Anomaly Detection and Predictive Maintenance on Sensor Data
Detecting abnormal behaviour in machine and process sensor time series to warn of faults before downtime.
Tier Grades
Tier A: Frontier AI labs
feasible
IBM, Microsoft and NEC build detectors and industrial monitoring products.
Tier B: Indonesian AI labs
feasible
Detectors of about 1M parameters (TSPulse) need little compute; Indonesian lab work is not in the sources.
Tier C: Institution units such as ITB
feasible
Anomaly Transformer came from a 4-author team.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
An AI predictive maintenance proof of concept has already run in Indonesia (NEC SIAT at Kalbe Farma, two of three targets met).
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeTSPulse: pre-trained models of about 1M parameters that run without a GPU; Anomaly Transformer gives no compute.
- DataTSPulse reports +20% on the TSB-AD anomaly detection leaderboard across over 75 datasets; NEC SIAT proof of concept at Kalbe Farma (April to June 2026) met two of three targets.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 8 authors on TSPulse; 4 on Anomaly Transformer.
- Time horizonMonths.
- Shelf-life signalsTiny pre-trained detectors (TSPulse, 1M parameters) beating models 10 to 100 times larger, Indonesian industry proofs of concept facilitated by the Ministry of Industry (NEC SIAT at Kalbe Farma)
- Pipeline roleTier A and vendors release detectors; Tier C benchmarks them on local plant data; Tier D runs predictive maintenance proofs of value in factories.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
NEC
Tsinghua University
Carnegie Mellon University
IBM
MLCommons
University of Notre Dame
Sungkyunkwan University
Microsoft AILinked models
Granite TSPulsetime seriesexemplarAnomaly Transformertime seriesexemplarSystem Invariant Analytics Technology (SIAT)time seriesexemplarMOMENT (MOMENT-1-large)time seriesproduced byTimesNettime seriesproduced bySpectral Residual (SR and SR-CNN)time seriesproduced byMSCREDtime seriesproduced byARTimetime seriesproduced byCAD OSEtime seriesproduced byConv-AEtime seriesproduced byDeepAnTtime seriesproduced byIsolation foresttime seriesproduced byMLPerf Tiny Dense AE (ToyCar)TinyMLproduced byMSET (Multivariate State Estimation Technique)time seriesproduced byNumenta HTMtime seriesproduced byOmniAnomalytime seriesproduced by
FOMO-AD (Visual Anomaly Detection)TinyMLproduced byPaAnotime seriesproduced byPCAtime seriesused by
ST arc fault detection models (st_conv, st_dense)TinyMLproduced by(Sub)-PCAtime seriesused byLinks
R&D TopicAnomaly Detection and Predictive Maintenance on Sensor Data
ModelsGranite TSPulseIBM, time seriesSystem Invariant Analytics Technology (SIAT)NEC, time seriesAnomaly TransformerTsinghua University, time seriesARTimetime series+17 more
Granite TSPulseIBM, time seriesSystem Invariant Analytics Technology (SIAT)NEC, time seriesAnomaly TransformerTsinghua University, time seriesARTimetime series+17 more
Models 21
System Invariant Analytics Technology (SIAT)NEC, time seriesexemplarFOMO-AD (Visual Anomaly Detection)Edge Impulse, TinyMLproduced byGranite TSPulseIBM, time seriesexemplarMOMENT (MOMENT-1-large)Carnegie Mellon University, time seriesproduced byMLPerf Tiny Dense AE (ToyCar)MLCommons, TinyMLproduced byMSCREDUniversity of Notre Dame, NEC, time seriesproduced byPaAnoSungkyunkwan University, time seriesproduced byST arc fault detection models (st_conv, st_dense)STMicroelectronics, TinyMLproduced byTimesNetTsinghua University, time seriesproduced byAnomaly TransformerTsinghua University, time seriesexemplarConv-AEtime seriesproduced byMSET (Multivariate State Estimation Technique)time seriesproduced by+9
Spectral Residual (SR and SR-CNN)Microsoft AI, time seriesproduced by(Sub)-PCAtime seriesused byARTimetime seriesproduced byCAD OSEtime seriesproduced byDeepAnTtime seriesproduced byIsolation foresttime seriesproduced byNumenta HTMtime seriesproduced byOmniAnomalytime seriesproduced byPCAtime seriesused byPapers
- LSTM-based Encoder-Decoder for Multi-sensor Anomaly DetectionPankaj Malhotra et al., ICML 2016 (per arXiv comment), 2016 founding
- TranAD: Deep Transformer Networks for Anomaly Detection in Multivariate Time Series DataShreshth Tuli et al., arXiv preprint, 2022 state of the art
- Anomaly Transformer: Time Series Anomaly Detection with Association DiscrepancyJiehui Xu et al., arXiv preprint, 2021 state of the art
- Current Time Series Anomaly Detection Benchmarks are Flawed and are Creating the Illusion of ProgressRenjie Wu et al., 38th IEEE International Conference on Data Engineering (ICDE), 2022, pp. 1479-14, 2020 state of the art
- Deep Learning for Time Series Anomaly Detection: A SurveyZahra Zamanzadeh Darban et al., arXiv preprint, 2022 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2505.13033 https://arxiv.org/abs/2505.13033
- arxiv.org/abs/2110.02642 https://arxiv.org/abs/2110.02642
- Industri Farmasi Masuk Era AI, Kemenperin Fasilitasi Uji Coba Solusi Industri 4.0 suarapemerintah.id https://suarapemerintah.id/2026/07/industri-farmasi-masuk-era-ai-kemenperin-fasilitasi-uji-coba-solusi-industri-4-0/
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Anomaly Detection and Predictive Maintenance on Sensor Data", https://aurion.semai.id/topics/sensor-anomaly-detection-and-predictive-maintenance/, records of October 2026.