NDT 2025 , 3 , 7
4of 12
While raw parameters such as voltage, current, and speed provide valuable operational insights, calculated values offer deeper system-level diagnostics [6]. For instance, apparent power (kVA) and efficiency (%) can be used to assess energy performance, and torque estimation can be used for mechanical load analysis as shown in Table 3 and applied in the later Python example [5,6]. Thus, VFD data are not only a snapshot of motor operation but a comprehensive diagnostic tool for fault detection, performance optimization, and predictive maintenance [4].
Table3. Examples of calculated metrics.
Calculated Metric
Formula
V × I × 3 √ 1000
Apparent power (kVA)
Apparent power =
Power output ( kW ) Apparent power ( kVA ) ×
E f f iciency =
Efficiency (%) Torque (Nm)
100
Power output ( kW ) × 9500 Motor speed ( RPM )
Torque =
Energy consumption (kWh) Consumed = Power output ( kW ) × Run time ( hours ) Rich data, whether in their raw or derived state, present a significant challenge for condition monitoring due to the imbalance between normal operational data, which com- prise the majority of collected data and the rare occurrences of abnormal conditions [8,9]. To address this issue, ML, a subset of AI, leverages training data to compare with collected data, enabling pattern recognition through selected algorithms. Models based on anomaly detection algorithms, recurrent neural networks (RNNs), and convolutional neural net- works (CNNs) analyze historical VFD datasets to identify failure patterns [8,9]. This data imbalance can be mitigated through techniques such as synthetic data generation or by transferring relevant data from adjacent applications [8,9]. Beyond data transformation, traditional ML classifiers, which include supervised, unsupervised, and semi-supervised models, have a crucial role in predictive maintenance. Decision trees, random forests, and support vector machines (SVMs) are widely used for anomaly detection in industrial systems [6]. Clustering algorithms such as k-means and Gaussian mixture models (GMMs) provide unsupervised methods for identifying abnormal motor behavior, while semi-supervised learning techniques utilize healthy operational data to detect faults in real time, offering solutions when labeled fault data are scarce [8,9]. Deep learning further enhances predictive maintenance and NDT applications by autonomously extracting patterns from high-dimensional sensor data. Unlike traditional ML, deep learning models continuously train on real-time motor and VFD parameters, improving their ability to detect anomalies before they escalate into critical failures [8,9]. While CNNs and RNNs analyze time-series VFD data to detect failure patterns, autoen- coders and generative adversarial networks (GANs) enhance anomaly detection in highly imbalanced datasets. A deep learning-based fault diagnosis system for centrifugal pumps, examined by Turunen et al. [5], successfully detected cavitation and metal-to-metal contact using only VFD torque estimation data, demonstrating the untapped potential of drive- based diagnostics. By integrating multimodal sensor data and deploying lightweight deep learning models on edge devices, industries can enhance predictive maintenance strategies, reduce false positives, and optimize equipment reliability. Beyond fault detection, these expanded metrics reconcile the gap between electrical and mechanical failure analysis in NDT [9]. Rather than relying solely on external sensors, trends in torque fluctuations, power loss, and efficiency drops serve as early indicators of potential failures such as bearing degradation, misalignment, or insulation wear [4,18]. Sensorless fault detection using VFD data has effectively diagnosed cavitation in pumps and metal-to-metal wear in bearings, illustrating how critical failure modes can be identified
Made with FlippingBook interactive PDF creator