In traditional manufacturing, quality control operates retrospectively. A part is cast, printed, or machined, and only after it is completed does it undergo inspection. This approach introduces a fundamental vulnerability: post-process inspection latency.
When inspection happens after the fact, the manufacturing system operates as a black box. If a mechanical fault occurs early in a production run, it remains undetected until the process finishes, leading to massive material waste, expensive manual rework, or catastrophic batch failures. To eliminate this latency, a rapidly growing body of research focuses on in-situ defect detection—shifting quality assurance from a reactive, post-process evaluation to an active, real-time, in-process mechanism.
The Cost of Looking Backward
Why do manufacturers accept inspection latency? Historically, the highest-resolution diagnostic tools—such as scanning electron microscopy, X-ray computed tomography, or precise coordinate metrology—are inherently ex-situ. As Wang et al. note in their work on laser materials processing, these high-fidelity techniques require removing the fabricated part from the machine environment, creating an insurmountable feedback delay.
This delay forces undesirable operational trade-offs across diverse domains:
- Industrial Scale: Slow measurement processes make 100% inspection impossible, forcing manufacturers to rely on statistical sampling, which allows localized machining errors to slip through to final assembly (Wang).
- Heavy Industry: In blast furnace ironmaking, critical quality indicators often require manual sampling and laboratory testing, preventing real-time closed-loop control and causing energy inefficiencies (Luo et al.).
- Bio-Manufacturing: Standard analytical workflows for monitoring cell states are performed offline, preventing operators from making immediate adjustments to culture conditions to maintain optimal cell states (Azari-Pour et al.).
- High-Throughput Computing: Even in digital environments like high-energy physics simulations, waiting months for retrospective data validation can result in millions of wasted CPU hours (Harenberg et al.).
The core insight of modern in-situ research is that we do not need to move the electron microscope inside the machine. Instead, we can treat the manufacturing environment as a dynamic system whose internal states can be inferred from real-time, surface-level emissions.
Reconstructing Reality via Sensor Fusion
The primary challenge of in-situ monitoring is that manufacturing environments are harsh, noisy, and incredibly fast. A single sensor stream is rarely enough to distinguish a true defect from environmental noise. In-situ systems solve this by using multi-sensor data fusion to build a composite, real-time representation of the process physics.
Consider laser welding or additive manufacturing. By combining sensors that capture different physical phenomena—such as plasma emissions, light reflections, acoustic waves, and thermal radiation—machine learning models can cross-reference signals to isolate anomalies.
[Physical Process] ───┬───> Plasma Sensor ──────┐
├───> Optical Camera ─────┼───> [Attention-Based Fusion] ───> Real-Time Health Estimate
└───> Thermal/Acoustic ───┘
For instance, Deng developed a system called IDDNet (A Multi-Sensor Data Fusion System for Laser Welding Process Monitoring) that fuses plasma, light, and temperature signals. Because the model processes these combined inputs in under 0.8 milliseconds, it can flag defects fast enough to match high-speed industrial production lines. Similarly, Wang combined visible light, infrared, and ultrasonic sensors with an attention-based fusion algorithm (Development of an Online Defect Detection System for Additive Manufacturing Based on Multi-Sensor Fusion), demonstrating that multi-spectral awareness can slash defect rates from 18.5% to 5.2% via closed-loop intervention.
Crucially, this approach does not always require expensive instrumentation. When paired with intelligent feature extraction, low-cost sensors can achieve comparable results. Han demonstrated that a standard, consumable camera analyzing "spatter" patterns (sparks and droplets) in Selective Laser Melting (SLM) can classify defect-inducing conditions with 97% accuracy (A novel image feature descriptor for SLM spattering pattern classification using a consumable camera). In battery manufacturing, Tian et al. built WeldMon (WeldMon: A Cost-effective Ultrasonic Welding Machine Condition Monitoring System), utilizing a $120 audio hardware setup to track ultrasonic welding health on the edge, bypassing the need to halt production lines for offline tool measurements.
The Inversion Operator: Inferring Hidden States
How do we detect an internal, subsurface defect—like porosity inside a metal part or micro-cracks in a brittle resonator—using only surface observations?
The answer lies in treating machine learning as an inversion operator. The physical process maps an internal state (e.g., melt pool depth, material density) to an observable surface emission (e.g., thermal radiation, vibration). A well-trained neural network can invert this mapping, decoding the hidden internal state from the surface data in real time.
