Energy-Efficiency Boundaries of Edge Shallow ML versus 1D Deep Learning

How do runtime energy dissipation (microjoules per inference) and execution latency compare between MVMD-envelope spectrum shallow classifiers and direct 1D CNNs when deployed on Cortex-M microcontrollers processing non-stationary bearing vibration signals under varying motor speeds?

1. Introduction

The field initially organized around the urgent industrial need to prevent severe, untimely machinery breakdowns by continuously monitoring rotating equipment, particularly rolling element bearings, which are central points of failure [1, 2, 3]. Early diagnostic pipelines relied exclusively on manual feature extraction, operating under the assumption that expert-driven analytical techniques—such as wavelet and Hilbert-Huang transforms—were required to isolate specific fault patterns from mechanical vibrations [4, 5]. During the early 2000s, the observational scope expanded when researchers demonstrated that bearing damage also introduces distinct radial rotor movements and measurable load torque variations into electrical currents, prompting the field to monitor stator current alongside physical vibration [6].

As empirical validation deepened around 2010, researchers rigorously mapped the physical links between mechanical faults and signal components by modeling incipient damage as torque disturbances [7]. However, condition monitoring consistently faced a critical bottleneck: the diversity of operating conditions and heavy background noise made traditional, manual feature extraction highly labor-intensive and susceptible to human bias [8, 9]. This operational barrier caused the field to pivot toward early artificial intelligence methods. In 2015, the introduction of optimization deep belief networks, which automatically fine-tuned neural connection weights using stochastic gradient descent, demonstrated that complex faults could be identified robustly without relying on prior, manually engineered fault information [8].

2. Theoretical Background

Building on these early methods, the transition to automated, data-driven models fundamentally reorganized the discipline [10]. Around 2016 and 2017, Prognostics and Health Management (PHM) became the dominant conceptual framework, shifting the field’s objective from merely diagnosing existing faults based on statistical indicators [11] to constructing continuous health indicators that predict the remaining useful life (RUL) of mechanical components [12]. To achieve this automation, analysts stopped processing vibration solely as a one-dimensional time series. By converting raw data into two-dimensional time-frequency images using short-time Fourier and wavelet transforms, the field unlocked the ability to feed machine data directly into deep Convolutional Neural Networks (CNNs), bypassing the need for human-engineered parameters entirely [9, 13].

With automated feature extraction established, the field's focus expanded to address data transmission and measurement limits in industrial settings. Because high-frequency continuous monitoring generates massive datasets, researchers integrated compressed sensing (CS) techniques, changing the diagnostic pipeline by proving that deep sparse autoencoders could accurately classify bearing faults directly from highly compressed measurements without reconstructing the original signal [14]. Simultaneously, predictive methodologies expanded to alternative machine signatures, applying supervised machine learning regression models to acoustic emissions to correlate signal intensity with natural component wear [15].

By 2019, the assumption that deep learning models could be seamlessly deployed in real-world environments was challenged by the realization that training data (the source domain) and operational testing data (the target domain) often exhibit entirely different feature distributions [13]. This realization pushed researchers to adopt ensemble learning methods, combining architectures like random forests with CNNs to ensure classification generalization across varying industrial domains [13]. Furthermore, the immense computational cost of standard CNNs and the unreliability of manual parameter adjustments drove the development of lightweight deep residual networks based on separable convolutions [16] and automated parameter tuning utilizing optimization frameworks like simulated annealing [17].

Today, vibration-based condition monitoring is thoroughly defined by deep learning, marking the endpoint of the discipline's transition away from manual signal processing [3, 18]. The current momentum revolves around optimizing these intelligent diagnostic models to be "small, light, and fast" for highly efficient, end-to-end edge deployment on factory equipment [16]. As the discipline moves forward, researchers are actively shifting attention toward next-generation network architectures, working to integrate graph-based neural networks, transformers, and physics-informed machine learning to further refine predictive accuracy [18].

3. Statement of the Problem

Despite these conceptual advances, a critical methodological problem persists in practical implementation. The core unresolved debate is whether edge bearing monitoring should rely on classical signal processing features with shallow ML or end-to-end deep learning models. End-to-end deep learning eliminates manual feature engineering for complex faults but demands memory and compute resources that strain ultra-low-power edge microcontrollers.

Currently, there is limited evidence resolving this challenge. The retrieved literature relies heavily on standard benchmark datasets, such as CWRU bearing signals, or synthetic sensor data, offering limited validation on resource-constrained microcontrollers operating under real-world industrial noise and non-stationary loads. Furthermore, several deep learning frameworks rely on short-time Fourier transform spectrograms rather than pure raw time-series inputs, obscuring the performance boundaries between manual feature engineering and true end-to-end representation learning.

