REAL-TIME ANOMALY DETECTION AND PREDICTIVE MAINTENANCE IN HIGH-THROUGHPUT MOLECULAR DIAGNOSTIC CARTRIDGE MANUFACTURING USING MULTI-MODAL SENSOR FUSION AND ENSEMBLE MACHINE LEARNING

Volume-08 Issue 12, December-2024 ISSN: 2456-9348 Impact Factor:7.936 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [596] REAL-TIME ANOMALY DETECTION AND PREDICTIVE MAINTENANCE IN HIGH-THROUGHPUT MOLECULAR DIAGNOSTIC CARTRIDGE MANUFACTURING USING MULTI-MODAL SENSOR FUSION AND ENSEMBLE MACHINE LEARNING Adetayo Folasole Department of Computing, East Tennessee State University, United States of America ABSTRACT The increasing global demand for rapid, high-throughput molecular diagnostic testing has placed unprecedented performance and reliability expectations on cartridge manufacturing systems. These production environments operate under stringent precision, sterility, and quality requirements, where even minor deviations in thermal control, reagent dispensing, microfluidic sealing, or mechanical assembly can compromise diagnostic accuracy. From a broader perspective, real-time anomaly detection and predictive maintenance have become essential pillars of modern diagnostics manufacturing, ensuring uninterrupted output while reducing scrap rates, equipment failures, and compliance risks. This study situates advanced monitoring strategies within the wider context of Industry 4.0, where multi-modal sensing, automated analytics, and continuous validation workflows are transforming operational reliability. Narrowing the focus, the paper presents a specialized framework tailored for high-throughput molecular diagnostic cartridge production, integrating multi-modal sensor fusion with ensemble machine learning techniques. The proposed approach collects and harmonizes diverse data streams including vibration signatures, thermal profiles, laser micrometry, optical imaging, pressure curves, and reagent-flow metrics to capture the full operational envelope of assembly and verification processes. These heterogeneous data sources are processed through an ensemble of machine learning models combining random forests, gradientboosted methods, one-class SVMs, and temporal sequence models for dynamic anomaly detection. A predictive maintenance layer leverages fused sensor features and historical fault patterns to forecast component degradation and proactively schedule interventions before deviations escalate into quality failures or line downtime. The resulting system enhances real-time defect detection, reduces false alarms, and provides interpretable alerts for quality engineers and process technicians, thereby accelerating root-cause isolation and improving regulatory compliance. By merging multi-modal sensing with robust ensemble analytics, this study demonstrates a scalable pathway toward fault-tolerant, data-driven manufacturing in molecular diagnostics. Keywords: Predictive maintenance; Multi-modal sensor fusion; Ensemble machine learning; Real-time anomaly detection; Diagnostic cartridge manufacturing; Industry 4.0 1. INTRODUCTION AND INDUSTRY CONTEXT 1.1 Global expansion of molecular diagnostic cartridge manufacturing The rapid global growth of molecular diagnostic cartridge manufacturing has been driven by rising demand for portable, decentralized, and rapid-testing platforms capable of supporting clinical decisions across diverse healthcare environments [1]. As nations continue to scale point-of-care molecular systems, cartridge production has shifted from niche, laboratory-restricted fabrication to a major industrial activity involving high-throughput assembly lines and sophisticated quality-control pipelines [2]. Manufacturers now operate in an increasingly competitive landscape where speed, reliability, and contamination-free workflows determine both regulatory success and market positioning. This expansion is also marked by a broader transition from manual or semiautomated processes to automation-intensive factories powered by intelligent robotics and electro-mechanical subsystems [3]. Diagnostic cartridges themselves have become more complex, integrating microfluidics, biochemical reagents, magnetic bead modules, optical detection chambers, and precision-molded plastics [4]. Each of these components introduces intricate tolerance requirements, heightening the sensitivity of the production environment to physical, mechanical, and environmental variations [3]. Simultaneously, global supply chains for polymers, reagents, and Volume-08 Issue 12, December-2024 ISSN: 2456-9348 Impact Factor:7.936 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [597] microelectronic components remain vulnerable to disruptions, forcing manufacturers to compensate through tighter process control and real-time monitoring [5]. Facilities producing millions of cartridges per month must ensure consistent sterility, performance stability, and batch reproducibility under stringent international standards such as ISO 13485 and ISO 14644 [4]. This collective push toward larger production volumes, tighter tolerances, and stricter compliance frameworks has increased the strain on legacy operational models. As production scales upward, factories must adopt more adaptive intelligence to prevent defects and ensure uninterrupted throughput, creating a strong rationale for sensorrich, data-driven manufacturing ecosystems [6]. 