A research team has introduced an open-source Python framework designed to automate adaptive sampling for industrial condition monitoring, demonstrating data reductions of up to 31 percent on motor current measurements while maintaining time-domain waveform fidelity. The software package, named DSRA-PMLO, implements an automated variant of the Dynamic Sampling Rate Algorithm optimized through parametric machine learning techniques. Developed by Hatem Algabroun, Sisi Wu, Abdelhakim Deboucha, and Muntaser Mohamed Nuttah, the tool addresses the data deluge generated by continuous high-rate sensing in industrial environments. By adjusting the sampling frequency dynamically based on signal variations rather than capturing continuous streams at a static rate, the system selectively allocates acquisition resources where rapid changes occur while thinning samples during steady-state intervals.
Industrial manufacturing accounts for roughly 16.18 trillion United States dollars in economic value, representing approximately 15 percent of global gross domestic product. Within these production environments, unplanned machine downtime and operational interruptions can impose substantial financial burdens, with unscheduled downtime estimated to cost manufacturers approximately 50 billion dollars each year. Predictive maintenance strategies have emerged as an essential approach to mitigate these losses by improving asset availability and reducing maintenance expenditures. Published estimates indicate that predictive maintenance can boost asset productivity by 25 percent, reduce mechanical breakdowns by 70 percent, and decrease overall maintenance costs by 25 percent. The core operational foundation of predictive maintenance is continuous condition monitoring, which provides timely data regarding equipment health to facilitate early anomaly and fault detection before catastrophic mechanical degradation occurs.
Condition monitoring is especially critical for electrical induction motors, which serve as foundational workhorses across modern industrial infrastructure and consume a substantial share of total industrial electrical power. Among various diagnostic methods, Motor Current Signature Analysis has become widely adopted because internal mechanical and electrical defects generate characteristic frequency components within stator current signals. Traditional current signature analysis relies almost entirely on fixed-rate sampling protocols, continuously recording sensor readings at a uniform sampling interval regardless of underlying signal dynamics. While uniform acquisition ensures consistent temporal coverage, it produces massive volumes of redundant data during steady-state machine operation. This excessive data acquisition strains network bandwidth, exhausts storage capacity, and increases processing workloads and energy consumption, particularly across battery-powered wireless nodes and resource-constrained Industrial Internet of Things deployments.
Adaptive sampling presents an alternative by dynamically tuning the temporal spacing between consecutive measurements according to real-time signal characteristics. By increasing measurement frequency across regions displaying rapid transients or abrupt shifts and reducing sampling density along smooth or steady trajectories, adaptive protocols compress overall data volume while preserving diagnostically vital waveform features. Although parametric methods such as the baseline Dynamic Sampling Rate Algorithm offer lightweight computation suitable for edge deployment, their practical adoption has been hindered by complex manual configuration requirements. Tuning operational parameters historically required domain expertise and extensive historical data records. While an optimization extension known as DSRA-PMLO demonstrated that parameter selection could be automated under reconstruction-error limits, configuring initial search boundaries still demanded manual oversight, leaving an unmet need for fully accessible, automated, and publicly verifiable implementations.
To overcome these practical barriers, the researchers developed DSRA-PMLO as a modular, object-oriented software package written entirely in Python and released under the permissive MIT license. Rather than relying on heavyweight deep learning frameworks such as TensorFlow or PyTorch, the implementation is built upon lightweight, standard scientific computing libraries including NumPy, SciPy, Pandas, and Matplotlib. The software architecture draws inspiration from the classical three-layered architectural model, dividing execution tasks into six specialized functional modules: Signal Handler, Error Controller, Optimization Engine, Sampler, Reconstructor, and Visualizer. The architecture cleanly separates computationally intensive offline parameter optimization from runtime data acquisition, ensuring that the resulting online sampling decisions rely only on simple algebraic expressions executed with constant time complexity per sample.
The package processes univariate continuous signals provided in standard comma-separated or text formats through a single configuration interface script. The Signal Handler reads raw measurements and partitions the series into dedicated training and testing segments using a fixed 60-to-40 percent split. Users define a reconstruction error ceiling using the Mean Arctangent Absolute Percentage Error metric, chosen for its scale-independent and intuitive interpretation. The Error Controller enforces this threshold constraint during the optimization phase, which offers two distinct operating paths. In automated mode, the Optimization Engine executes a two-stage search starting with a successive coarse-to-fine grid search to isolate promising regions in the exploration and sensitivity parameter space. Dual Simulated Annealing then refines these parameters by combining localized fast searches with generalized annealing mechanisms that explore broader candidate spaces at elevated virtual temperatures before converging upon an optimal configuration.