Advocates for feature-engineered shallow ML contend that pre-extracting spectral and wavelet features simplifies model complexity, enabling efficient execution on low-power IoT microcontrollers. The retrieved literature supports this perspective, demonstrating that pairing classical Fourier and wavelet descriptors with shallow neural networks or ensemble algorithms yields high anomaly detection performance with low energy overhead in embedded environments.

Conversely, proponents of end-to-end deep learning argue that deep architectures directly learn complex latent fault patterns from sensor data, avoiding human bias in manual feature design. Retrieved studies support this position by showing that lightweight deep networks and dual-stream architectures processing time-series and spectrogram inputs can effectively perform bearing fault diagnosis and remaining useful life prediction on edge devices.

4. Review of Prior Research

To fully understand this problem, it is necessary to examine the current state of condition-based monitoring and edge processing literature.

4.1 What is widely reported

The reliable operation of industrial manufacturing facilities depends critically on rotating machinery, wherein rolling element bearings are the primary source of motor failures and performance imperfections [1, 19, 20]. Consequently, condition-based monitoring systems that detect anomalous parameters before a failure occurs are essential for predicting breakdown situations, reducing maintenance costs, and maintaining production productivity [19, 21]. Vibration measurement remains the predominant method for diagnosing these malfunctions, providing direct insight into the mechanical condition of rotating equipment [18].

In recent years, the deployment of artificial intelligence algorithms has transitioned from centralized, cloud-based feature engineering and classification to edge processing frameworks [22]. Cloud-based computing poses substantial barriers for continuous industrial monitoring, including high data transmission costs, significant latency, and data privacy threats [22]. Processing vibration signals directly at the edge of the Industrial Internet of Things (IIoT) mitigates these issues. For example, edge processing systems that analyze and mix sensor signals locally before transmission can reduce the volume of transmitted data by approximately 95%, which heavily reduces power consumption and enables the use of battery-supplied IIoT nodes for continuous condition monitoring [23]. To achieve this, modern diagnostic frameworks integrate data from multiple sensors—including temperature, vibration, sound, and electrical current—with intelligent edge devices to identify distinctive fault patterns in real time [21]. Within this context, deep learning techniques are extensively applied to analyze vibration signals [18, 24], establishing an ongoing engineering challenge regarding how best to process raw sensor data on resource-constrained embedded systems.

4.2 What the competing positions rest on

Position A — Feature-Engineered Shallow ML

This position argues that manual, classical signal processing techniques combined with shallow machine learning algorithms offer an optimal pathway for early fault detection on edge devices. The approach relies on explicitly isolating diagnostic features from nonstationary vibration signals before applying classification algorithms. This perspective is supported by monitoring frameworks that deploy smart sensor-based data acquisition systems utilizing specialized signal decomposition [19]. Specifically, techniques such as modified variational mode decomposition (MVMD) are used to break down vibration signals into a series of intrinsic mode functions (IMFs) [19]. By applying correlation kurtosis to select the most representative IMFs, constructing an analytical signal, and performing envelope spectrum analysis, these systems can successfully predict bearing faults and identify representative features directly on smart sensors [19].

Position B — End-to-End Deep Learning

This position contends that explicitly extracting fixed features demands significant computational overhead that prevents effective real-time application, arguing instead for convolutional neural networks (CNNs) that combine feature extraction and classification into a single learning body [20]. Proponents implement 1D CNNs that accept raw vibration signals as direct inputs, completely eliminating the need to run separate feature extraction algorithms [20]. This end-to-end capability allows for lightweight, edge-implementable architectures that avoid the memory and processing speed constraints associated with complex, multi-stage pipelines [22]. Embedded deep learning systems, such as simple CNNs deployed on fan drives, achieve optimal accuracy within 30 epochs without conventional machine learning feature extraction [25]. To manage hardware limitations, some edge intelligence methods employ parameter transplantation, where lightweight network architectures (such as S-AlexNet) are trained offline on a PC to establish optimal weights and biases, which are subsequently transplanted into a low-cost embedded system for real-time condition monitoring [26].

4.3 Foundational work behind the debate

The debate over analytical methodologies builds upon a foundational history of diverse signal processing applications. Early diagnostic techniques heavily utilized data mining, expert systems, and intelligent diagnostics operating across distinct time, frequency, and time-frequency domains to monitor induction motor rolling bearings [1]. These classical methods required deep domain expertise to extract meaningful indicators of incipient damage.