1.2 Limitations of legacy quality-control and maintenance systems Traditional quality-control systems in cartridge manufacturing rely heavily on periodic manual inspection, sporadic sampling, and scheduled equipment maintenance routines that often fail to reflect real-time operational realities [7]. These legacy approaches operate on static thresholds and rigid procedural assumptions, making them poorly suited for dynamic production environments where micro-contamination, thermal fluctuations, equipment drift, and tooling wear can occur unpredictably [6]. Many high-throughput lines still depend on maintenance cycles based on fixed calendar intervals rather than actual equipment condition. As a result, organizations experience unplanned downtime, inconsistent product quality, and escalating operational costs. Manual inspection further introduces variability and human error, particularly when identifying subtle defects in microfluidic geometries or reagent-filled chambers [7]. In addition, traditional supervisory control systems often lack cross-line integration, meaning anomalies detected in one part of the assembly are not automatically correlated with upstream conditions. Furthermore, conventional alarm systems frequently generate high volumes of false positives or ambiguous alerts lacking root-cause clarity [8]. These limitations collectively highlight the need for predictive, interconnected, and adaptive systems capable of monitoring equipment, environmental conditions, and material flows in real time. As production complexity rises, legacy QC models no longer provide the speed, precision, or anticipatory capability required for modern cartridge manufacturing [9]. 1.3 Rise of sensor-driven smart factories in biomanufacturing Biomanufacturing facilities have increasingly adopted sensor-driven architectures to support continuous monitoring, high-resolution anomaly detection, and automated decision-making. These smart factories leverage distributed arrays of sensors acoustic, thermal, vibration, optical, airborne particle counters, flow sensors, and chemical detectors to generate dense data streams that characterize both equipment health and product quality in real time [10]. The transition toward sensor-driven digital infrastructure is motivated by the recognition that early detection of micro-contamination events, thermal drift, mechanical misalignment, or reagent degradation requires more frequent and granular data than humans can manually gather. Smart factories also strengthen regulatory compliance by enabling continuous documentation and traceability, which are essential in environments where deviations must be rapidly identified and corrected [7]. Connectivity technologies such as industrial Ethernet, OPC-UA, and edge computing solutions support seamless transport and analysis of sensor data across production units. This integrated ecosystem reduces latency, enabling decisions to occur at or near the point of data generation. It also creates the foundation for advanced analytics, including machine learning systems trained to detect early warning patterns invisible to human operators. By converting traditionally reactive workflows into proactive and predictive ones, sensor-driven factories redefine both the pace and reliability of high-throughput cartridge manufacturing. 1.4 Why multi-modal sensor fusion and ensemble ML are needed Although individual sensors provide valuable information, relying on single-signal monitoring is insufficient in a manufacturing context where faults often arise from interacting variables rather than isolated anomalies. Multimodal sensor fusion allows data from different physical domains thermal, acoustic, mechanical, optical, chemical, and particulate to be combined into a unified analytical framework that captures the true complexity of ISOcontrolled production lines [4]. When fused properly, these signals offer deeper insight into the latent conditions that precede equipment failure, contamination events, or microfluidic defects. Ensemble machine-learning models further enhance predictive power by integrating diverse classifiers, anomaly detectors, and forecasting algorithms. These ensembles reduce the weaknesses of individual models while Volume-08 Issue 12, December-2024 ISSN: 2456-9348 Impact Factor:7.936 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [598] amplifying their strengths, providing more stable and interpretable