Simulation maps how motor fault signatures shift across the full load range

Engineers who keep industrial motors running rely on telltale patterns in electrical and mechanical signals to catch faults before they cause breakdowns. A new simulation study argues that those patterns are far more fragile than many diagnostic schemes assume, because the values of most monitored features change substantially as the mechanical load on the motor changes. The work, published in Discover Electronics, systematically characterises fault signatures in a three-level neutral-point-clamped (NPC) inverter-fed induction motor drive across thirteen load levels, from 10 percent to 110 percent of rated torque.

The study, conducted entirely in MATLAB/Simulink by Sena Gemechis File, addresses two gaps the author identifies in the existing literature. Most published fault characterisation studies examine a single operating point, typically rated load, and most target two-level inverters rather than the three-level NPC topology that dominates medium-voltage and high-power applications. Because the three-level switching pattern alters the harmonic structure of stator voltage and current, diagnostic thresholds calibrated on two-level systems cannot simply be transferred to NPC drives. Likewise, features calibrated at rated load can produce missed detections or false alarms at partial load unless their load-dependent behaviour is explicitly quantified.

The simulation framework models a 37 kW squirrel-cage induction motor fed by a three-level NPC inverter under closed-loop voltage-to-frequency (V/f) control, with space vector PWM at a switching frequency of 2,500 Hz and a DC bus voltage of 665 V. Seven fault categories are injected through physics-motivated parameter perturbation: broken rotor bar (BRB), air gap eccentricity (AGE), bearing degradation (BD), DC voltage imbalance (DCI), DC undervoltage (DCU), DC overvoltage (DCO), and mechanical overload (MO). Each fault is applied at three severity levels across thirteen discrete load levels, yielding a nominal 273 faulty-condition runs. Thirty-one cases in which severe faults caused mechanical instability were excluded, leaving 242 faulty-condition observations, which together with 46 healthy-baseline runs produced a 288-observation dataset that has been made publicly available.

From each five-second simulation, features were extracted from the steady-state window between three and five seconds, using signals from stator phase current, rotor mechanical speed, and electromagnetic torque. Nine signal-derived features spanning mechanical, electrical, and statistical categories were characterised, including mean rotor speed, speed ripple, RMS stator current, total harmonic distortion (THD), mean torque, torque ripple, kurtosis, skewness, and crest factor. The author is careful to note that the single-parameter perturbations are first-order representations: a real broken rotor bar also produces localised electromagnetic asymmetry and characteristic sideband harmonics that a uniform resistance increase does not reproduce, and physical bearing faults generate localised defect frequencies rather than uniformly elevated viscous friction.

A central contribution is a formal load-invariance classification of every feature-fault pair. Features were labelled load-invariant if their coefficient of variation across load levels was below 2 percent, partially load-dependent between 2 and 10 percent, and load-dependent at 10 percent or above. Within the studied simulation framework, mean rotor speed satisfied the load-invariance criterion for every fault type, making it a candidate for load-independent thresholding in this drive configuration, although the author stresses that speed does vary physically with load through induction-motor slip; the classification reflects only that this variation is small relative to the mean speed. By contrast, RMS stator current, THD, skewness, and speed ripple were load-dependent for all fault types, meaning threshold-based systems using these features would need load normalisation or operating-point-indexed lookup tables.

The DC-side fault models produced the most dramatic and most distinctive signatures. The DC voltage imbalance model, in which one of the two series DC sources is reduced, generated the largest feature deviations of any fault type. Skewness of the stator current reversed sign, reaching −0.53 under severe imbalance at 50 percent load compared with roughly +0.03 to +0.04 for the healthy motor, a reversal unique among the seven fault types and physically explained by the asymmetric inverter output attenuating one half-cycle of the current waveform. THD increased by 92.4 percent of the healthy mean from mild to severe conditions, the largest THD sensitivity observed. The undervoltage and overvoltage models, which scale both DC sources proportionally, produced opposing ripple signatures: DCU suppressed torque ripple by 32.2 percent and speed ripple by 32.9 percent, while DCO amplified torque ripple by 50.8 percent and speed ripple by 52.7 percent, both with skewness remaining positive. Together the three DC fault models form a differentiable triad based on skewness sign and ripple direction, a three-way discrimination the author reports has not previously been published for NPC motor drives.

