Mechanism-AI Fusion for Intelligent Monitoring and Diagnosis of Precision Forming Processes
Targeting precision forming scenarios such as stamping and injection molding, multi-source time-series signals are co-modeled with physical mechanisms to close the loop from weak-anomaly online monitoring to root-cause analysis
Project Overview
Production disturbances readily translate into product defects: weak anomalies are difficult to catch in time, leading to recurring batch-level defects, while manual sampling inspection lags behind and wastes both material and time. This project acquires multi-dimensional time-series signals — acceleration, acoustic emission, stress-strain, displacement, pressure, temperature, and flow — and fuses them with physical mechanisms to build a closed-loop "monitor–identify–diagnose" system, now deployed on a seat-slide-rail production line at an automotive OEM and a housing injection-molding line at a leading 3C manufacturer.
Research Objectives
Co-modeling of multi-source time-series signals and physical mechanisms
Real-time time-series anomaly monitoring informed by physical priors
Closed-loop diagnosis spanning monitoring, identification, and root-cause analysis
Lightweight, full-pipeline online deployment
Methodology
A physics-prior-injected time-series modeling pipeline: mechanism models constrain the feature space while deep time-series networks amplify and separate weak anomalies; for injection molding, a dedicated chain for precise anomaly monitoring and root-cause analysis was built.
Technical Approach: How It Works
- 1
Multi-Source Sensing and High-Frequency Synchronized Acquisition
Acceleration, acoustic-emission, stress-strain, displacement, and pressure/temperature/flow sensors are deployed on the stamping/injection-molding equipment, synchronously acquired at high sampling rates to form high-dimensional time-series (profile) data spanning an entire forming cycle.
- 2
Physically-Informed Feature Construction and Prior Embedding
Hybrid feature construction: after low-pass filtering, both physical-mechanism features (based on forming-mechanics models) and global statistical features are extracted; physical priors then assign each candidate feature a confidence prior (process-critical / uncertain / irrelevant features are initialized with distinct Beta distributions), turning expert knowledge into a learnable probabilistic prior.
- 3
Bayesian Bi-Level Feature Optimization
The upper level performs feature-architecture learning — a Gumbel-Softmax continuous relaxation enables differentiable search over the discrete feature-subset space, converging to the optimal subset under multi-criteria objectives (discriminability, redundancy, stability); the lower level performs multi-model performance estimation across heterogeneous classifiers (linear / neural-network / kernel methods), ensuring the selected features are robust and model-agnostic.
- 4
Online Monitoring and Root-Cause Analysis Closed Loop
The optimized lightweight model is deployed online end-to-end, issuing millisecond-level judgments for every stamping stroke or molding cycle; once an anomaly triggers, mechanism templates automatically localize the root cause (skipped material, pitting, punch fracture, etc.), forming a "monitor–identify–diagnose–act" closed loop.
Figures: Methods & Results
Key Results
Detects anomalies as small as 0.02mm, with lightweight full-pipeline deployment
Large-batch anomaly rate cut to 0.22 incidents per 100,000 units, a 97.35% reduction
Customer-confirmed outgoing defect rate reduced by 78%
Recall rate >98% for skipped-material, pitting, and punch-fracture anomalies
Related Publications
Physically-informed Bayesian feature optimization for semi-supervised industrial anomaly detection, Measurement Science and Technology, 2026
Profile Abstract: An optimization-based subset selection and summarization method for profile data mining, IEEE Transactions on Industrial Informatics, 2022
Intelligent Monitoring of Manufacturing Processes