outputs. For instance, thermal drift may correlate with acoustic signatures of bearing wear, while airborne particulate spikes may align with vibration anomalies in sealing mechanisms [2]. Single-sensor models often overlook such interactions. Fusion-driven ensemble ML also supports spatiotemporal reasoning, enabling the detection of progressive degradation patterns across interconnected machines. As production lines operate continuously, small deviations can propagate, eventually leading to batch-level failures if not intercepted early [7]. The combined use of multi-modal sensor fusion and ensemble ML therefore transforms monitoring systems from static, reactive tools into adaptive, anticipatory intelligence capable of ensuring high throughput, product integrity, and compliance in complex diagnostic cartridge manufacturing environments [5]. 2. MOLECULAR DIAGNOSTIC CARTRIDGE MANUFACTURING ARCHITECTURE 2.1 Overview of high-throughput assembly line workflows Modern high-throughput molecular diagnostic cartridge manufacturing operates through tightly synchronized assembly-line workflows engineered to maintain speed, cleanliness, and mechanical precision across multiple sequential stages. These production lines typically begin with raw material preparation, where polymer pellets, microfluidic substrates, and reagent reservoirs are conditioned, cleaned, and verified before entering downstream equipment [8]. Injection molding units then shape microfluidic shells and reaction chambers, followed by automated inspection systems that screen each molded component for dimensional accuracy, wall thickness, and defect-free surfaces. Once structural components are verified, microfluidic subassemblies move into reagent dispensing stations where enzymes, buffers, primers, and magnetic bead mixes are deposited with microlitre-level accuracy. Small deviations in reagent volume can alter analytical sensitivity, making this step one of the most tightly controlled across the workflow [9]. Optical sensors, gravimetric checks, and temperature-regulated dispensing heads work in tandem to ensure biochemical consistency. Subsequent processes include foil sealing, bonding, ultrasonic welding, and precision assembly of functional modules such as valves, filters, and optical windows. Automated pick-and-place robots integrate these micro-scale components at speeds that require flawless synchronization and vibration isolation to avoid mechanical misalignment [10]. Final stages include high-speed packaging, sterilization, labeling, and serialization, all completed under ISO-classified cleanroom conditions. Because these workflows involve a mix of thermal cycles, mechanical stresses, airflow dynamics, and biochemical sensitivity, their performance depends on continuous stability across hundreds of interacting subsystems. Any upstream anomaly however minor can propagate rapidly, affecting yield, reagent integrity, and overall device reliability [11]. 2.2 Microfluidic sensitivity, biochemical stability, and mechanical tolerances Microfluidic diagnostics rely on precise channel geometries and highly stable biochemical environments, making the cartridges exceptionally sensitive to fluctuations in production conditions. Micro-scale channels, often narrower than a human hair, require smooth surfaces, defect-free corners, and exact tolerances to support consistent capillary flow and accurate thermocycling. Even microscopic burrs or resin inconsistencies can alter flow resistance or generate unwanted turbulence, leading to false results or device malfunction [12]. Biochemical stability is equally fragile. Enzymes, nucleic acid probes, and lyophilized reagents are temperatureand humidity-sensitive, meaning ambient conditions across dispensing, bonding, and curing stages must remain within extremely narrow ranges. Airborne particles, volatile organic compounds, or trace contaminants can inhibit enzymatic reactions long before failures become visible at QC checkpoints [13]. Mechanical tolerances in bonding, sealing, and alignment stages introduce additional layers of complexity. Ultrasonic welding parameters amplitude, pressure, frequency must be calibrated precisely to avoid microcracks or partial seals that compromise the integrity of reaction chambers. Even a slight drift in welding horn alignment can disrupt sealing uniformity. At high throughput, minor mechanical vibration or thermal expansion of tooling can accumulate into pronounced defects across entire batches [14]. The interplay between microfluidic geometry, reagent sensitivity, and mechanical precision means each subsystem must operate in harmony. As production speeds increase, so does the magnitude of risk: interactions between heat, airflow, mechanical stress, and biochemical degradation become more nonlinear and harder to detect using conventional monitoring systems [15]. Volume-08 Issue 