The study, led by Giulia Gorla of the University of the Basque Country together with Jon Ander Iturrioz, Kim Esbensen and Jose Manuel Amigo, takes aim at what the authors call the persistent misconception that data quantity alone ensures representativity. Their central message is blunt: advanced sensor technology and powerful data modelling cannot compensate for systematic measurement bias introduced at the moment of sampling. Data quality, they insist, originates upstream, before analysis, and the mathematical machinery of chemometrics can only extract information that is physically present in the measurements.

To make this case, the authors turn to the Theory of Sampling, a rigorous framework developed originally for mining and bulk materials by Pierre Gy and later expanded by Esbensen and colleagues. The theory classifies materials by their dimensionality, from zero-dimensional discrete objects like tablets to three-dimensional volumes such as reactors and silos, and identifies the specific errors that arise in each case. Constitutional heterogeneity reflects differences in composition between particles, distribution heterogeneity arises from segregation and clustering, and temporal heterogeneity captures fluctuations over time. In dynamic systems, flow-induced heterogeneity adds further complications through velocity gradients and varying residence times in vessels. Unless a sampling strategy explicitly counteracts these heterogeneities, the collected data will reflect the sampling geometry rather than the true process.

The authors demonstrate this with three striking datasets. The first involved a natural sodalite rock from Brazil examined with a portable Raman spectrometer. Fifteen spectra collected from different points on the rock’s two faces revealed dramatic differences: some spots showed pure sodalite signatures, while others carried clear evidence of calcite, a common inclusion. The lesson is that a single surface measurement can completely miss the three-dimensional character of a heterogeneous material, a situation directly analogous to industrial probes that interrogate only a thin layer of a powder bed or slurry.

The second dataset used short-wave infrared hyperspectral imaging of ground coffee deliberately adulterated with chicory or malt. When the components were left unmixed, principal component analysis readily detected both the adulterant and an added salt marker. Once the samples were homogenized, discrimination became far harder, partly because salt has no distinct near-infrared signal of its own and mainly registers through its water of hydration. Even more revealing, the calculated homogeneity curves depended strongly on which region of the image was analysed: the same mixed sample appeared markedly more inhomogeneous when a smaller region of interest was chosen. Field of view, the authors emphasize, is not a technical afterthought but a decisive factor in whether a measurement claims to represent the bulk.

The third dataset came from beer. Drawing on 377 near-infrared spectra from seven batches of Belgian Pale Ale spanning mashing, boiling, fermentation and carbonation, the team built multivariate statistical process control models. Even a sampling frequency of once per day was enough to track the general arc of fermentation, with the first principal component capturing time evolution. But two abnormal batches drifted differently, and detecting such anomalies early would demand higher temporal resolution. The work also exposed an uncomfortable truth: in the early brewing stages the variability was so large that no clear trends could be distinguished among batches, and the researchers could not tell whether this stemmed from sampling, composition, or genuine chemical information. Poorly planned sampling in timing, location or sensor choice distorts model performance and can fabricate misleading cause-and-effect relationships when transport lags delay the recorded signal.

Perhaps the most provocative part of the paper is its critique of fashionable chemometric remedies. The authors tested SMOTE, a widely used synthetic data augmentation technique, on the beer monitoring models. Synthetic spectra generated by interpolation between existing observations did inflate the training set, yet the geometric structure of the monitoring space remained essentially unchanged. The out-of-control samples still fell outside the same limits. The explanation is geometric: projection-based models are built from the covariance structure of real data, and interpolating between points adds no new directions of variance. Regions of the process that were never measured remain unobserved, and heterogeneity that never reached the sensor cannot be reconstructed algorithmically. Worse, artificial densification can give a false sense of coverage and inflate validation statistics without ever testing the model’s ability to extrapolate to genuinely new conditions.

The review is equally skeptical of Interactive Optimization Technologies, promoted recently as a lean alternative to traditional partial least squares calibration. By using spectra of pure raw materials to track blend homogeneity in pharmaceutical manufacturing, these methods slash the calibration burden, a real advantage when active ingredients are scarce or expensive. But the authors note that this is essentially a return to multiple linear regression on pure component spectra, an approach abandoned decades ago because real industrial and natural samples are never pure components. The verdict: useful in the pharmacopeia-driven world of pharmaceutical manufacturing, but no universal panacea.

