Triplet Feature Fusion for Equipment Anomaly Prediction : An Open-Source Methodology Using Small Foundation Models

Takato Yasuno

cs.LGFeb 16, 2026

Triplet Feature Fusion for Equipment Anomaly Prediction : An Open-Source Methodology Using Small Foundation Models

Authors: Takato Yasuno

Abstract

Predicting equipment anomalies before they escalate into failures is a critical challenge in industrial facility management. Existing approaches rely either on hand-crafted threshold rules, which lack generalizability, or on large neural models that are impractical for on-site, air-gapped deployments. We present an industrial methodology that resolves this tension by combining open-source small foundation models into a unified 1,116-dimensional Triplet Feature Fusion pipeline. This pipeline integrates: (1) statistical features (x in $R^{28}$ ) derived from 90-day sensor histories, (2) time-series embeddings (y in $R^{64}$ ) from a LoRA-adapted IBM Granite TinyTimeMixer (TTM, 133K parameters), and (3) multilingual text embeddings (z in $R^{1024}$ ) extracted from Japanese equipment master records via multilingual-e5-large. The concatenated triplet h = [x; y; z] is processed by a LightGBM classifier (< 3 MB) trained to predict anomalies at 30-, 60-, and 90-day horizons. All components use permissive open-source licenses (Apache 2.0 / MIT). The inference-time pipeline runs entirely on CPU in under 2 ms, enabling edge deployment on co-located hardware without cloud dependency. On a dataset of 64 HVAC units comprising 67,045 samples, the triplet model achieves Precision = 0.992, F1 = 0.958, and ROC-AUC = 0.998 at the 30-day horizon. Crucially, it reduces the False Positive Rate from 0.6 percent (baseline) to 0.1 percent - an 83 percent reduction attributable to equipment-type conditioning via text embedding z. Cluster analysis reveals that the embeddings align time-series signatures with distinct fault archetypes, explaining how compact multilingual representations improve discrimination without explicit categorical encoding.

Figures & tables

Panel of Figure 1 — Figure 1: Triplet Feature Fusion architecture. Statistical features $\mathbf{x}$ (orange), TTM embeddings $\mathbf{y}$ (blue), and text embeddings $\mathbf{z}$ (green) are independently computed and concatenated into a 1,116-dimensional vector $\mathbf{h}$ fed to LightGBM. The text embedding is shared across all time windows for the same (equipment_id, check_item_id) pair and cached as .npz .

Model	Hor.	Prec	Rec	F1	AUC	FPR
v2-0	30d	.910	.940	.920	.995	0.6%
	60d	.930	.940	.930	.995	0.5%
	90d	.950	.880	.910	.995	1.1%
v3-0	30d	.992	.926	.958	.998	0.1%
	60d	.970	.963	.967	.999	0.3%
	90d	.962	.940	.951	.998	0.4%

Table 1: Performance comparison: v2-0 (Hybrid

\mathbf{x}

\mathbf{y}

) vs. v3-0 (Triplet Fusion

\mathbf{x}

\mathbf{y}

\mathbf{z}

). FPR = False Positive Rate. Bold = improvement over v2-0.

Figure 2: Confusion matrices (top) and ROC curves (bottom) at the 30-, 60-, and 90-day horizons on the test set ( $n=8{,}745$ ). All three horizons achieve ROC-AUC $\geq 0.998$ with FPR $\leq 0.4\%$ , confirming near-perfect separation between normal and anomalous samples. At 30 days, only 6 false alarms occur among 7,961 normal samples (FPR = 0.1%).

Figure 3: Multi-modal Fusion Feature importance Top-50 by LightGBM gain (30d / 60d / 90d). Orange: statistical features ( $\mathbf{x}$ ); Blue: TTM embeddings ( $\mathbf{y}$ ); Green: text embeddings ( $\mathbf{z}$ ). The three modalities occupy distinct rank ranges, confirming mutually complementary roles.

Version	Feature set	Prec (best)	Key lesson
v1.0	Granite TS only (5 equip)	71% (90d)	Foundation model alone is insufficient for small $N$ ; equipment-identity context is missing
LGB-Stat	Stats only (28-dim)	79–87%	Hand-crafted explicit statistics provide a strong discriminative baseline
v2-0	Stats + TTM (92-dim)	91–95%	TTM embeddings capture latent temporal structure; hybrid beats any single modality
v3-0	Stats + TTM + Text (1116-dim)	96–99%	Text embeddings encode equipment-type context, suppressing FPR via type-specific decision boundaries

Table 2: Model version history and design lessons. Each row represents one generation of the HVAC anomaly detector, tracing the progression from a Transformer-only prototype to the full Triplet Feature Fusion system.

Figure 4: UMAP (left) and t-SNE (right) projections of 278 unique $(e,c)$ text embeddings. Colours indicate DBSCAN clusters ( $\varepsilon=0.24$ , min_samples =3, cosine); 20 clusters found + 26 noise points ( $\times$ ). Marker shapes: $\blacksquare$ Air Conditioning. Stars ( $\bigstar$ ) mark pseudo-centroids. The well-separated geometry confirms that free-form text labels encode genuine semantic structure—equipment category and check-item type—providing rich background context for anomaly discrimination.

Figure 5: Cluster-representative 90-day sensor series (Clusters 0–4). Red background: anomaly window ( any_anomaly =1); blue background: normal window. y-axis: z-scored value. C0 exhibits rapid oscillation with a sustained positive excursion after day 40 (heat-exchanger fouling).

Cluster	$n$	Sensor Measurement Type	Anom.	Time series pattern signature
C0	20	Chilled/Cooling Water Temp	$\checkmark$	Rapid oscillation + sustained positive excursion after day 40 (heat-exchanger fouling)
C1	5	Mixed HVAC		Near-zero, low-variance; stable setpoint with no significant drift
C2	5	Mixed HVAC		Slow sinusoidal oscillation ( $\lesssim$ 0.5 $\sigma$ ); seasonal load variation
C3	5	Mixed HVAC		Step-like plateaus; multi-stage setpoint controller cycling between fixed modes
C4	7	Main Motor Current [A]		Flat baseline with minor periodic spikes; motor within healthy $\pm 1\sigma$ band

Table 3: Cluster signatures (Clusters 0–4).

n

: unique

(e,c)

pairs.

\checkmark

: anomaly window present.

Figure 6: Cluster-representative series (Clusters 5–9). C6 shows a gradual monotonic increase across the 90-day window (progressive degradation). C7 exhibits two prominent spike bursts, indicating intermittent mechanical stress events.

Cluster	$n$	Sensor Measurement Type	Anom.	Time series pattern signature
C5	3	Mixed HVAC		Sparse intermittent readings at two discrete values; binary/ON-OFF sensor
C6	5	Mixed HVAC	$\checkmark$	Progressive upward drift across window followed by sharp spike (gradual degradation)
C7	9	Mixed HVAC	$\checkmark$	High-frequency noise on rising mean; classic early-warning pattern for mechanical wear
C8	3	Differential Pressure [Pa]		Stable, slightly negative plateau ( $\approx-0.5\sigma$ ); duct pressure well-controlled
C9	6	Operating Current [A]		Smooth near-constant trace; steady motor load over 90 days

Table 4: Cluster signatures (Clusters 5–9).

Figure 7: Cluster-representative series (Clusters 10–14). C12 ( $n=85$ , Ammeter/Diff. Pressure) shows near-flat behaviour, confirming class-level homogeneity of the largest cluster.

Cluster	$n$	Sensor Measurement Type	Anom.	Time series pattern signature
C10	4	Refrigerant/System Pressure	$\checkmark$	Step-down at day 20 + sustained low-pressure plateau (refrigerant leak / valve failure)
C11	15	Flow Rate + Temperature	$\checkmark$	Gradually increasing negative deviation; cooling flow falls below design spec
C12	85	Ammeter / Diff. Pressure		Near-flat trace over 90 days; largest and most homogeneous cluster
C13	28	Temperature [°C] (HVAC)		Gentle sinusoidal pattern ( $\approx$ 0.3 $\sigma$ ); routine diurnal/weekly cycling
C14	14	Temperature [°C] (sub-system)		Flatter than C13 with brief spikes; tighter sub-system control

Table 5: Cluster signatures (Clusters 10–14).