12, December-2024 ISSN: 2456-9348 Impact Factor:7.936 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [599] 2.3 Identifying critical failure points and contamination risks Failed cartridges typically originate from a small subset of critical failure points distributed across the assembly line. These points frequently include reagent dispensing inaccuracies, microfluidic bonding defects, airborne particulate intrusions, and thermal instability during enzymatic handling. Failures may also emerge from misaligned robotic components, worn sealing fixtures, or microleaks introduced during packaging [16]. Contamination risks biological, chemical, or particulate are among the most challenging hazards to control. ISO 5 and ISO 7 cleanrooms remain vulnerable to micro-contamination events caused by improper gowning, equipment vibration, airflow turbulence, electrostatic particle attraction, or off-gassing from materials. Environmental disruptions such as pressure drift or filter inefficiencies can introduce contaminants that remain undetected until downstream QC stages, where entire batches must be quarantined or discarded. Identifying these risks often requires correlating signals from multiple subsystems, making single-sensor monitoring ineffective. Figure 1, “End-to-End Workflow of Molecular Diagnostic Cartridge Production and Critical Control Points,” illustrates how failure modes cluster around key transitions between molding, reagent loading, bonding, and sterile packaging, where mechanical, biochemical, and environmental conditions intersect most sharply [17]. Figure 1, “End-to-End Workflow of Molecular Diagnostic Cartridge Production and Critical Control Points 2.4 Operational bottlenecks in scaling cartridge production As manufacturers scale output, bottlenecks increasingly emerge from tooling fatigue, reagent throughput limitations, robotic cycle-time constraints, and cleanroom airflow balancing. Minor delays in microfluidic molding or reagent replenishment can cascade into large line slowdowns, especially when upstream and downstream buffers are insufficiently sized [12]. Additionally, manual interventions during maintenance introduce variability and downtime that interrupt continuous production rhythms. These operational constraints highlight the need for intelligent, predictive, and fully integrated monitoring frameworks capable of stabilizing throughput across rapidly scaling cartridge manufacturing operations [9]. Volume-08 Issue 12, December-2024 ISSN: 2456-9348 Impact Factor:7.936 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [600] 3. MULTI-MODAL SENSOR FUSION IN ISO 5 CLEANROOM MANUFACTURING 3.1 Sensor ecosystem: optical, acoustic, vibration, thermal, pressure, flow, chemical Modern cartridge-manufacturing environments require an expansive sensor ecosystem capable of capturing subtle variations across mechanical, thermal, environmental, and biochemical domains. Optical sensors form the backbone of high-precision inspection, enabling detection of microfluidic channel defects, reagent-level inconsistencies, sealing misalignments, and particulate intrusions through high-resolution imaging and structuredlight methods [14]. These systems can also validate transparency, colorimetric stability, and internal cartridge geometry, which are essential for downstream assay performance. Acoustic sensors contribute complementary diagnostic value by measuring high-frequency signatures generated by motors, sealing presses, and robotic actuators. Deviations in acoustic profiles often indicate emerging wear, lubrication failures, or vibration coupling across mechanical assemblies [15]. Vibration sensors capture these mechanical anomalies directly, documenting misalignment, resonance shifts, or unstable motion in pick-and-place robots and sealing dies. Thermal sensors provide critical insight into both environmental conditions and equipment performance by monitoring temperature gradients in reagent-handling modules, curing stations, and molding units. Even minor temperature drift can accelerate reagent degradation or alter polymer behavior during bonding processes [16]. Pressure sensors ensure stability in pneumatic systems driving liquid dispensing, valve operations, and microfluidic routing. Meanwhile, flow sensors monitor reagent delivery rates, air displacement accuracy, and uniformity in microfluidic test channels, allowing the system to detect clogs, fouling, or inconsistent dosing. Chemical sensors enhance contamination control by detecting volatile organic compounds, residual cleaning agents, humidity fluctuations, or off-gassing events that may compromise biochemical reagents [17]. Combined, these sensor modalities form a multi-layered surveillance framework capable of identifying early deviation patterns that would otherwise go undetected in traditional manufacturing workflows [18]. 3.2 Sensor placement logic across robotics, sealing units, and microfluidic subsystems The diagnostic value of sensors depends not only on their type but on their strategic positioning across manufacturing subsystems. In robotic assembly units, vibration and acoustic sensors are typically placed near joint housings, end-effectors, and linear actuators to capture localized mechanical stress and alignment variations [19]. Optical systems are positioned above conveyor pathways or integrated into robotic grippers for real-time inspection during pick-and-place operations. In sealing units, thermal and pressure sensors are distributed along heated sealing plates, bonding tools, and pneumatic drivers to track uniformity during lamination and heat-pressing steps. Deviations detected here often correlate with long-term sealing failures or reagent leakage during assay activation [20]. Microfluidic assembly subsystems require the most delicate placement logic, with flow sensors embedded near reagent-loading ports and pressure sensors positioned along micro-channels to detect abnormal resistance patterns. Chemical sensors are placed near reagent-handling enclosures to monitor environmental stability. This distributed arrangement creates a dense spatial map of the manufacturing line, ensuring that anomalies arising in upstream steps can be immediately associated with downstream quality signatures. It also enables integrated decision-making models that tie mechanical behavior to biochemical consistency across the full assembly lifecycle [21]. 3.3 Time-series synchronization and feature alignment challenges A key challenge in multi-modal sensing is synchronizing highly heterogeneous time-series streams collected at different sampling frequencies, latencies, and noise profiles. Optical systems generate data in intermittent bursts, while vibration or thermal sensors produce continuous high-frequency signals [22]. Aligning these streams requires robust timestamp harmonization, interpolation, and buffering strategies to ensure temporal coherence. Feature alignment becomes even more complex when combining data from sensors positioned across separate subsystems with physically offset event triggers. For instance, an acoustic anomaly detected at the robotic feeder may correlate with an optical defect seen several seconds later at the inspection module. Addressing these issues requires standardized synchronization schemes tied to the manufacturing execution system (MES), as well as embedding explicit time markers into all sensor outputs. These alignment methods are crucial for the integration of analytics models and directly influence the integrity of downstream fusion algorithms used to interpret Table 1 (“Sensor Categories, Measured Variables, and Diagnostic Relevance in Cartridge Manufacturing”) [23]. Volume-08 Issue 12, December-2024 ISSN: 2456-9348 Impact Factor:7.936 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [601] Table 1. Sensor Categories, Measured Variables, and Diagnostic Relevance in Cartridge Manufacturing Sensor Category Measured Variables Diagnostic Relevance in Cartridge Manufacturing Optical Sensors Surface integrity, micro-cracks, alignment, bonding quality, reagent volume, seal uniformity Detect visual defects such as incomplete sealing, scratched surfaces, misaligned components, and reagent underfill/overfill; essential for microfluidic transparency checks. Acoustic Sensors Acoustic signatures, resonance patterns, frequency shifts Identify anomalies in sealing pressure, bonding failures, crack propagation, and improper seating of components based on acoustic deviation patterns. Vibration Sensors (Accelerometers) Amplitude, frequency spectra, mechanical oscillation patterns Detect robotic misalignment, motor wear, bearing degradation, and abnormal friction events that lead to assembly errors or premature equipment failure. Thermal/Infrared Sensors Temperature gradients, heat distribution, thermal drift Reveal hotspots in sealing units, thermal bonding irregularities, reagent degradation, and unstable heating elements during cartridge assembly. Pressure Sensors Pneumatic pressure, seal compression forces, vacuum levels Identify insufficient sealing forces, leakage pathways, improper vacuum suction in pick-and-place systems, and pressure instability in microfluidic units. Flow Sensors Flow rate, microfluidic stability, turbulence signatures Detect clogging, reagent flow obstruction, microchannel constriction, and variation in reagentdispensing precision within cartridges. Chemical Sensors Gas composition, VOCs, humidity, contaminant presence Monitor micro-contamination, humidity-sensitive reagent stability, polymer curing quality, and airborne by-products generated during assembly. Force/Torque Sensors Applied load, torque variation during fastening or sealing Identify over-compression or under-tightening of cartridge components, mechanical stress accumulation, and tool miscalibration. Position/Encoder Sensors Robotic arm coordinates, actuator displacement, motion precision Detect micro-misalignment during robotic assembly cycles, identify drift in automated positioning, and ensure consistency in repetitive motion tasks. 