On the quantitative side, the authors integrate sampling into the framework of measurement uncertainty, distinguishing the total sampling error, the total analytical error, and a newly proposed total data modelling error that captures preprocessing artefacts, overfitting and model misspecification. The hierarchy is stark: sampling error often dominates the analytical error by factors of five to ten, and in highly heterogeneous materials by up to twenty-five. The greatest gains in reliability therefore come from better sampling, not better algorithms. Practical tools exist to quantify this: replication experiments that duplicate the entire lot-to-aliquot pathway, heterogeneity tests, variograms and homogeneity curves, all codified in standards such as the Danish DS 3077.

The authors close with concrete recommendations. Sensor type, position, field of view and penetration depth should be characterized against the physics of the process, not chosen by convenience. Sampling frequency should match process dynamics, since oversampling merely adds redundancy while undersampling misses critical events. Sampling design should be embedded in Quality by Design workflows and evaluated during risk analysis, as regulators are increasingly expected to demand. Emerging technologies, from adaptive sampling that lets exploratory models decide where to measure next, to digital twins that simulate processes virtually, can help, but only when built on informative primary sensors that comply with the Theory of Sampling. In regulated sectors like pharmaceuticals, where multivariate models define the design space under ICH guidelines, sampling representativity is not best practice but a prerequisite. The paper’s conclusion is a warning worth repeating across every industry drowning in sensor data: analysis can refine what was measured, but it can never recover information that was never observed.

Subject of Research: The role of the Theory of Sampling in ensuring representative data acquisition for reliable Process Analytical Technology and chemometric process monitoring

Article Title: Getting “sampling” right for reliable process monitoring: main challenges and solutions

Article References: Gorla, G., Iturrioz, J. A., Esbensen, K. H., & Amigo, J. M. (2026). Getting “sampling” right for reliable process monitoring: main challenges and solutions. Frontiers in Analytical Science, 6, Article 1839552. https://doi.org/10.3389/frans.2026.1839552

Image Credits: AI Generated

DOI: 10.3389/frans.2026.1839552

Keywords: Process Analytical Technology, Theory of Sampling, chemometrics, representative sampling, near-infrared spectroscopy, Raman spectroscopy, hyperspectral imaging, data augmentation, measurement uncertainty, heterogeneity, process monitoring, Quality by Design

Friction plug welding is a solid-state technique used to repair high-performance aluminum alloy structures, particularly thick 2219 aluminum alloy components common in aerospace applications. During the process, a rotating plug is forced into a pre-machined hole in the workpiece. Friction between the plug and the hole’s sidewall generates intense localized heating, softening the material into a thermoplastic state. Once the plug reaches a prescribed depth, rotation stops and a heavy forging load consolidates the bond. Because the material never melts, the classic fusion-welding problems of porosity, segregation, and hot cracking are avoided. The process also produces dynamically recrystallized, fine equiaxed grains near the interface, a narrow heat-affected zone, low residual stress, and requires no filler metal, flux, or radiation shielding.

Yet friction plug welding has an Achilles heel: the quality of the bond is notoriously difficult to verify. The mechanics depend on sidewall friction, so the applied axial force is not transmitted directly perpendicular to the bonding interface, and the softened metal’s plastic flow is hard to control precisely. Local regions can receive too little or too much material flow, producing lack-of-bonding defects that are hidden beneath extruded flash. Worse, different depths within a thick joint experience very different histories of frictional contact and heat accumulation. The lower interface, which engages friction earlier and for longer, is especially prone to solute enrichment and the formation of brittle eutectic phases, while insufficient local deformation can leave microscopic voids known as kissing bonds. These weak spots are nearly invisible to non-destructive testing and can cut tensile strength by 50 to 100 megapascals depending on location through the thickness.

The stakes are illustrated starkly by prior fatigue research. Friction stir welds in 5083 aluminum alloy containing kissing bond defects showed fatigue lives 21 to 43 times shorter than sound welds, with characteristic strength at two million cycles dropping from 100.24 to 65.57 megapascals. In friction stir spot welded aluminum-copper joints, an interfacial Al2Cu eutectic layer provided a preferential crack path and reduced fatigue strength to a mere 2.0 megapascals at ten million cycles. Signal-based monitoring has already proven itself in friction stir welding and additive friction stir deposition, where machine learning models built on force and torque features have achieved defect detection accuracies exceeding 95 percent. The Tianjin-led team set out to bring the same diagnostic power to friction plug welding.

