Answers are grounded in this paper and cite their sources.
Energy-Efficient Edge AI for Real-Time Anomaly Detection in IoT-Based Embedded Systems A Reproducible Comparison of Lightweight Detectors on Public Sensor Benchmarks Rachit Debbad [India] Rishabh Misra[rishabheco@gmail.com] Abstract—Detecting abnormal sensor readings on the node itself saves bandwidth and latency, yet learned detectors are rarely compared on energy and memory cost alongside accuracy. This paper benchmarks eight lightweight unsupervised detectors — a rolling z-score, an EWMA control chart, PCA reconstruction, Isolation Forest, Local Outlier Factor (LOF), a one-class SVM, and a small autoencoder in FP32 and INT8 form on two public datasets: 52 labelled series of the Numenta Anomaly Benchmark (NAB; mixed server, traffic and sensor signals) and 34 runs of the Skoltech Anomaly Benchmark (SKAB; an eight-channel water-circulation sensor testbed). Models are fitted on the first 15% of each series without anomaly labels, and alarm thresholds come from training scores only. We report threshold-free and fixed-threshold metrics, bootstrap confidence intervals, paired tests, model memory, operation counts and a first-order analytic energy estimate (modelled, not measured on hardware). On NAB the learned detectors reach macro AUROC 0.56–0.58 and cannot be told apart statistically, but all clearly exceed the statistical baselines (0.49–0.50). On SKAB they reach 0.74–0.85 against 0.49–0.52. Cost differs by up to two orders of magnitude: under our model the INT8 autoencoder uses about 6× less energy than its FP32 form (4.6–20× across memory-cost assumptions) and 30–200× less than LOF, while weight quantization changes per-series AUROC by about 0.001. Fixed percentile thresholds transferred poorly on both datasets. The results favour compact reconstruction-based models for edge anomaly detection and identify hardware measurement and drift-aware thresholding as the necessary next steps. Keywords: edge AI; TinyML; anomaly detection; Internet of Things; embedded systems; autoencoder; quantization; energy efficiency; Numenta Anomaly Benchmark; Skoltech Anomaly Benchmark 1. Introduction IoT deployments place sensors where power is scarce and connectivity is costly. Streaming every raw reading to a cloud service for analysis consumes bandwidth and radio energy and adds latency, which is the usual motivation for moving analytics to the network edge [4]. For abnormal-reading detection (a failing bearing, a drifting thermistor, a blocked valve) the node itself must often decide, in real time, whether the latest window of samples is unusual. The anomaly-detection literature offers many methods [1]–[3], but evaluations typically rank them by accuracy alone. On a microcontroller the questions are different: how many bytes of model must live in flash or SRAM, how many arithmetic operations and memory fetches each decision costs, and whether the method survives reduced numeric precision. Work on resource-constrained learning, such as compact tree and nearest-neighbour models [13], [14], embedded inference runtimes [15] and the MLPerf Tiny benchmark [16], addresses the deployment side. It does not directly tell a designer how common detector families compare on public sensor-style data under one protocol, with quality and cost in the same table. This study provides such a hardware-free, reproducible comparison. Its contributions are: 1. A chronological, label-free training protocol and a common evaluation harness applied to two public datasets: 52 series of NAB [17] and 34 runs of the multivariate SKAB sensor testbed [26]. 2. A head-to-head comparison of eight lightweight detectors on threshold-free and fixed-threshold metrics, with bootstrap confidence intervals and paired significance tests. 3. A transparent cost model (memory footprint, multiply and add counts, estimated energy) that exposes the quality–cost trade-off, including an INT8 autoencoder and a sensitivity analysis of both the energy coefficients and the main design parameters. 4. An explicit account of limitations: modest absolute accuracy on NAB, poor transfer of percentile thresholds, small training sets on SKAB, and energy values that are modelled rather than measured. Section 2 reviews related work, Section 3 describes the methods and cost model, Section 4 the experimental setup, Section 5 the results, Section 6 discusses them together with limitations, and Section 7 concludes. 