the enterprise resource planning system as the benchmark reference data, and the expected set of materials to be used under the current work order is established to obtain the standard BOM dataset. The standard BOM dataset is compared item by item with the actual collected material identification codes. The BOM matching degree is calculated, and the matching result is processed by the set intersection-union ratio algorithm to generate a BOM matching metric value and output the BOM matching degree index. Based on the batch information associated with the material identification code, the production date, expiration date and current system timestamp of the batch material corresponding to the batch information are extracted from the warehouse management system. The remaining valid days are calculated and normalized to a unified range to generate a batch validity period remaining index. Align the process route parameters corresponding to the current work order with the preset standard process route template, calculate the path deviation distance, and then map it into the process route consistency deviation amount through reverse normalization, and output the process route consistency deviation amount index. Threshold ranges and grading functions are set for the BOM matching degree index, the batch shelf life remaining index, and the process route consistency deviation index, respectively. The rule confidence score corresponding to each index is calculated to obtain three independent rule confidence score sub-items. The three rule confidence score sub-items are integrated and processed to generate a unified rule confidence score; The BOM matching index, the batch validity period margin index, the process route consistency deviation index, and their corresponding rule confidence score sub-items are packaged and encapsulated into a structured rule judgment parameter set, which is then transmitted to the dynamic arbitrator module via a message middleware.

6. The intelligent management method for material traceability and error prevention throughout the entire circuit board production process according to claim 5, characterized in that, Step S2 further includes obtaining the standard BOM corresponding to the work order from the enterprise resource planning system, using GS1 barcode unique mapping and cross-validation of process parameters, comparing actual materials used with set intersection-union comparison algorithm, normalizing the batch validity period margin with time difference, comparing the consistency of process route with standard process template using dynamic time warping algorithm, generating rule confidence scores through hierarchical function and weighted fusion respectively, and packaging them into a structured rule judgment parameter set.

7. The intelligent management method for material traceability and error prevention throughout the entire circuit board production process according to claim 1, characterized in that, Step S3 specifically includes: Acquire multimodal raw input data related to current material status change events during the circuit board production process, including material identification images, work order structured text information, and operation log time sequence data; Perform preprocessing operations on the material identification image to generate a standardized image tensor; An image feature extraction submodule is constructed based on a residual convolutional neural network architecture. The standardized image tensor is input into the residual convolutional neural network, and deep visual feature maps are extracted through layer-by-layer convolution and pooling operations. Finally, the image feature vector is output after the last layer of global average pooling. The structured text information of the work order is semantically encoded to output a contextualized word vector sequence, and a max pooling strategy is used to generate the semantic feature vector of the work order. Sequence modeling is performed on the time-series data of the operation log, which is then aligned by timestamps and converted into an event-encoded sequence. This sequence is then input into a bidirectional long short-term memory network for context-aware time-series feature learning, and the hidden state feature vector of the operation behavior is output. A cross-modal feature fusion unit is constructed, which receives the image feature vector, the work order semantic feature vector, and the operation behavior latent state feature vector as input, maps the three to a unified joint semantic space, calculates the relevant weights between each modality, and generates a comprehensive feature representation. Based on the comprehensive feature representation, three parallel fully connected networks with output branches are constructed to predict the probability of material type mismatch, deviation of process parameters, and label integrity index, respectively. After normalization, a structured set of unstructured prediction results is formed. The confidence level is calibrated on the unstructured prediction result set, and the probability estimates of each output branch are adjusted by a temperature factor based on the historical validation sample set to generate the AI ​​prediction result set.

8. The intelligent management method for material traceability and error prevention throughout the entire circuit board production process according to claim 7, characterized in that, The preprocessing operations include grayscale normalization, noise filtering and ROI region cropping. The image is uniformly scaled to 224×224 pixels using bilinear interpolation algorithm, and histogram equalization is used to enhance local contrast.

9. The intelligent management method for material traceability and error prevention throughout the entire circuit board production process according to claim 1, characterized in that, Step S4 specifically includes: The device sensor stability index is obtained as input. Based on the signal-to-noise ratio and data packet loss rate sequence within the sliding time window, the original sampled data is smoothed and the intermediate value of device sensor stability is output. A threshold mapping transformation is performed on the intermediate value of the device sensor stability, and it is quantified into a standardized score within a unified range according to a preset three-level grading standard to generate a device sensor stability index, which is used as the first dimension input of the environmental feature vector. Collect the identity information and job history data of the operators in the current shift, calculate the experience level score of the shift personnel based on the historical training records and error operation statistics in the human resources system, compress it into a uniform range, and generate the experience level index of the shift personnel as the second dimension input of the environmental feature vector; The raw data stream of process cycle fluctuation is obtained. Based on the actual cycle time series of the most recent N production cycles, its relative dispersion is calculated to obtain the basic value of process cycle fluctuation. The basic value of process cycle fluctuation is standardized to generate the process cycle fluctuation index, which is used as the third dimension input of the environmental feature vector. Read the timestamp record of the most recent manual review, calculate the most recent manual review time interval based on the time difference between the current system time and the timestamp record, use the negative exponential decay function to perform nonlinear mapping on the time difference, generate a continuous variable within a unified interval that reflects the freshness of supervision, and output the most recent manual review time interval index as the fourth dimension input of the environmental feature vector; The standardized indicators of the four dimensions are combined in a predetermined order to form a four-dimensional environmental feature vector. The four-dimensional environmental feature vector is then normalized to output an environmental feature vector at a uniform scale.

10. The intelligent management method for material traceability and error prevention throughout the entire circuit board production process according to claim 1, characterized in that, The dynamic arbitrator model includes an input layer, a single hidden layer, and a dual-output Softmax layer. The single hidden layer contains 16 ReLU activated neurons.

Images