The researchers welded 20-millimeter-thick 2219 aluminum alloy plates using 25 different combinations of plug taper angle, ranging from 12 to 20 degrees, and pre-machined hole taper angle, from 24 to 32 degrees, while recording axial force and torque at 30 hertz. The resulting cross-sections revealed three distinct forming states. Small plug angles paired with large hole angles produced upper lack-of-bonding defects with no visible flash. Large plug angles with small hole angles shifted the defects to the bottom of the joint. Only moderately matched geometries yielded macroscopically defect-free joints with balanced flash extrusion on both surfaces.

The crucial discovery was that each defect type imprints a characteristic signature on the force and torque curves. Joints with upper lack-of-bonding defects showed axial force and torque rising continuously throughout the rotational friction phase, because the frictional contact region never expanded through the full workpiece thickness, so resistance kept climbing as the plug fed deeper. Numerical simulations showed the uncontacted upper region stayed at roughly 333 degrees Celsius while the contact zone heated to nearly 545 degrees, far too cool for adequate plasticization. Lower defects told the opposite story: after force and torque peaked and stabilized, both signals declined. Simulations revealed that the lower plug region, having endured friction longest, reached about 572 degrees Celsius, softened, necked, and locally separated from the workpiece, exactly the contact-separation behavior that produces the rise-stabilization-decrease pattern.

For the ten macroscopically sound joints, the team extracted variables from the steady friction stage and found they correlated tightly with layer-dependent strength. Average axial force during the steady stage, governed mainly by plug taper angle, rose from 34.92 to 50.11 kilonewtons as the angle increased from 12 to 20 degrees, and this force proved decisive for the upper interface. Higher force raised normal stress from roughly 26.7 to 62.7 megapascals and interfacial strain from 0.69 to 1.07 millimeters per millimeter, compressing shut micropores, enabling dynamic recrystallization across the interface, and eliminating kissing bonds entirely once force exceeded about 43.55 kilonewtons. Upper-layer ultimate tensile strength climbed correspondingly from 246.1 to 335.4 megapascals, while middle and lower layers stayed largely indifferent to force.

Heat told the reverse tale at the bottom of the joint. The team defined an effective heat input combining torque and friction duration, which was dominated by the contact angle between plug and hole. Once effective heat input exceeded roughly 0.62 times ten to the fifth joules, Al-Al2Cu eutectic networks began forming at the lower interface and widened up to 14.12 micrometers as heat input rose to its maximum of 1.54 times ten to the fifth joules. Electron backscatter diffraction showed interfacial temperatures climbing from 521 to 562 degrees Celsius, driving fragmentation of Al2Cu particles and copper diffusion toward the interface until the eutectic reaction became thermodynamically favorable. During tensile loading, the hard eutectic phases created stress concentrations and brittle, rock-candy-like fracture surfaces, dragging lower-layer strength down from 323.8 to 238.3 megapascals as heat input increased.

The team then tested whether simply adjusting axial feed could fix defective joints. Increasing feed eliminated upper lack-of-bonding defects and produced a steady friction stage, while decreasing feed suppressed lower defects, confirming the link between plug position and macroscopic quality. But feed adjustment alone proved insufficient: joints freed of upper defects still harbored kissing bonds because steady-stage force remained too low, and joints freed of lower defects accumulated excessive heat and wide eutectic regions. The solution was a synergistic strategy, selecting a joint geometry, a 16-degree plug taper with a 32-degree hole taper and 8-degree contact angle, that promised high steady-stage force with low heat input, then tuning axial feed to 12 millimeters. The optimized joint achieved 43.85 kilonewtons of average steady force with only 0.19 times ten to the fifth joules of heat input, eliminated both kissing bonds and lower-interface eutectic, and delivered tensile strengths above 310 megapascals in every layer.

Beyond its immediate implications for repairing rocket tanks and airframe structures, the framework establishes force and torque signals as quantitative indicators of bonding quality in friction plug welding for the first time, closing a gap that had limited the technique’s adoption. The authors caution that critical thresholds for kissing bonds and eutectic formation will vary with joint thickness, geometry, and alloy system, and that optimal axial feed still requires determination for each configuration. But the trajectory is clear: the same signal-based monitoring that revolutionized friction stir welding could soon enable online quality assessment and dynamic parameter regulation in plug welding, with lessons extending to additive friction stir deposition and other solid-state processes where the decisive action happens far below the surface, hidden from every camera and probe.