Among the motor-side faults, the broken rotor bar model was identified primarily through mean rotor speed, which fell by 5.3 percent of the healthy mean from mild to severe conditions, amplified at higher loads, while RMS current and THD showed negligible severity sensitivity in this configuration. The air gap eccentricity model behaved in nearly the opposite manner: rotor speed stayed within 0.03 percent of the healthy baseline, but RMS current rose by 23.4 percent while THD fell by 20.7 percent, an opposing-trend feature pair unique among the seven fault types. The bearing degradation model, represented as elevated viscous friction, produced a systematic increase in mean electromagnetic torque of about 1.9 percent per severity step, a signature absent in all other fault models, along with a 12.8 percent rise in RMS current. Because the AGE and BD models share similar current and THD trends, the study suggests that torque-based features should accompany current-based ones to discriminate between them.

The most sobering result concerns mechanical overload. Because this fault is modelled purely as an increase in applied load torque, with no modification of any electrical or mechanical motor parameter, the motor simply moves to a higher operating point along the healthy characteristic curve. No extracted feature changed by more than 0.8 percent of the healthy mean across the full severity range, making overload functionally indistinguishable from normal heavy loading within the studied steady-state feature set. The author notes this conclusion is model-dependent: transient features during load changes, or fundamental-frequency current measurements rather than spectrum-wide RMS values, could potentially distinguish overload in a physical drive. Within the modelled scenario, however, overload protection would require direct current or torque threshold monitoring or transient-based detection rather than the steady-state features examined here.

The study carries substantial caveats that the author states explicitly. All results derive from simulation; no physical test bench or experimental measurements were used, and experimental validation on hardware is described as an essential next step. The dataset was generated under deterministic conditions without sensor noise, measurement noise, thermal drift, or mechanical tolerances, all of which would broaden the distributions of statistical features and reduce separability for faults with small sensitivities, such as mild bearing degradation and broken rotor bars. The severity levels are simulation scenarios whose physical plausibility is supported by prior literature rather than calibration on real hardware. All findings are specific to the three-level NPC inverter under scalar V/f control with space vector PWM; signatures may differ under field-oriented control, direct torque control, model predictive control, or different modulation strategies, and the generalisation of the load-invariance classifications to other control architectures has not been assessed.

Within those limits, the reference signature tables, feature sensitivity matrix, and load-invariance taxonomy offer practitioners a quantitative starting point for designing diagnostic thresholds and selecting feature subsets for data-driven classifiers in NPC-fed drives. The headline messages are practical ones: a single fixed threshold on a load-dependent feature invites false alarms or missed detections; mean rotor speed and, with one exception, kurtosis offer relatively stable indicators across the load range in this configuration; and DC-link faults leave rich, physically interpretable traces in current statistics that could be exploited without direct voltage measurement, provided the simulation-based findings survive experimental scrutiny.

Subject of Research: Load-dependent evolution of fault diagnostic signatures in a simulated three-level NPC inverter-fed induction motor drive

Article Title: Load-dependent fault signature evolution in three-level NPC inverter-fed induction motor drives via physics-based characterisation

Article References: File, S. G. (2026). Load-dependent fault signature evolution in three-level NPC inverter-fed induction motor drives via physics-based characterisation. Discover Electronics, 3(1), Article 146. https://doi.org/10.1007/s44291-026-00300-0

Image Credits: AI Generated

DOI: 10.1007/s44291-026-00300-0

Keywords: induction motor, fault diagnosis, NPC inverter, load dependence, broken rotor bar, air gap eccentricity, bearing degradation, DC-link faults, total harmonic distortion, skewness, condition monitoring, MATLAB simulation