Surface-to-Subsurface Inference
In metal 3D printing, "lack-of-fusion" defects happen beneath the surface layer, making them invisible to standard optical cameras. To expose these hidden flaws, Ogoke et al. designed a hybrid deep learning framework (Deep Learning for Melt Pool Depth Contour Prediction From Surface Thermal Images via Vision Transformers). By pairing a ResNet (to capture spatial surface features) with a Vision Transformer (to capture temporal dynamics), their model predicts the hidden, subsurface depth contour of the melt pool using only surface thermal videos. This allows the system to flag internal structural failures before they are permanently buried under subsequent layers of metal powder.
Cross-Domain Mapping
Sometimes the challenge is not looking beneath the surface, but making sense of a highly distorted, off-angle perspective inside a machine. Niu et al. addressed this by using a Semi-Siamese network architecture (Semi-Siamese Network for Robust Change Detection Across Different Domains with Applications to 3D Printing). By utilizing unshared encoders, the network maps noisy, real-time camera images directly to idealized digital schematics, performing pixel-level change detection that remains robust to shifting lighting conditions and camera angles.
Virtual Sensing (Soft Sensors)
When physical sensors cannot survive the extreme environments of heavy manufacturing, data-driven virtual instruments—or soft sensors—step in. As surveyed by Luo et al. (Data-driven soft sensors in blast furnace ironmaking: a survey), soft sensors use easily measurable variables (like external temperature or gas utilization) to estimate hard-to-measure internal variables in real-time, bypassing manual laboratory delays.
This principle applies equally to high-precision component fabrication. Alexos et al. demonstrated that by pairing raw vibration signals with unsupervised domain adaptation (Nuclear Fusion Diamond Polishing Dataset), machine learning models can predict nanometer-scale surface roughness in real-time, eliminating hours of manual, labor-intensive inspection halts.
From Reactive Detection to Proactive Control
The ultimate goal of eliminating inspection latency is not merely to stop a machine when a failure occurs, but to adjust process parameters dynamically to prevent the failure from happening at all. This requires moving from reactive detection to proactive forecasting and feedforward control.
Proactive Forecasting
Instead of identifying a defect after it forms, advanced in-situ models analyze the stability of process dynamics to predict failures before they manifest. A prime example is the VIMPS framework developed by Hassan et al. (Monitoring Variability in Melt Pool Spatiotemporal Dynamics (VIMPS): Towards Proactive Humping Detection in Additive Manufacturing). By tracking the statistical variability in melt pool solidification dynamics, VIMPS can predict geometric "humping" defects up to 10 seconds before they physically occur, giving the control system a window to adjust laser power or speed.
Digital Twins and Feedforward Control
To sustain this level of proactive intervention across complex, multi-stage manufacturing lines, researchers rely on Digital Twins to maintain spatial and temporal context. Reisch et al. demonstrated that integrating machine, temporal, and spatial data into a Digital Twin framework (Context awareness in process monitoring of additive manufacturing using a digital twin) allows for the detection of complex, propagating anomalies—such as "pore nests"—that single-point sensors miss entirely.
When these real-time predictive models are embedded into control loops, they enable active compensation. Upadhyay et al. introduced a feedforward control scheme (Deep Koopman-based Control of Quality Variation in Multistage Manufacturing Systems) that leverages a Stochastic Deep Koopman model. By linearizing complex, non-linear manufacturing dynamics in real-time, this framework actively compensates for quality variations as a part moves through successive stages of production.
[In-Situ Sensors] ───> [Digital Twin / Koopman Model] ───> [Predictive Adjustment] ───> [Actuator Control]
▲ │
└─────────────────────────────── Real-Time Feedback Loop ───────────────────────────────┘
By treating quality control as a real-time, data-driven inversion problem, in-situ defect detection dismantles the traditional separation between fabrication and inspection. The result is a manufacturing paradigm that no longer documents failures after the fact, but actively works to prevent them.
Key Foundational Work
Robust Defect Segmentation in Woven Fabrics (1998) — contributed 1 concept elements to this direction.
Estimation of Elution Induce Time and 80% Elution Time of Polymer-Coated Manure by a Neural Network and Ck Interpolation Scheme. Development of a Neural Network Simulator for Structure-Activity Correlation of Molecules: Neco (5). (2001) — contributed 1 concept elements to this direction.