…

The distinction between local vibration responses and machine-level load variations reflects a fundamental division in how rotating and reciprocating machinery distributes its dynamic behavior across spatial scales. A high-pressure abrasive waterjet machine tool is not a single homogeneous structure but a chain of loosely coupled subsystems: an intensifier pump that converts electrical power into pressurized water, an abrasive delivery mechanism that meters garnet particles into the stream, a cutting head that focuses the jet, and a motion system that positions the nozzle relative to the workpiece. Each subsystem generates characteristic dynamic signatures at its own location, while simultaneously drawing electrical current that varies with the aggregate mechanical load it imposes on the supply circuit. A fault localized to one component, such as nozzle erosion or abrasive feed irregularity, produces a strong vibration response near its physical seat but only a modest change in total current draw. Conversely, a pump-related anomaly that alters the pressure pulsation delivered to the entire system may register clearly in supply current while exciting vibration at all four measurement points to different degrees. The multimodal framework evaluated in this study exploits precisely this complementarity, and the magnitude of the reported gains over unimodal baselines suggests that neither modality alone captures the full state space of the machine.

The choice of a support vector machine for the feature-level fusion stage deserves some contextualization. Support vector machines construct decision boundaries by maximizing the margin between class clusters in a transformed feature space, and they remain competitive with deep neural networks when the number of training samples is small relative to the feature dimensionality. In industrial condition-monitoring experiments conducted on physical test rigs, dataset sizes are almost always constrained by the cost of running controlled fault scenarios, since each fault condition must be physically induced or allowed to develop under supervision. Under such constraints, kernel-based classifiers with well-engineered features frequently outperform end-to-end learned models, which require far more data to generalize. The 98.27 percent macro-F1 score achieved by the fused feature representation with an SVM classifier is consistent with this pattern, and the fact that decision-level fusion pushed performance above 99.6 percent indicates that the two modalities carry largely non-redundant information that can be combined either at the feature vector stage or at the classifier output stage with comparable success.

Macro-F1 as an evaluation metric warrants brief explanation for readers accustomed to overall accuracy figures. Macro-averaging computes the F1 score independently for each class and then takes the unweighted mean, which means that a classifier cannot achieve a high macro-F1 by performing well on frequent classes while neglecting rare ones. In fault diagnosis, where the six classes here comprise one normal state and five distinct component faults, macro-F1 is a more demanding criterion than accuracy because it penalizes poor performance on any individual fault category equally, regardless of how many samples that category contributes. The reported improvements of 7.06 percentage points over the current-only model and 14.50 percentage points over the vibration-only model therefore represent gains that are distributed across all fault categories rather than concentrated in the easiest ones, which strengthens the claim that fusion provides broadly useful discriminative information.

The fivefold contiguous time-block cross-validation protocol is a methodologically important detail that distinguishes careful condition-monitoring studies from those that overstate their results. In many published machine-learning diagnostics, samples from a single recording session are randomly shuffled and split into training and test folds, which allows the model to encounter nearly identical signal segments in both sets and produces inflated performance estimates. Contiguous time-block validation instead partitions the recording timeline into consecutive blocks and holds out entire blocks for testing, ensuring that the model must classify signal segments that are temporally separated from anything it saw during training. This design tests whether the learned feature representations capture stable physical signatures of each machine state rather than transient artifacts of a particular moment. The use of this protocol here, while still confined to a single experimental batch, means the reported scores reflect genuine temporal generalization within the batch, which is a more meaningful benchmark than random-split results would be.

The feature-group ablation findings align with established understanding of vibration-based diagnostics. Time-domain features, which include statistical moments, peak values, and crest factors computed directly from the raw acceleration signal, are sensitive to the overall energy and impulsiveness of the vibration and respond quickly to gross changes in machine behavior. Frequency-domain features, typically derived through spectral estimation methods such as Welch’s averaged periodogram, resolve the signal into its constituent frequency components and can isolate characteristic frequencies associated with specific rotating elements, pump pulsation harmonics, or structural resonances. Cross-axis information captures the directional structure of the vibration, since different fault mechanisms excite different geometric directions: a bearing defect may manifest predominantly in the radial plane, while an axial misalignment or thrust-related anomaly produces stronger axial response. The ablation results confirming that all three feature groups contribute to recognition performance suggest that the fault conditions studied produce distinguishable signatures across multiple descriptive dimensions, and that no single conventional feature family is sufficient on its own.

The physical rationale for combining vibration with supply current rests on the relationship between electrical power consumption and mechanical load. In an induction-motor-driven high-pressure pump, the current drawn from the supply reflects the torque demanded by the pump, which in turn depends on the pressure being generated and the volumetric flow being delivered. When abrasive flow rate changes, when the nozzle orifice wears and alters the discharge coefficient, or when internal pump leakage develops, the mechanical load shifts and the current signature follows. Vibration, by contrast, responds to the dynamic forces and impacts within each component, including pressure pulsations, particle impacts, mechanical looseness, and rubbing or bearing distress. Because these two quantities respond to overlapping but distinct sets of physical mechanisms, their combination provides a more complete description of machine state than either alone. The current modality acts as a slowly varying envelope that contextualizes the faster vibration fluctuations, helping the classifier distinguish, for example, a vibration pattern caused by a genuine component fault from a similar pattern caused by a temporary change in operating load.