3.4 Calibration, drift correction, noise filtering, and data-quality assurance Long-term deployment of sensors in high-throughput environments leads to calibration drift, signal decay, and noise accumulation. Thermal sensors may display baseline drift due to heating cycles, while vibration sensors become susceptible to mechanical resonance shifts as components age [14]. Filtering these distortions requires automated drift-correction routines built into edge-processing units, enabling recalibration without halting production. Noise filtering is equally critical, especially for acoustic and vibration channels that are inherently sensitive to factory-wide disturbances. Techniques such as adaptive filtering, wavelet decomposition, and frequency-domain noise suppression ensure that only meaningful signatures are forwarded to analytics pipelines [17]. Data-quality assurance frameworks further monitor signal completeness, dropout events, and sudden out-of-range values. When combined, these processes guarantee that fused inputs retain statistical integrity, minimize false alarms, and maintain traceability for regulatory audits. Without rigorous quality assurance, even sophisticated AI models risk misclassifying benign fluctuations as anomalies or overlooking early indicators of equipment degradation [20]. 3.5 Fusion methodologies across heterogeneous streams Multi-modal sensor fusion involves integrating diverse data streams into a unified analytical representation capable of capturing cross-sensor correlations and emergent fault signatures. Low-level fusion merges raw signals to enhance temporal and spatial resolution, often combining vibration and acoustic signatures to detect micromechanical changes [18]. Mid-level fusion aggregates engineered features such as thermal gradients, pressure deviations, and optical intensity values into structured vectors suitable for machine-learning models [21]. Volume-08 Issue 12, December-2024 ISSN: 2456-9348 Impact Factor:7.936 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [602] High-level fusion integrates outputs from multiple predictive models, enabling ensemble decision-making that reduces false positives and strengthens root-cause interpretability [19]. Bayesian fusion, Kalman-based methods, and deep-learning architectures such as multimodal autoencoders are commonly applied to integrate heterogeneous information streams and capture latent relationships across subsystems. These fusion strategies allow the manufacturing line to transition from siloed sensing to holistic situational awareness, supporting predictive maintenance, contamination prevention, and real-time quality control. They form a critical foundation for robust ensemble machine-learning frameworks used in modern cartridge production environments [15]. 4. ENSEMBLE MACHINE LEARNING FOR ANOMALY DETECTION AND PREDICTIVE MAINTENANCE 4.1 Supervised defect detection models for cartridge assembly anomalies Supervised machine-learning models have become foundational tools for detecting manufacturing defects during cartridge assembly, particularly where labelled historical datasets exist to capture representative failure patterns across sealing, dispensing, robotic positioning, and microfluidic construction steps. These models learn discriminative boundaries that differentiate normal conditions from defective states by analysing multi-modal features derived from optical, vibration, pressure, and thermal sensors. Convolutional neural networks are often applied to high-resolution optical data to classify surface defects, contamination spots, misaligned laminates, or microfluidic channel distortions [22]. Gradient-boosted decision trees, meanwhile, excel in processing engineered features from vibration and pressure sensors, enabling accurate recognition of anomalies such as improper clamp force, pneumatic leakage, or mechanical drift in high-throughput loading units. A key advantage of supervised approaches lies in their ability to incorporate domain-specific labels that encode the severity, causality, and operational context of each defect, enabling more granular quality-assurance logic [24]. However, labelled data is rarely exhaustive, particularly for low-frequency anomalies related to contamination, reagent instability, or mechanical fatigue. This limitation necessitates robust cross-validation schemes that prevent overfitting and ensure generalizable performance during production variability, tool wear, and machine recalibration cycles [26]. When carefully trained, supervised models deliver reliable real-time alerts that support batch release decisions, equipment adjustments, or immediate workstation isolation, thereby significantly reducing the probability of defective cartridges entering downstream diagnostic workflows [25]. 