The eventual shift toward artificial intelligence fundamentally altered this workflow. Researchers began leveraging machine learning and deep learning algorithms, matching success seen in image processing and speech recognition to the domain of machinery fault detection [24]. This transition was heavily facilitated by the widespread availability of open-source reference datasets. In particular, the Case Western Reserve University (CWRU) bearing dataset became a central pillar for the field, serving as the accepted standard reference for validating various deep learning fault detection models [24]. The reliance on standardized, pre-collected data allowed researchers to compare pure algorithmic performance, setting the stage for the current competition between manual feature extraction pipelines and unified, automated deep learning architectures.

4.4 Adjacent debates and neighbouring evidence

Similar methodological debates exist in adjacent condition monitoring and edge processing domains, offering transferable parallels to vibration-based bearing diagnosis.

In the monitoring of acoustic emission (sound) signals, supervised machine learning techniques are applied to predict the remaining useful life of components. Methods such as support vector machine regression, Gaussian process regression, and multilayer artificial neural networks are utilized to correlate specific acoustic emission features—such as root mean square and signal intensity estimators—with early damage from natural wear [15]. When deployed on embedded devices with IoT connectivity, shallow machine learning models processing acoustic emissions, such as fine decision trees, achieve classification accuracies of 96.1% for fault detection in rotating machinery [27].

In other edge-based settings, deep learning frameworks still rely on explicit time-frequency transformations rather than raw time-series inputs. For instance, intelligent microcontroller-unit (MCU) grippers designed for object shape identification process acceleration and vibration signals by first transforming the raw data into images using the short-time Fourier transform (STFT) before adopting a CNN to extract features and classify the grasped objects [28].

Parallel advancements in smart factory fleet monitoring also demonstrate the viability of edge AI. Generic real-time fault diagnosis systems for autonomous transfer vehicles utilize edge AI integrated with data distributor middleware (such as FIWARE) to facilitate continuous equipment monitoring and the early generation of fault alarms [29]. Furthermore, in the domain of wind turbine planetary gearboxes, early fault signals are particularly difficult to detect. Here, deep belief networks (DBNs) optimized by quantum particle swarm optimization (QPSO) algorithms are applied to self-powered wireless sensors to extract deeper and more identifiable features from slight fault signals, which are then passed to optimized least squares support vector machines for final diagnosis [30].

4.5 Contradictory and inconsistent findings

The retrieved literature presents distinct contradictions regarding the computational cost and necessity of feature extraction on edge devices. Eren [20] explicitly argues that extracting fixed features at every evaluation imposes a significant computational cost that prevents the use of traditional decision support systems in real-time applications, thereby necessitating unified 1D CNNs. Contradicting this assessment, Zhang et al. [19] successfully propose and implement a highly complex signal processing pipeline—comprising modified variational mode decomposition, correlation kurtosis, and envelope spectrum analysis—expressly designed for real-time bearing fault detection on a smart wireless sensor data acquisition system.

Furthermore, there is inconsistency regarding what constitutes a true "end-to-end" deep learning architecture for edge devices. While some researchers emphasize feeding pure raw vibration data directly into 1D CNNs to bypass any pre-processing [20], other frameworks advocating for lightweight CNN edge implementation rely heavily on transforming sensor measurements into 2D representations prior to neural network processing. For example, Mukherjee and Tallur [22] utilize vibration spectrograms to feed their lightweight CNN for fault extraction.

5. Research Gap and Purpose of the Study

The synthesis of prior research highlights a critical gap in hardware-level evaluation. The unresolved debate centers on whether explicitly pre-computing classical time-frequency decompositions consumes more or less microcontroller energy than executing 1D end-to-end convolutional neural networks under non-stationary operating conditions. Prior studies evaluate execution performance on static benchmark datasets or rely on offline PC training, lacking real-time hardware microjoule-per-inference energy measurements under fluctuating speed and load conditions on ultra-low-power microcontrollers. Existing evidence leaves unresolved whether pre-extracting frequency features or evaluating multi-layer convolutional networks consumes more energy on ultra-low-power microcontrollers, as studies rely heavily on desktop training or static reference datasets [19, 20, 24, 26].

This study presents an opportunity for researchers to construct a physical testbed with embedded microcontrollers measuring dynamic current draw to directly benchmark the power consumption and latency profiles of MVMD-shallow classifiers against raw 1D-CNN architectures under transient motor speed profiles.

6. Research Questions and Hypotheses

To address this gap, the following research questions and hypotheses are proposed. How do runtime energy dissipation (microjoules per inference) and execution latency compare between MVMD-envelope spectrum shallow classifiers and direct 1D CNNs when deployed on Cortex-M microcontrollers processing non-stationary bearing vibration signals under varying motor speeds?

7. Method

The empirical investigation will follow a structured methodology to test these hypotheses.