Subject of Research: Signal-based monitoring of joint formation and interfacial bonding quality in friction plug welding of thick 2219 aluminum alloy

Article Title: Relationship of axial force–torque responses with joint formation behavior and interfacial bonding quality in friction plug welding of thick aluminum alloy

Article References: Relationship of axial force–torque responses with joint formation behavior and interfacial bonding quality in friction plug welding of thick aluminum alloy. (n.d.). https://doi.org/10.1007/s44500-026-00022-z

Image Credits: AI Generated

DOI: 10.1007/s44500-026-00022-z

Keywords: friction plug welding, 2219 aluminum alloy, axial force, torque, kissing bond, lack-of-bonding defects, Al-Al2Cu eutectic, dynamic recrystallization, solid-state welding, aerospace structures, process monitoring, tensile strength

The research, carried out by Xiaofei Cui of TWI Limited together with Akram Zitoun and Sadik Omairey of the Brunel Composites Centre at Brunel University London, examines a photonic-integrated sensor system developed under the European Union-funded SEER project. The sensor is designed to be embedded directly within composite tooling, where it monitors refractive index, temperature, and pressure in real time. These three parameters are the vital signs of composite production: refractive index reveals how the resin is flowing and curing, temperature governs the chemistry of the cure cycle, and pressure indicates whether the part is being consolidated properly. Because the sensor sits inside the tool surface, it can detect anomalies as they happen rather than after the part has been removed and inspected, offering the prospect of fewer scrapped components and more repeatable processes.

At the heart of the device is a photonic-integrated circuit, or PIC, fabricated using silicon-on-insulator technology. The prototype chips were produced through a multi-project wafer run, starting from a 725-micrometre silicon substrate topped with a 2-micrometre buried oxide layer. Two types of silicon waveguides, each 220 nanometres thick, guide light through the chip: a strip waveguide and a rib waveguide with a 70-nanometre etched slab. A 520-nanometre low-temperature oxide layer and a thin silicon nitride cap complete the stack. Light travelling through these waveguides interacts with its surroundings, so changes in the refractive index of the resin, or in temperature and pressure at the sensor surface, shift the optical signal in ways that can be read out with high precision. The chip is then post-processed, with reactive ion etching or wet etching used to create the refractive index sensing features and anisotropic etching or substrate bonding used to form the pressure-sensing membrane.

What makes the study unusual is that the authors did not stop at demonstrating that the sensor works. They built a combined life cycle costing and life cycle assessment framework to evaluate the device from prototype fabrication through to industrial scale-up. The economic side used a hybrid bottom-up and activity-based costing model, in which the total cost of production is the sum of the costs of every individual manufacturing activity, each broken down into labour, materials, energy, and equipment. The environmental side was modelled in GaBi software in accordance with the ISO 14040 and ISO 14044 standards, taking a single sensor as the functional unit and covering raw material supply, manufacturing, distribution, use, and end-of-life pathways including reuse and landfill.

The cost findings are striking. Across the prototype stage, labour was the dominant expenditure in both recurring and non-recurring costs, dwarfing the contributions of materials, energy, and equipment. At the process level, assembly dominated the recurring costs, substantially exceeding PIC fabrication, testing and validation, and coating. This is not surprising when the assembly sequence is examined in detail: the PIC with its micro-lens must be manually inserted into a metal adapter, a separate lens mounted on a ceramic ferrule holding the optical fibre must be aligned using a six-degree-of-freedom positioner, and each joint must be secured with epoxy and cured. These delicate optical alignment steps are precisely the operations that photonic packaging engineers identify as the bottleneck in moving integrated photonics from the laboratory to volume production.

To understand how costs might evolve as production scales up, the team applied learning curve analysis, a well-established manufacturing economics technique that captures how labour hours per unit fall as workers gain experience and processes mature. The analysis assumed learning rates of 80, 85, and 87 per cent, drawn from the literature, and revealed a pattern familiar from aircraft and semiconductor manufacturing. Early gains are dramatic: at low volumes, suitable investment can improve productivity by around one per cent per unit. But the improvements saturate quickly. After the 191st unit, the potential gain drops to just 0.1 per cent, and after the 365th unit it falls to 0.05 per cent. Beyond that endpoint, further investment in training and tooling will not pay for itself, a conclusion with direct consequences for how manufacturers should time their automation investments.

The sensitivity analysis added a geopolitical dimension to the economics. Labour accounted for 77 per cent of total recurring costs in the scale-up scenario, so regional wage differences matter enormously. Applying Eurostat data from 2022, which showed average hourly labour costs ranging from €8.2 in Bulgaria to €50.7 in Luxembourg and €55.6 in Norway, the researchers found that in high-wage countries labour represents more than 80 per cent of total production costs, while in lower-wage regions it falls to around 40 per cent. For a technology whose commercial viability hinges on labour-intensive assembly, decisions about where to manufacture and how much to automate could determine whether embedded photonic sensing reaches the factory floor at all.

The environmental assessment told a complementary story. PIC fabrication and post-processing emerged as the most energy-intensive stages, together accounting for more than 63 per cent of the total energy required to manufacture and use the sensor, a burden attributed to the energy-hungry clean-room operations and the complexity of photonic device fabrication. The same fabrication stage also returned the highest global warming potential and the highest ozone depletion rate in the life cycle impact results, partly because prototype production was performed manually and involved multiple research iterations. At the other extreme, sensor testing and validation consumed less than two per cent of the total energy, using small, low-emission equipment. Downstream processes such as coating, packaging, and integration contributed minimally to the overall footprint.

Perhaps the most valuable insight of the study is that the economic and environmental hotspots largely coincide. The labour-intensive assembly operations that dominate recurring costs and the PIC fabrication and post-processing stages that dominate energy demand and emissions are the same targets that automation, wafer-level batch processing, and improved first-time-right manufacturing would address. The authors recommend focusing automation specifically on the manual insertion, alignment, epoxy dispensing, and packaging steps; prioritising reductions in repeated prototype iterations during chip fabrication; comparing etching routes not only technically but in terms of energy demand and yield; and concentrating investment in productivity improvements before the learning curve saturates. They also caution against over-investing in environmental improvements to testing and validation, which is simply too small a contributor to matter.

The broader significance extends beyond one sensor design. As industries under pressure to decarbonise turn to lightweight composites, and as smart manufacturing promises to squeeze waste and defects out of production, embedded sensing is likely to become standard practice. But this study demonstrates that the sustainability case for such technologies cannot be assumed; it must be quantified. By showing that cost and carbon concentrate in the same early manufacturing stages, and that learning curves, labour rates, and automation strategies jointly determine commercial viability, the researchers have provided a reusable framework for evaluating emerging photonic sensing technologies as they make the difficult transition from research prototype to industrial product. For composite manufacturers weighing whether to embed intelligence in their tooling, the answer, it turns out, depends as much on economics and energy as on optics.

Subject of Research: Environmental and economic life cycle assessment of an embedded photonic sensor system for real-time composite manufacturing monitoring

Article Title: Environmental and economic assessment of an embedded photonic sensor system for composite manufacturing monitoring

Article References: Cui, X., Zitoun, A., & Omairey, S. (2026). Environmental and economic assessment of an embedded photonic sensor system for composite manufacturing monitoring. Clean Technologies and Environmental Policy, 28(10), Article 265. https://doi.org/10.1007/s10098-026-03614-5

Image Credits: AI Generated

DOI: 10.1007/s10098-026-03614-5

Keywords: composite manufacturing, photonic sensors, photonic integrated circuits, life cycle assessment, life cycle costing, learning curve, industrial automation, sustainability, silicon photonics, process monitoring, labour costs, clean technology

The research, led by Bei Sun with co-authors Yao Wu, Peng Kong and Yijun Wang, addresses a long-standing tension in industrial artificial intelligence. Data-driven models can be remarkably accurate, but they often behave as black boxes, offering predictions without reasons. Mechanistic models built from first principles are interpretable, but they struggle to capture every quirk of a messy, uncertain plant environment. The team’s answer is a hybrid: a classification pipeline that fuses spatial and temporal information about the process, optimizes itself with Bayesian methods, and then anchors its clustering results back to the underlying chemistry and physics of the plant.

At the heart of the challenge is what the authors describe as spatio-temporal coupling. In a complex industrial process, variables are not independent readings on a dashboard. They form a web of cause-and-effect relationships shaped by chemical reactions, the physical flow of materials between units, and disturbances from the outside environment, such as ambient temperature swings or fluctuations in feed composition. These relationships unfold over time, so the state of the process now depends on both the current configuration of variables and the history of how that configuration has evolved. Traditional classification methods, which typically treat each sensor reading or each time window in isolation, lose much of this structure.