2. Related Work Anomaly-detection foundations. Chandola et al. [1] organise the field by data type and technique, and Blázquez-García et al. [2] do the same for time series, distinguishing point, subsequence and pattern anomalies. Pang et al. [3] review deep approaches. The detectors used here are classical representatives: control charts such as the EWMA chart [5]; isolation-based [6], density-based [7] and support-based [8] outlier detectors; and reconstruction-error methods, including autoencoders for sensor data [9]. Lightweight and on-device learning. Quantization-aware integer inference [11] and pruning with compression [12] reduce neural-network cost. Bonsai [13] and ProtoNN [14] are compact learners aimed at devices with kilobytes of memory. TensorFlow Lite Micro [15] provides an inference runtime for microcontrollers, and MLPerf Tiny [16] standardises tiny-model benchmarking, with an anomaly-detection task among its workloads. Kitsune [10] shows that ensembles of small autoencoders can run online for network intrusion detection. These works motivate our cost axes, but none compares the classical detector families above on public sensor-style data with quality and modelled cost side by side. Benchmarks and evaluation. The Numenta Anomaly Benchmark [17] provides labelled real and artificial streams with an application-oriented scoring scheme and was used to evaluate a streaming detector in [18]. SKAB [26] offers multivariate recordings from a water-circulation testbed with labelled fault episodes. Wu and Keogh [19] argue that several widely used time-series anomaly benchmarks suffer from triviality, mislabelling and other flaws that can exaggerate progress. We take that critique seriously: we avoid claims of state-of-the-art performance, report threshold-free metrics with confidence intervals, and caution against over-reading small differences. Energy modelling. Direct power measurement needs target hardware. A common first-order alternative counts arithmetic operations and memory accesses and weights them by per-operation energies, such as those summarised by Horowitz [20]. We adopt this approach and state its limits in Section 6.2. 3. Methodology 3.1 Problem formulation and pipeline A sensor emits a stream xt (a vector of C channels for SKAB). At each step the node normalises the sample with statistics from the training segment, forms a window of the last W normalised samples, and computes an anomaly score st. An alarm is raised when st exceeds a threshold τ (Eqs. 1–3). zt = (xt − μtr) / σtr (1) wt = (zt−W+1, …, zt) (2) alarmt = 1[ st > τ ] (3) Here μtr and σtr are the training-segment mean and standard deviation (per channel for SKAB), and τ is the 99th percentile of the scores st over the training segment. No anomaly labels are used for fitting or for choosing τ. Fig. 1 shows the pipeline. Fig. 1. Processing pipeline of the embedded anomaly detector. 3.2 Detectors Table 1 lists the eight detectors. The two statistical baselines process the normalised stream directly (for SKAB, per channel, with the maximum over channels as the score); the other methods use the window as a feature vector (W = 16 samples for NAB; W = 8 samples × 8 channels = 64 features for SKAB). Implementations use scikit-learn [21] defaults except where stated. Hyperparameters were fixed a priori and not tuned on the evaluation data. Table 1. Detector configurations (NAB / SKAB where they differ). Detector Configuration Anomaly score Rolling z-score trailing 64 samples, previous samples only |z − mean| / (std + 10⁻³); max over channels EWMA residual [5] smoothing factor 0.1 |z − EWMA| / (EW std + 10⁻³); max over channels PCA reconstruction 4 / 8 components mean squared reconstruction error Isolation Forest [6] 50 trees, 64 samples per tree negated scikit-learn score LOF [7] k = 20, novelty mode, ≤ 1000 stored windows negated LOF score One-class SVM [8] RBF kernel, gamma = "scale", ν = 0.05, ≤ 1000 windows negated decision function Autoencoder (FP32 / INT8) 16-8-4-8-16 / 64-16-8-16-64, ReLU hidden layers, linear output, Adam [22], learning rate 0.005, ≤ 300 epochs mean squared reconstruction error When a NAB series has more than 1000 training windows, 1000 are drawn at random (fixed seed) for fitting. SKAB training segments contain 104–192 windows (median 164) and are used in full. The z-score and EWMA baselines need no fitting. 3.3 INT8 autoencoder The INT8 variant applies post-training symmetric per-tensor quantization to every weight and bias matrix of the trained FP32 autoencoder, with scale s = max|W|/127 and values round(W/s) clipped to [−127, 127]. Accuracy is evaluated with the de-quantized weights. This simulates weight quantization only; activations stay in floating point, so end-to-end integer inference on a device may behave differently. The cost model treats the INT8 model as using 8-bit multiply-accumulates and 8-bit parameter storage. 3.4 Cost model For each detector we count stored bytes, multiplications, additions (comparisons count as additions) and 32-bit-word parameter reads per inference. The estimated energy per inferen
No takes yet. Share an insight, caveat, or question.
Debbad et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: