Production process tracing method based on cooperation of large and small models

Abstract

The invention discloses a production process tracing method based on large and small model collaboration, and relates to the technical field of industrial intelligent manufacturing, and the method achieves the whole process optimization through an end-cloud collaboration architecture: 1) deploying a multi-modal sensor (vibration / hyperspectrum / PLC) at an edge end, and employing a MobileNetV3 + GRU lightweight model to extract layered features; 2) constructing a BERT-GAT-driven dynamic knowledge graph at the cloud side, and supporting new node creation triggered by feature similarity and TCN time sequence weight updating; 3) designing a bidirectional knowledge distillation mechanism, generating decision rules through Grad-CAM pruning in the forward direction, and aligning cross-process features by using DTW in the reverse direction; 4) fusing A * path search and credible traceability of Monte Carlo simulation, and combining COMSOL physical verification to realize error lt; and 7% root cause positioning. The implementation effect shows that the cross-process anomaly tracing time is shortened by 96%, the end side model resource occupation is reduced by 60%, the dynamic rule generation speed is increased by 36 times, and the complex coupling anomaly tracing credibility is greater than or equal to 90%.

Description

Technical Field

[0001] The present invention relates to the field of industrial intelligent manufacturing technology, and in particular to a production process tracing method based on collaboration between large and small models. Background Art

[0002] Industrial production often results in defective products. Tracing and analyzing the causes of these products is a major challenge in industrial manufacturing. Production process traceability is a key component in ensuring product quality. Production lines often involve hundreds of devices and tens of thousands of parameters that require monitoring and management. Currently, defect analysis relies primarily on manual experience, product disassembly, and the use of testing equipment. This consumes significant human resources and hinders accurate attribution, hindering timely optimization of production line equipment parameters. This results in companies suffering financial losses from scrapped products.

[0003] In view of this, we propose a production process traceability method based on the collaboration of edge computing and cloud-side large models. It deeply integrates the cognitive capabilities of large models with the specialized capabilities of small models, deploys small models on hundreds of devices for real-time monitoring, and interacts with large cloud models to understand the changes in product form caused by parameter changes. This forms a knowledge dissemination chain from the edge to the cloud and then from the cloud to the edge, achieving rapid root cause location and knowledge self-optimization of complex production process anomalies. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a production process tracing method based on the collaboration of large and small models. It deeply integrates the cognitive capabilities of large models with the specialized capabilities of small models, deploys small models on hundreds of devices for real-time monitoring, and interacts with large models in the cloud to understand the changes in product form caused by parameter changes, forming a knowledge dissemination chain from the end to the cloud and then from the cloud to the end, thereby realizing the rapid root cause location and knowledge self-optimization of complex production process anomalies.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a production process tracing method based on large and small model collaboration, comprising the following steps:

[0006] Step S1: End-cloud collaborative data collection and feature extraction

[0007] (1) Multi-source data collection

[0008] ① Device-side deployment: Deploy edge computing nodes on key equipment such as PCB drilling machines and electroplating tanks, integrating:

[0009] 1) Structured data collector: real-time collection of temperature (±0.1°C), pressure (±0.5kPa), and rotation speed (±10rpm);

[0010] 2) Unstructured data collector: A 5μm-class industrial camera captures microscopic images of the drilled hole wall;

[0011] 3) Timing data collector: The vibration sensor records the operating waveform at a sampling rate of ≥10kHz.

[0012] (2) Lightweight feature extraction

[0013] ① Small model architecture: Using MobileNetV3+GRU hybrid network, where:

[0014] 1) Image branch: MobileNetV3 inverted residual structure (FLOPs < 100M), output 1024-dimensional features;

[0015] 2) Temporal branch: two-layer GRU (hidden layer 64 dimensions), outputting 128-dimensional features.