4.2 Unsupervised anomaly detection (autoencoders, clustering, isolation forest) Unsupervised anomaly-detection models have become increasingly essential in cartridge manufacturing due to the inherent scarcity of labelled fault examples and the continuous evolution of equipment behaviour. Autoencoders represent a powerful class of models that learn compressed representations of normal sensor patterns and measure reconstruction deviation to detect irregularities. When fed multi-modal data, deviations in reconstruction error often indicate micro-defects, unusual acoustic signatures, or atypical thermal gradients that would not be captured using rule-based logic [28]. Cluster-based approaches such as k-means or density-based methods allow grouping of normal operational states while isolating outliers that may signal drift, contamination, or emerging mechanical instability. Isolation Forest algorithms complement these approaches by isolating anomalies based on their statistical rarity within high-dimensional feature spaces [23]. These models excel at detecting abrupt changes in vibration envelopes, irregular reagent-dispensing flow profiles, or shifts in pneumatic pressure behaviour that may not yet manifest visually. Collectively, unsupervised methods operate continuously without requiring predefined labels, making them particularly valuable for discovery of novel fault modes or unknown interactions across robotic, sealing, and microfluidic systems. A central benefit of unsupervised techniques is their suitability for early detection under conditions where anomalies accumulate gradually, such as slow heater degradation or subtle actuator misalignment [30]. Their flexibility and independence from curated defect sets position them as indispensable tools for modern highthroughput production lines, enabling factories to identify risk conditions before they escalate into clinical-quality failures [27]. 4.3 Hybrid ensemble modelling for multi-modal fusion Hybrid ensemble architectures combine supervised and unsupervised models to enhance robustness, improve cross-modal reasoning, and reduce false positives in real-time detection pipelines. Because cartridge Volume-08 Issue 12, December-2024 ISSN: 2456-9348 Impact Factor:7.936 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [603] manufacturing involves heterogeneous sensor streams optical, acoustic, vibration, thermal, flow, and chemical no single model optimally captures all fault signatures. Ensembles integrate multiple weak learners or specialized feature extractors into a unified decision engine, enabling complementary detection capacity across diverse operational contexts [29]. In practice, supervised CNN classifiers may be paired with unsupervised autoencoder anomaly scores, while gradient-boosted trees integrate statistical features derived from vibration and flow signals. Decision-level fusion then aggregates outputs using weighted voting, Bayesian updating, or calibrated confidence scoring, ensuring that atypical events emerging in one sensor modality are cross-validated by others [24]. This greatly reduces the likelihood of noise-driven false alarms or blind spots. Hybrid ensembles also allow real-time incorporation of sensor redundancy, recognizing that some modalities occasionally fail or enter suboptimal calibration states. The ensemble thus maintains resilience during high-speed production cycles. Figure 2: Ensemble Machine Learning Pipeline for Real-Time Anomaly Detection and Predictive Maintenance. As shown in Figure 2, ensemble architectures form the interpretative centrepiece of predictive maintenance workflows, linking multi-modal fusion to both anomaly detection and equipment-degradation forecasting [22]. 4.4 Spatiotemporal learning for early detection of equipment degradation Spatiotemporal learning models expand beyond static feature analysis by capturing temporal evolution and spatial interdependencies in sensor data. Recurrent neural networks (RNNs), gated recurrent units (GRUs), and long short-term memory networks (LSTMs) are commonly applied to vibration, acoustic, and thermal time-series streams to uncover degradation signatures indicative of bearing wear, actuator backlash, resonance drift, or pneumatic deterioration [25]. These models track how signals evolve over operational cycles, identifying subtle sequential changes that typically precede catastrophic machine failure. On the spatial dimension, models such as 2D-CNNs or graph-based neural networks can interpret correlations across sensor arrays distributed along sealing dies, conveyor segments, and robotic joints. This is particularly Volume-08 Issue 12, December-2024 ISSN: 2456-9348 Impact Factor:7.936 