To capture the spatial side of the problem, the framework employs graph neural networks, a class of models that learn from relationships rather than raw vectors. By representing industrial variables as nodes in a graph and their physical or statistical interdependencies as edges, the network can extract features that reflect how a change in one part of the plant propagates to another. This is a meaningful departure from conventional approaches that flatten all variables into a single list, discarding the topology that makes an industrial process coherent. Graph-based learning has gained traction across science and engineering in recent years precisely because so many real systems, from molecules to power grids, are naturally relational.

The temporal dimension is handled by a combination of state space models and attention mechanisms. State space models, which have recently surged in popularity in machine learning, describe a system through hidden states that evolve over time, offering an efficient way to model long sequences of data. Attention mechanisms, best known from their role in transformer architectures, allow the model to weigh which moments in the process history matter most for the current classification decision. The spatial features extracted by the graph network are fed into this temporal pipeline as a form of compensation, enriching the time-series representation with structural context. The result is an integrated spatio-temporal feature set that encodes both where the process stands in its network of variables and how it got there.

These fused features then flow into a Self-Organizing Map, a classic neural network technique dating back to the 1980s that projects high-dimensional data onto a low-dimensional grid while preserving topological relationships. Self-Organizing Maps are prized in industrial applications because their neighborhoods are visually interpretable: similar operating conditions land near each other, making the resulting clusters easier to inspect. The twist here is Bayesian optimization, a sample-efficient strategy for tuning hyperparameters that would otherwise require exhaustive trial and error. By searching the configuration space intelligently, the Bayesian-optimized SOM achieves more reliable and accurate classification of operating conditions than a naively configured version would.

Interpretability, the study’s central selling point, enters at the final stage. The framework is built around a comprehensive state-space description of the process that integrates inlet conditions, reaction conditions and output states. This structure ensures the model sees the full breadth of operating states represented in the collected data, from raw material entry to final product characteristics. When the SOM produces its clusters, the researchers map those clusters back onto process mechanisms, giving each group of operating conditions a physical reading. Instead of an opaque label such as cluster three, an engineer can see that a particular region of the map corresponds, say, to a specific balance of feed properties and reaction environment. That kind of translation from statistical pattern to process meaning is exactly what has been missing from many data-driven monitoring systems.

The team demonstrated the approach on an arsenic salt cobalt removal process, a hydrometallurgical operation used in zinc purification where cobalt impurities are precipitated with the help of arsenic salts. Such processes are notoriously difficult to monitor because reaction kinetics, additive dosing and solution chemistry interact in ways that shift the plant between subtly different operating regimes. According to the authors, the proposed method showed both effectiveness and interpretability under the evaluated conditions, correctly partitioning the process into meaningful operating states while remaining transparent about why each classification was made. The researchers are careful to note, however, that practical deployment and broader applicability to other processes still require further validation, a candid caveat that reflects the gap between a strong case study and proven industrial technology.

The significance of this work extends beyond one metallurgical process. Industries from steelmaking to pharmaceuticals are under pressure to squeeze more efficiency out of existing plants while meeting stricter environmental and safety standards, and intelligent monitoring is widely seen as a key enabler. Yet adoption has been slowed by the trust problem: operators are reluctant to act on recommendations from models they cannot interrogate. By combining modern deep learning components, graph networks, state space models and attention, with older, more transparent tools like self-organizing maps and mechanism-based interpretation, the Central South University framework offers a template for building AI that speaks the language of the plant floor. If follow-up work confirms its generality, interpretable spatio-temporal classification could become a standard layer in the digital infrastructure of complex industry, turning the flood of sensor data into decisions that engineers can understand, question and, ultimately, trust.

Subject of Research: Interpretable spatio-temporal machine learning for classification of operating conditions in uncertain complex industrial processes

Article Title: An interpretable spatio-temporal fusion operating conditions classification model for uncertain complex industrial processes

Article References: Sun, B., Wu, Y., Kong, P., & Wang, Y. (2026). An interpretable spatio-temporal fusion operating conditions classification model for uncertain complex industrial processes. Applied Intelligence, 56(15), Article 434. https://doi.org/10.1007/s10489-026-07443-3

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07443-3

Keywords: machine learning, industrial processes, spatio-temporal fusion, self-organizing map, graph neural networks, state space models, attention mechanisms, Bayesian optimization, interpretability, process monitoring, cobalt removal, clustering

Researchers at Chungnam National University in Daejeon, South Korea, working with a colleague at Gadjah Mada University in Indonesia, set out to test whether a short-wave infrared (SWIR) hyperspectral imaging system could reliably predict moisture content in pork meat emulsions while simultaneously revealing how uniformly the batter had been mixed. Their system swept across 1,000 to 1,700 nanometers — a range where water, fat, and protein all leave distinctive spectroscopic fingerprints — and, combined with machine learning models, produced moisture maps of the emulsion in roughly 20 seconds per sample. For an industry racing to build smart factories, the implications are considerable.