[0016] Step S2: Dynamic process knowledge graph construction

[0017] (1) Entity relationship extraction

[0018] ① Cloud-side large model analysis: using the BERT-GAT architecture:

[0019] 1) Text parsing: Extracting entities (e.g., “abnormal pH value of electroplating solution”) from maintenance logs, with an F1-score ≥ 0.92;

[0020] 2) Image parsing: ResNet-50 detects defect patterns (such as burrs on hole walls), with mAP ≥ 0.85.

[0021] (2) Dynamic update of the map

[0022] ①New node creation conditions: When the small model detects an unrecorded anomaly:

[0023] Create a new node

[0024] ②Edge weight update: Analyze historical event frequency through TCN Step S3: Bidirectional knowledge distillation and collaborative reasoning

[0025] (1) Positive knowledge injection (large model → small model)

[0026] ① Subgraph extraction: When the abnormal index Ia = (current value - baseline value) / (threshold - baseline value) > 0.85, extract the associated subgraph G8

[0027] ② Rule distillation: Use the GAT pruning algorithm to retain the first 20% of edges and convert them into decision tree rules.

[0028] (2) Reverse experience feedback (small model → large model)

[0029] ① Cross-process correlation calculation: Aggregate end-side abnormal features and construct a matrix:

[0030]

[0031] ② Graph expansion: If Mcoe[i,j]>0.8 and there is no edge eij, add an edge and update the weight

[0032] Step S4: Production process traceability execution

[0033] (1) Abnormal source location

[0034] ① Fast positioning on the end side: The small model matches the distillation rules (such as vibration exceeding the limit → clamping cylinder failure), and an alarm is triggered when the confidence level is greater than 0.8.

[0035] ② Cloud-side root cause derivation: The large model calculates cross-process path scores through GAT:

[0036]

[0037] (2) Traceability report generation

[0038] ① Visual annotation: highlight the critical paths with edge weight > 0.7 (such as "clamping force ↓ → aperture deviation ↑ → electroplating turbulence").

[0039] Step S5: Closed-loop knowledge optimization

[0040] (1) Health assessment: H = 0.6 × (number of covered processes / total number of processes) + 0.4 × (number of newly added edges / total number of edges)

[0041] (2) Incremental learning:

[0042] ① Data replay pool: stores recent abnormal cases (capacity ≥ 10,000);

[0043] Update strategy: Full graph update every month, incremental update of edge weights every day.

[0044] The present invention provides a production process tracing method based on the collaboration of large and small models. Compared with the existing technology, it has the following advantages:

[0045] (1) Adopting a real-time, deep, two-way interactive architecture that collaborates with both end and cloud, compared to the traditional single-model architecture, the cross-process root cause tracing time is shortened from 8 hours to 5 minutes, improving efficiency by 96%; and the end-side model resource usage is reduced by 60%;

[0046] (2) Through the adaptive evolution capability of the dynamic knowledge graph, the problem of delayed rule updates in traditional static knowledge bases is solved. The rule generation time for new abnormal patterns is reduced from 3 days with manual intervention to 2 hours, and the response speed is increased by 36 times.

[0047] (3) Combining multimodal credibility verification with physical-data fusion, the misjudgment rate is reduced from 25% in traditional single data verification to below 7%, and the credibility of the tracing results of complex coupled anomalies is ≥90%. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a schematic diagram of the process of the present invention;

[0049] Figure 2 This is a schematic diagram of the network structure of a small model of the present invention;

[0050] Figure 3 This is a schematic diagram of the cloud-side large model network structure of the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] Example

[0053] This embodiment is based on a case in which an abnormal drilling process in a PCB manufacturing plant results in uneven copper thickness on the hole wall after electroplating. The root cause needs to be quickly located and process parameters adjusted.

[0054] See also Figure 1-3 The present invention provides a technical solution: a production process tracing method based on large and small model collaboration, comprising the following steps:

[0055] Step S1: End-cloud collaborative data collection and feature extraction

[0056] (1) Device configuration and data collection on the end side

[0057] ① Sensor deployment:

[0058] Drilling machine:

[0059] Vibration sensor (model XYZ-2000, range ±2mm / s 2 , sampling rate 10kHz);

[0060] 5μm industrial camera (resolution 1920×1080, frame rate 30fps);

[0061] Temperature sensor (range 0-100℃, accuracy ±0.1℃).