Once optimal exploration and sensitivity parameters are determined, the Sampler applies the resulting algebraic rules to downsample the held-out test signal dynamically. The Reconstructor subsequently rebuilds the complete continuous waveform through linear or quadratic interpolation, and the Visualizer computes final performance statistics alongside graphical plots comparing reconstructed traces against raw sensor signals. The deployed sampling rule requires only a small constant state and a fixed number of basic arithmetic calculations, achieving an online computational complexity of order one per decision and order N across an entire monitoring window. This low footprint ensures that resource-constrained sensing hardware or embedded microcontrollers can execute sampling interval updates in real time without exceeding strict latency or energy budgets.
The authors evaluated the framework across two distinct testing environments, examining performance across five user-specified reconstruction error thresholds ranging from 1.0 percent to 5.0 percent. The first evaluation utilized a synthetic benchmark consisting of twenty-five sinusoids with randomized phases and unit amplitudes, spanning frequencies from 10 Hz to 2000 Hz to simulate complex motor current spectra. Evaluated at a 20 kHz baseline sampling rate over a one-second interval, the synthetic sequence yielded 12,000 training samples and 8,000 testing samples. Across five independent random seeds, the automated engine delivered consistent sampling reductions that grew from 14 percent at a 1.0 percent error threshold up to 25 percent at a 5.0 percent threshold. The retained measurements clearly mirrored signal dynamics, concentrating sampling points within sharp gradient transitions while sparsely populating smoother crests and troughs.
The second evaluation investigated physical motor current data acquired from a 0.38 kW laboratory induction motor instrumented with an ACS712 Hall-effect current sensor connected to an STM32 microcontroller. The physical measurements were recorded at a baseline rate of 1 kHz across three operational conditions: a healthy no-load state, an unladen condition with a severed broken rotor bar, and a light-load condition with the broken rotor bar. From steady-state operating records, segments of 400 samples were partitioned into 240 training points and 160 testing points. When evaluated on this physical benchmark, DSRA-PMLO achieved sampling reductions between 6 percent and 28 percent for the no-load baseline, between 8 percent and 28 percent for the no-load faulty motor, and between 9 percent and 31 percent for the loaded motor exhibiting a broken rotor bar.
The experimental trials highlighted important performance trade-offs and fundamental operational limitations inherent to offline parameter optimization. On held-out testing partitions, the measured reconstruction error slightly exceeded the user-defined threshold in several test cases. In the physical dataset, the no-load broken rotor bar condition registered a 2.4868 percent error against a 2.0 percent target and 4.8140 percent against a 4.0 percent target, while the no-load healthy motor reached 7.0632 percent error at the 5.0 percent setting. The researchers noted that these deviations illustrate a generalization gap: because parameters are optimized strictly over training segments, unseen test sequences displaying faster local variations or sharp zero crossings can yield coarser approximations. Consequently, the authors caution that users should validate optimized parameters on representative held-out signals and consider lowering input thresholds before field deployment to compensate for potential error expansion.
Subject of Research: Open-source adaptive sampling software for industrial condition monitoring
Article Title: DSRA-PMLO: An open-source Python package for adaptive sampling in industrial condition monitoring
Article References: Algabroun, H., Wu, S., Deboucha, A., & Nuttah, M. M. (2026). DSRA-PMLO: An open-source Python package for adaptive sampling in industrial condition monitoring. SoftwareX, 36, Article 103214. https://doi.org/10.1016/j.softx.2026.103214
Image Credits: AI Generated
DOI: 10.1016/j.softx.2026.103214
Keywords: adaptive sampling, condition monitoring, predictive maintenance, motor current signature analysis, induction motors, dynamic sampling rate algorithm, dual simulated annealing, signal reconstruction, data reduction, industrial internet of things, edge computing, broken rotor bar, MAAPE, Python, NumPy, SciPy, open-source software
News Source: Denise Maddox. (November 12, 2026). Open-Source Python Tool Automates Adaptive Sampling for Industrial Motor Monitoring. Scienmag.