The four monitoring locations were selected to cover the principal functional units of the machine. The high-pressure pump is the heart of the system, converting mechanical power into water pressure through an intensifier mechanism, and its dynamic behavior dominates the pressure waveform delivered downstream. The nozzle end sits closest to the cutting action, where jet formation, abrasive entrainment, and mixing occur, making it sensitive to nozzle wear, abrasive feed irregularities, and cutting-head blockage. The machine-tool spindle carries the motion system’s structural dynamics and reflects positioning accuracy and mechanical integrity of the drive train. The abrasive-tank bottom pressure-port region monitors the abrasive delivery system, where flow disruption or tank pressurization anomalies originate. Placing triaxial accelerometers at these four points yields twelve vibration channels that span the machine’s process chain, allowing the classifier to learn location-specific signatures and, through the rule-based module, to verify that predicted fault states are physically consistent with the monitoring locations where the strongest deviations are observed.

The deterministic rule-based post-classification module addresses a persistent gap between academic diagnostic models and practical industrial deployment. In real maintenance workflows, a bare classification output, however accurate, provides limited actionable guidance: maintenance personnel need to know what evidence supports the diagnosis, which other locations should be checked to confirm or refute it, and what verification steps to take. By applying a fixed topology matrix that encodes the physical relationships between components and monitoring points, together with predefined evidence rules that organize observed feature deviations into structured records, the module converts a probabilistic classifier output into an auditable diagnostic report. Crucially, the module preserves the original classifier outputs rather than modifying them, which maintains a clean separation between statistical inference and rule-based reasoning. This separation means that errors in the rule layer cannot corrupt the classification itself, and that the rule layer can be inspected, audited, and updated independently of the trained model, which is valuable in regulated or safety-critical industrial environments.

The explicit acknowledgment that the reported performance represents within-batch recognition under fixed sensor placement is an important caveat that deserves elaboration. Sensor remounting, even with magnetic attachment at the same nominal location, introduces variability in coupling stiffness, orientation, and contact geometry that can shift the vibration transfer function between the source and the sensor. Different machines of the same model exhibit manufacturing tolerances, assembly variations, and wear histories that alter their dynamic signatures. Operating batches may differ in material type, cutting parameters, abrasive characteristics, and ambient conditions, all of which modulate the signals available to the classifier. A model trained on one batch may therefore lose accuracy when applied to another, a phenomenon related to domain shift in machine learning. The authors’ careful scoping of their claims to within-batch recognition, rather than claiming cross-machine or cross-batch generalization, reflects a rigorous understanding of these limitations and provides an honest baseline against which future transfer-learning and domain-adaptation studies can be measured.

The abrasive waterjet machining context itself shapes the diagnostic challenge in specific ways. Unlike conventional machining, where tool wear progresses gradually and cutting forces provide a direct measure of tool-workpiece interaction, AWJ machining involves a high-velocity mixture of water and abrasive particles whose cutting action depends on jet coherence, particle velocity distribution, and standoff distance. The absence of a rigid tool contact means that force-based monitoring is less straightforward, and the high-pressure environment limits sensor placement options near the cutting zone. Vibration sensing on the machine structure offers a non-intrusive alternative that captures the structural response to jet formation and component operation without exposing instruments to the abrasive stream. Combined with electrical current measurement, which requires only a clamp meter on the supply circuit and involves no contact with the process at all, the two modalities form a practical sensing pair that can be retrofitted to existing machines with minimal modification.

The sampling configuration also merits attention. Vibration was acquired at 1 kHz per channel, a rate sufficient to capture frequency content up to the Nyquist limit of 500 Hz, which encompasses the dominant structural resonances and pump pulsation frequencies of typical high-pressure systems. The current signal was sampled at a median interval of 0.75 seconds, a much slower rate appropriate for a modality that varies slowly with load changes rather than with mechanical oscillations. Aligning these heterogeneous sampling rates into common 2-second analysis windows required careful synchronization, and the resulting dataset pairs each vibration snapshot with the corresponding current reading, enabling the fusion models to learn joint patterns across timescales. This multi-rate acquisition strategy mirrors practical industrial deployments, where vibration monitoring systems and power monitoring systems typically operate independently with different sampling characteristics, and demonstrates that meaningful fusion is achievable without forcing all modalities onto a common high-rate clock.

Looking toward future development, the framework establishes several natural directions for extension. Cross-batch validation, in which the model is trained on one experimental campaign and tested on a separate campaign conducted under different conditions, would directly probe the generalization limits that the current study deliberately excludes. Transfer-learning techniques, including domain adaptation and fine-tuning strategies, could address the variability introduced by sensor remounting and machine-to-machine differences. Expanding the fault taxonomy beyond the five predefined conditions to include progressive degradation states rather than binary fault categories would enable prognostic capabilities alongside diagnostic ones. The structured evidence records produced by the rule-based module could additionally serve as training material for higher-level reasoning systems, bridging the gap between statistical classification and the maintenance decision-making processes used in industrial practice, while the deterministic architecture ensures that such integration does not compromise the transparency and auditability of the diagnostic chain.

Subject of Research: Multimodal vibration and current fusion for recognizing component faults in an abrasive waterjet machine tool

Article Title: Vibration–current multimodal fusion for within-batch recognition of predefined component faults of a high-pressure abrasive waterjet machine tool

Article References: Vibration–current multimodal fusion for within-batch recognition of predefined component faults of a high-pressure abrasive waterjet machine tool. (n.d.). https://doi.org/10.3389/fmech.2026.1970496

Image Credits: AI Generated

DOI: 10.3389/fmech.2026.1970496

Keywords: abrasive waterjet machining, fault diagnosis, multimodal fusion, vibration monitoring, machine current, support vector machine, cross-validation, condition monitoring, feature ablation, rule-based evidence, Vibration, current