[0062] Plating tank:

[0063] pH sensor (range 0-14, accuracy ±0.05);

[0064] Infrared thermal imager (resolution 640×480, temperature measurement accuracy ±1°C).

[0065] ②. Data synchronization:

[0066] Using the drilling machine spindle encoder pulse as the time reference, align the vibration waveform and image frame (synchronization error <1ms);

[0067] ③. Time series data segmentation: The vibration signal uses a sliding window (window length 200ms, overlap rate 50%).

[0068] (2) Lightweight feature extraction: The extracted feature vector Ve is compressed (the dimension is reduced from 1152 to 256) and uploaded to the cloud platform through the 5G private network.

[0069] Step S2: Dynamic knowledge graph construction and update

[0070] (1) Knowledge graph initialization

[0071] ①. Entity extraction:

[0072] A. Text parsing: The cloud-side BERT model extracts entities (such as "spindle bearing wear" and "aperture deviation") from historical maintenance logs and generates triplets (confidence level > 0.9).

[0073] B. Image analysis: ResNet-50 detects hole wall burrs (threshold area > 0.1mm 2 ), mark the defect type.

[0074] ②. Initial graph: Contains 4 types of nodes (equipment, parameters, defects, processes) and 12 edges (causal relationships, temporal associations).

[0075] (2) Dynamic update mechanism

[0076] ① New node creation: When a new high-frequency vibration pattern (5kHz) is detected on the client side, feature similarity is calculated:

[0077]

[0078] ② Edge weight optimization:

[0079] A. Based on TCN analysis of historical data, update the edge weight of "vibration anomaly → aperture deviation": (Original value 0.82)

[0080] Step S3: Bidirectional knowledge distillation and cross-process reasoning

[0081] (1) Cloud-to-end rule distillation

[0082] ①, Subgraph pruning: When the vibration abnormality index The cloud side extracts the associated subgraph (3 nodes, 2 edges);

[0083] ②. Device-side rule generation

[0084] IF vibration RMS>0.15mm / s 2 THEN:

[0085] IF spindle temperature ≤ 65°C → trigger aperture re-measurement (threshold ±5μm)

[0086] ELSE→Alarm code E101 (spindle overheating)

[0087] (2) End-to-cloud feature feedback

[0088] ①. The terminal detects the combined anomaly of "aperture deviation + uneven electroplating copper thickness" and calculates the cross-process correlation matrix: Mcorr[drilling, electroplating] = 0.87 (threshold > 0.8).

[0089] ②. Added edge "Aperture Deviation → Influence → Copper Thickness Uniformity" (initial weight Wij = 0.82)

[0090] Step S4: Multimodal traceability verification and optimization

[0091] (1) Root cause location

[0092] ① Path derivation: The cloud-side GAT model calculates the optimal path: Path score = 0.91 × 0.88 × 0.85 = 0.68 (threshold > 0.5)

[0093] ②Physical verification:

[0094] ANSYS simulations show that a 5 μm increase in pore size leads to a 12% increase in the plating solution flow rate, which deviates from the measured value by 3.2%;

[0095] Metallographic section analysis confirms the correlation between hole wall burr area and copper thickness CV value R 2 =0.89

[0096] (2) Process optimization

[0097] ①. Parameter adjustment: The drilling speed was reduced from 12krpm to 10.5krpm, and the RMS value of the spindle vibration decreased by 42%;

[0098] ② Model update: Add joint detection rules to the client-side model

[0099] IF the vibration frequency is in the 5kHz band AND the spindle temperature is >60°C → Early warning (code W202)

[0100] Implementation effect verification

[0101]

[0102] The above-described embodiments merely represent several implementation methods of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A production process tracing method based on collaboration between large and small models, characterized in that: include: Collect multi-source heterogeneous data through edge computing nodes and perform lightweight feature extraction; Build an updateable process knowledge graph based on large model analysis; Realize knowledge injection from large models to small models and experience feedback from small models to large models; Complete anomaly location and traceability report generation through end-cloud collaborative reasoning; Use health assessment and incremental learning mechanisms to continuously optimize graph quality.