The experimental design was deliberately systematic. The team prepared pork ham and back fat batters at three lean meat-to-fat ratios: Group A at 6:2, Group B at 5:3, and Group C at 4:4. Each formulation was mixed in a silent cutter operating at 1,440 rpm for seven different durations: 0, 1, 2, 5, 7, 10, and 15 minutes, with processing temperatures held at approximately 10 to 15 degrees Celsius. Over three batches, the researchers generated 84 samples, each imaged with a line-scan SWIR camera capable of capturing 275 spectral bands across 894 to 2,505 nanometers. Only the cleanest 120 bands, spanning 1,000 to 1,700 nanometers, entered the analysis. Reference moisture values came from the classic, laborious route: oven-drying one-gram subsamples at 105 degrees Celsius for 24 hours, repeated in triplicate for every sample.

The chemical results told a clear story before any photons were analyzed. Group A, richest in lean meat, held the most water at 63.4 percent, followed by Group B at 57.2 percent and Group C at 52.4 percent — differences that were highly statistically significant. This aligns with established meat science: lean muscle proteins form networks that bind water effectively, while increasing fat dilutes that capacity. Mixing time mattered too. Freshly assembled batter at zero minutes averaged 61.8 percent moisture, but after just one minute of cutting the value dropped and then stabilized at roughly 56 to 58 percent through the full 15 minutes. Two-way analysis of variance confirmed both factors as significant main effects, with no meaningful interaction between them — formulation and mixing time independently shape the water content of the final batter.

The spectral data told a parallel story in reflected light. Batters with more moisture absorbed more strongly, since water exhibits intense O–H absorption features in the near infrared; fattier batters reflected more light because fat particles scatter it. Group A showed the lowest overall reflectance, Group C the highest, with the most informative differences appearing near 1,100 nanometers, across 1,250 to 1,350 nanometers, and from 1,600 to 1,700 nanometers — regions corresponding to water O–H overtones and fat C–H overtones. Even more striking was what happened over time: at the start of mixing the spectra varied widely from sample to sample, but as mixing progressed the curves converged into stable, nearly uniform profiles, a direct optical signature of the batter homogenizing.

Turning spectra into numbers required serious modeling. The team extracted 840 mean spectra and trained four different regression approaches — partial least squares regression (PLSR), random forest (RF), Elastic Net, and a weighted ensemble of all three. To avoid the pitfall of a single lucky or unlucky train-test split, they used Monte Carlo random sampling, repeating the random 60/40 calibration-validation partition 100 times for each model and reporting averaged performance metrics. Ten-fold cross-validation guided model selection, with final models judged on the lowest root mean square error of cross-validation. Seven spectral preprocessing techniques, from standard normal variate to Savitzky–Golay derivatives, were tested against raw spectra as well.

The winner was, somewhat counterintuitively, the simplest treatment. Elastic Net regression applied to raw, unpreprocessed spectra delivered the best predictive performance, achieving a prediction coefficient of determination of 0.76 with a root mean square error of prediction of 2.54 percent. The ensemble model, which weighted each base model by the inverse of its cross-validation error squared, performed nearly as well and proved the most stable overall. Random forest posted the highest calibration fit at 0.87 to 0.89, a hallmark of mild overfitting, while PLSR and Elastic Net landed in a comparable band across validation and prediction sets. The researchers also noted that raw spectra often beat preprocessed ones in their experiments — a finding consistent with earlier literature showing that preprocessing is not automatically beneficial when baseline drift and scattering are minimal during acquisition.