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [604] valuable when mechanical imbalances propagate across discreet modules, such as when misalignment in a primary actuator affects downstream optical positioning consistency [30]. The fusion of spatial and temporal inference allows the system to generate early-warning indicators that traditional threshold systems would not capture. Spatiotemporal models are especially effective for predicting “slow-burn” degradation modes, where micro-level anomalies accumulate gradually rather than manifesting as abrupt deviations [28]. Their predictive ability directly supports preventive maintenance, enabling manufacturing teams to schedule repairs before operational reliability is compromised [26]. 4.5 Model interpretability, explainability, and confidence scoring As machine-learning systems increasingly influence quality-control decisions and maintenance scheduling, interpretability becomes essential for regulatory compliance, operational trust, and root-cause diagnosis. Explainable AI techniques enable operators to understand why a model flagged a defect or predicted equipment degradation, ensuring transparency in decision-making pipelines. Feature-importance maps, saliency overlays, and SHAP-value decompositions help clarify which sensor inputs contributed most to a given prediction [23]. For optical systems, visualization tools highlight defective regions, sealing inconsistencies, or microfluidic anomalies responsible for classification results [29]. For vibration and acoustic anomalies, frequency-domain heatmaps help engineers pinpoint resonance peaks or harmonic distortions linked to mechanical failures. Confidence scoring complements interpretability by quantifying prediction certainty. Calibrated probability outputs, ensemble-based voting consistency scores, and Bayesian uncertainty estimates assist operators in distinguishing between high-confidence alerts requiring immediate action and low-confidence signals that may warrant further investigation [27]. This prevents unnecessary line stoppages while maintaining vigilance against genuine threats. By combining explainability with confidence metrics, factories can strengthen decision governance, ensure traceability, and create auditable pathways aligned with quality-management standards [30]. Ultimately, interpretability transforms ML models from opaque systems into trusted partners within the broader predictivemaintenance ecosystem [24]. 5. REAL-TIME EDGE-AI DEPLOYMENT AND OPERATIONAL INTEGRATION 5.1 Edge-AI inference for low-latency QC and equipment monitoring Edge-AI computing has become a foundational enabler of low-latency quality control and real-time equipment monitoring in high-throughput cartridge-manufacturing environments. At the workstation level, embedded inference modules process multi-modal sensor streams including optical, vibration, acoustic, thermal, and flow data to detect deviations within milliseconds, thereby avoiding delays associated with centralized server routing [28]. By performing inference physically closer to robotic manipulators, sealing heads, dispensing units, and microfluidic assembly stations, the system ensures that decisions are made at production-line speed and without interruption. This architecture significantly reduces dependency on bandwidth-heavy upstream transmission, allowing manufacturing lines to maintain peak throughput even when network load fluctuates or cloud access is temporarily constrained [34]. Edge intelligence also enables selective data offloading, where only high-value events such as confirmed anomalies, sensor drifts, or predictive-maintenance signals are transmitted for further analysis. This reduces unnecessary data movement and preserves computational resources across distributed plants. A major benefit of edge-AI inference is its ability to support adaptive control loops, in which detected anomalies automatically trigger real-time machine adjustments, robotic path recalibration, or pneumatic pressure corrections before defects propagate downstream [31]. Moreover, edge processing strengthens operational resilience during transient failures by maintaining autonomous anomaly detection even when cloud connections degrade, positioning it as a core capability for future-ready diagnostic manufacturing [36]. 5.2 Cloud–edge hybrid pipelines for scalable analytics Cloud–edge hybrid architectures provide the scalability required to analyse large volumes of historical sensor data, enabling deeper trend discovery and long-horizon predictive modelling beyond what on-device inference modules can support. The cloud layer aggregates time-series information from optical stations, thermal arrays, vibration modules, and flow meters across multiple production shifts, allowing large-scale analytics, retraining of anomaly models, and longitudinal asset-health assessments [38].