2. The production process tracing method based on large and small model collaboration according to claim 1 is characterized in that: The end-cloud collaborative data collection includes: Deploy edge computing nodes to integrate structured data collectors, unstructured data collectors, and time series data collectors; A MobileNetV3+GRU hybrid network is used for hierarchical attention feature extraction, where image features are extracted through the SE module to extract channel attention, and time series features are extracted through causal convolution to extract 3-5kHz frequency band energy features.

3. The production process tracing method based on large and small model collaboration according to claim 1 is characterized in that: The construction of the dynamic process knowledge graph includes: Parsing unstructured data using the BERT-GAT architecture to generate entity-relationship triples, including text parsing and image parsing; A dynamic graph update mechanism is established. When the small model detects an unrecorded anomaly, a new node is created through feature similarity calculation, and a temporal convolutional network is used to adjust the edge weight.

4. The production process tracing method based on large and small model collaboration according to claim 1 is characterized in that: The bidirectional knowledge distillation includes: During forward injection, when the process abnormality index Ia>0.85, the associated subgraph is extracted, the top 20% weight edges are retained through graph attention pruning and converted into decision tree rules; During reverse feedback, the abnormal features of edge nodes are aggregated to construct a correlation matrix. When the cross-process correlation coefficient Mcorr is greater than 0.8, the graph edge is expanded and the GAT weight is updated.

5. The production process tracing method based on large and small model collaboration according to claim 1 is characterized in that: Production process traceability implementation includes: The client side quickly locates abnormal patterns by matching them with distillation rules, while the cloud side uses the GAT algorithm to calculate the cross-process impact path. Generate a weighted topology traceability report, highlight the critical paths with edge weights greater than 0.7, and calculate failure probability confidence intervals through Monte Carlo simulation.

6. The production process tracing method based on large and small model collaboration according to claim 1 is characterized in that: The closed-loop knowledge optimization includes: Calculate the graph health H every week. When H is less than 0.6, start PPO reinforcement learning to reconstruct the graph. Establish a data replay pool to store ≥10,000 abnormal cases, and implement a hybrid strategy of monthly full updates and daily incremental edge weight updates.

7. The production process tracing method based on large and small model collaboration according to claim 2 is characterized in that: The multi-source data collection includes: The vibration sensor collects the vibration signal of the equipment; Hyperspectral industrial cameras capture process surface morphology with a spatial resolution of 10μm / pixel; The equipment PLC interface reads structured parameters in real time and adopts a dual-threshold trigger mechanism.

8. The production process tracing method based on large and small model collaboration according to claim 3 is characterized in that: The knowledge graph construction includes: The entity disambiguation module calculates entity similarity using the BERT-wwm model and merges duplicate nodes when cos(v1,v2)>0.9; Define spatiotemporal constraints between processes and use graph comparative learning to optimize the embedding representation of newly added nodes.

9. The production process tracing method based on large and small model collaboration according to claim 4 is characterized in that: The knowledge distillation process includes: Subgraph pruning uses Grad-CAM heatmap to locate critical paths and retains edges with gradient contribution > 15%; When aligning features across processes, the dynamic time warping algorithm is used to compensate for timing differences, and a lag compensation factor is introduced to calculate the correlation.

10. The production process tracing method based on large and small model collaboration according to claim 5 is characterized in that: The generation of the traceability report includes: The A* algorithm is used to search the graph, and the path cost function is Cost = 0.6 × (1-edge weight) + 0.4 × node hop count; Implement multi-dimensional verification, including numerical verification with a historical library matching degree of >85% and physical verification through COMSOL simulation calculations.

Images