Just as revealing was where each model looked. PLSR and Elastic Net concentrated their attention on the lower SWIR wavelengths, around 1,000 to 1,224 and 1,200 to 1,400 nanometers, where water’s O–H second overtone and combination bands dominate — exactly the regions expected to encode moisture. Elastic Net also flagged bands from 1,288 to 1,341 nanometers, hinting at additional fat-related C–H information. Random forest, by contrast, emphasized mid-to-high wavelengths between 1,359 and 1,641 nanometers, corresponding to the first overtone of hydrocarbon C–H bonds and to protein N–H and lipid C–H vibrations, reflecting its nonlinear capacity to capture composite spectral changes. The ensemble blended both perspectives, highlighting water bands near 1,200 and 1,400 nanometers alongside fat and protein regions extending to 1,700 nanometers.

Perhaps the most visually compelling result came from the chemical imaging. Applying the winning Elastic Net model pixel by pixel, the team generated false-color maps of moisture distribution across each batter block. At zero to two minutes of mixing, all groups showed scattered hotspots of locally high moisture exceeding 70 percent — evidence of poorly dispersed muscle proteins and fat particles. By 5 to 15 minutes the maps became visibly homogeneous, mirroring the spectroscopic convergence and the physical reality of proteins encapsulating fat globules into a stable emulsion. Fattier formulations, especially Group C, showed progressively declining moisture across the mixing timeline. The authors caution that the study used a modest sample set under controlled laboratory conditions, and that industrial deployment will demand validation on larger datasets in real processing environments, along with model simplification for low-latency, real-time operation. Still, the demonstration that a 20-second camera scan can replace a 24-hour oven test — while also showing whether a batch is evenly mixed — marks a meaningful step toward moisture monitoring that food processors can actually run on the factory floor.

The SWIR region occupies a particularly informative slice of the electromagnetic spectrum for food analysis. While visible and near-infrared instruments have long been used to assess meat quality, wavelengths beyond 1,000 nanometers probe stronger overtone and combination vibrations of the O–H, C–H, and N–H bonds that define water, lipid, and protein chemistry. This gives SWIR systems inherently richer contrast among the major constituents of a meat batter, though at the cost of weaker detector sensitivity and higher instrument expense, which has historically limited their adoption on processing lines.

The imaging approach also differs fundamentally from the point-sensor methods used in earlier emulsion studies. Fiber-optic probes and benchtop spectrometers average the signal over a small spot, so a single reading can mask pockets of unmixed fat or free water within a batch. By scanning an entire 18 by 10 centimeter block pixel by pixel, the hyperspectral camera turns heterogeneity itself into a measurable quantity, which is why the moisture maps could reveal localized hotspots above 70 percent that a spot measurement would likely have smoothed away.

The choice of reference chemistry matters as well. Oven drying at 105 degrees Celsius for 24 hours remains the AOAC gold standard, but it is destructive, slow, and impractical for in-line control. A prediction error of 2.54 percent moisture, while larger than the 0.291 percent reported in an earlier multispectral study of sausage emulsions, was achieved across a broader range of formulations and mixing times, and with the added benefit of spatial visualization rather than a single averaged value.

The finding that raw spectra outperformed preprocessed ones carries practical weight for industry. Preprocessing pipelines add computational overhead and can amplify noise when scattering conditions are already stable, so a model that tolerates uncorrected spectra simplifies deployment in smart-factory environments where low-latency predictions are essential.

Looking ahead, the authors’ emphasis on validation under real processing conditions points to the next hurdles: variable raw material lots, fluctuating line temperatures, and the calibration transfer needed to move a laboratory model onto multiple production lines without retraining from scratch.

Subject of Research: Non-destructive prediction of moisture content and mixing uniformity in meat emulsions using SWIR hyperspectral imaging and chemometric modeling

Article Title: Monitoring of moisture content in meat emulsion using an SWIR hyperspectral imaging system

Article References: Kim, J., Rho, T.-G., Park, E.-S., Kwon, O.-T., Lee, S.-J., Faqeerzada, M. A., Joshi, R., Amanah, H. Z., & Cho, B.-K. (2026). Monitoring of moisture content in meat emulsion using an SWIR hyperspectral imaging system. Food Science of Animal Resources, 46(1), Article 100. https://doi.org/10.1007/s44463-026-00101-9

Image Credits: AI Generated

DOI: 10.1007/s44463-026-00101-9

Keywords: SWIR hyperspectral imaging, meat emulsion, moisture prediction, chemometrics, Elastic Net regression, PLSR, random forest, ensemble modeling, process monitoring, meat batter quality, food analysis, non-destructive testing