7.1 Theoretical Framework

The design is grounded in the theoretical framework of Condition-Based Monitoring (CM) and edge computing constraints in the Industrial Internet of Things (IIoT). CM theory posits that mechanical degradation in rotating equipment manifests as distinct physical anomalies in vibration signatures, which can be computationally isolated to predict failure [1, 15, 24]. The debate centers on how this computation should be distributed within highly constrained hardware environments. Position A rests on classical signal processing theory, positing that explicitly deconstructing non-stationary signals into intrinsic mode functions (e.g., MVMD) isolates pure fault features, theoretically requiring simpler, lower-energy shallow classifiers to diagnose [19, 31]. Position B relies on deep representation learning theory, arguing that explicitly computing classical transformations introduces prohibitive overhead, whereas unified 1D Convolutional Neural Networks (CNNs) inherently learn robust feature maps from raw sequences, streamlining the inference pipeline to save energy and memory [16, 20, 22].

7.2 Variables

7.3 Operational Definitions

7.4 Data

7.5 Methodology

7.6 Analysis Plan

Full specification

The primary models are linear mixed-effects models estimated via Restricted Maximum Likelihood (REML) to account for the repeated-measures nature of the hardware benchmark:

Energy Dissipation Model:

$E_{ij}={\beta }_0+{\beta }_1{\text{Model}}_{ij}+u_{0j}+u_{1j}{\text{Model}}_{ij}+{ϵ}_{ij}$

Execution Latency Model [established, Inferred]:

$L_{ij}={\gamma }_0+{\gamma }_1{\text{Model}}_{ij}+v_{0j}+v_{1j}{\text{Model}}_{ij}+{\eta }_{ij}$

Robustness checks: Additional checks will involve adding fixed effects for specific compiler optimization levels (e.g., `-O2` vs `-O3`) if different flags are piloted.

Power-analysis assumptions: The power analysis will require pilot testing on the specific Cortex-M board to estimate the intra-class correlation (ICC) of energy consumption across signal windows.

8. Expected Contributions

The outcomes of this methodology are anticipated to yield significant industrial implications. If the prevailing assumption that direct end-to-end neural networks are inherently more energy-efficient on microcontrollers than classical feature decomposition paired with shallow models proves false, firmware engineers will continue selecting sub-optimal model architectures for edge diagnostics. Plant maintenance teams deploying self-powered wireless monitoring nodes would experience unexpected power depletion during non-stationary motor operations, resulting in premature device failures or costly maintenance cycles. Consequently, hardware integrators would misallocate compute budgets between signal pre-processing routines and inference execution, failing to maximize battery longevity on factory floors.

9. Limitations

Despite its rigorous design, the proposed study operates within several defined boundaries. The design cannot identify the predictive accuracy of these models relative to true mechanical failure rates in the wild, how energy consumption scales on drastically different architectures like FPGA or Cortex-A, or long-term model degradation or drift due to hardware aging.

Furthermore, the available evidence cannot settle whether differences in energy dissipation stem from inherent algorithmic complexity or uneven compiler optimization (e.g., `-O3` flags affecting standard FFT routines differently than neural network weights). This requires researcher judgement and methodological review to ensure compiled firmware binaries are uniformly optimized. Hardware benchmarking is only meaningful if both models complete diagnostic validation. Because current evidence often relies on static benchmarks, pilot testing is required to establish that both candidate architectures successfully detect faults before their power profiles are compared. Diagnostic validation is strictly a pre-experimental go/no-go qualification step and will not be analyzed or reported as a dependent outcome [Inferred].

10. Feasibility

The successful execution of this research depends on specific practical requirements. Access to a hardware testbed equipped with Cortex-M microcontrollers is required. This is not settled because laboratory inventory is unknown; it must be verified by checking for suitable STM32 or equivalent development boards. Access to dynamic power measurement instrumentation with microjoule resolution is required, which is unverified and must be checked by confirming access to a high-frequency power analyzer, current sense amplifier, or advanced oscilloscope. A dataset of non-stationary bearing vibration signals under varying motor speeds must be accessible; this remains uncertain and must be verified by locating an open dataset containing transient motor profiles or preparing a physical motor rig to generate the data. Source code or pre-trained models for MVMD-shallow and 1D CNN architectures must be obtainable, requiring verification by reviewing open-source repositories for existing lightweight implementations. Finally, verification is required to ensure that any synthetic dataset manipulation genuinely replicates the physical transients of varying motor speeds and does not merely add stationary noise to static data [Inferred].

Overall feasibility rests on confirming access to physical measurement instruments and representative datasets; if the hardware testbed or non-stationary data cannot be sourced, the study can proceed via cycle-accurate simulation and synthetic non-stationary datasets provided they are verified to genuinely replicate physical transients [Inferred].

References