Figure 8: Cluster-representative series (Clusters 15–19). C15 (Inverter Frequency) exhibits a step-function waveform reflecting discrete variable-speed stages, invisible without equipment-type context.

Cluster	$n$	Sensor Measurement Type	Anom.	Time series pattern signature
C15	10	Inverter Frequency [Hz]		Step-function waveform; discrete VAV speed stages
C16	5	Temperature [°C]	$\checkmark$	Sustained excursion above $+1\sigma$ in second half (slow-onset insulation degradation)
C17	15	Flow Rate + Temperature		Irregular bounded oscillation ( $\pm$ 0.8 $\sigma$ ); normal demand variability
C18	4	Diff. Pressure / Ammeter		Low-amplitude flat signal; stable normal behaviour shared by two sensor types
C19	4	Mixed HVAC		Near-zero flat with single brief negative excursion; likely sensor dropout

Table 6: Cluster signatures (Clusters 15–19).

Component	Model / Library	License	Role
TS encoder	IBM Granite TTM [ 7 ]	Apache 2.0	$\mathbf{y}$
LoRA adapter	PEFT (HuggingFace)	Apache 2.0	Fine-tuning
Text encoder	multilingual-e5-large [ 16 ]	MIT	$\mathbf{z}$
Classifier	LightGBM [ 17 ]	MIT	Prediction

Table 7: Full open-source stack. All components are freely usable in commercial settings.

Anomaly Detection

Anomaly detection finds rare or unexpected patterns in data, from defective products and system faults to fraud and security threats.

See all 652 papers

FactoryNet: A Large-Scale Dataset toward Industrial Time-Series Foundation Models
May 9, 2026Jonas Petersen, Pratik Jawahar, Riccardo Maggioni +11Manufacturing Industrial Anomaly Detection
Forgis · ETH Zurich · Cairo University +3
Cross-Machine Anomaly Detection Leveraging Pre-trained Time-series Model
Apr 7, 2026Yangmeng Li, Kei Sano, Toshihiro Kitao +4Anomaly Detection Industrial Anomaly Detection
Program of Operations Research and Industrial Engineering, The University of Texas at Austin, 204 E. Dean Keeton Street, Austin, TX 78712-1139, USA. · Equipment Intelligence & App R&D Department, Tokyo Electron Ltd., Daido Seimei Sapporo Building, 1-3 Kita 3-jo Nishi, Chuo-ku, Sapporo, Hokkaido 060-0003, Japan. · Walker Department of Mechanical Engineering, The University of Texas at Austin, 204 E. Dean Keeton Street, Austin, TX 78712-1139, USA.
Do Time-Series Foundation Models Pay Off for Industrial Monitoring? A Cost-Aware Empirical Study
Aug 24, 2026Guan-Hua Wen, Kuan-Yu ChenTime-Series Foundation Models Time-Series Anomaly Detection
Department of Computer Science and Information Engineering National Taiwan University of Science and Technology Taipei, Taiwan

May 9, 2026J/K move · Enter open · S save

FactoryNet: A Large-Scale Dataset toward Industrial Time-Series Foundation Models

Jonas Petersen, Pratik Jawahar, Riccardo Maggioni, Alexandre Huou+10

Forgis · ETH Zurich · Cairo University+3

Manufacturing Industrial Anomaly Detection Cross-Domain Transfer Learning Time-Series Anomaly Detection Time-Series Foundation Models Industrial Automation

Open paper AskPDF arXiv

Industrial machines are actuated systems, yet most open time-series corpora record sensor outcomes alone; the few that do log commanded signals expose them in vendor-specific forms that differ from machine to machine, so commanded intent and measured response cannot be compared across embodiments. We introduce FactoryNet, a pretraining corpus for industrial time-series: 113M datapoints across 57k end-to-end task executions (15k real, 42k synthetic) on seven embodiments and four tasks, with 27 annotated anomaly types alongside healthy baselines and counterfactual pairs. Every signal in every source is mapped into a single control-theoretic schema, Setpoint, Effort, Feedback, Context (S-E-F-C), which expresses any actuated system in a common representational frame and makes commanded-versus-realized dynamics readable as explicit prediction residuals. Using the corpus, we find that a task-phase model trained on simulation, with no real phase labels, recovers roughly two thirds of the real-data ceiling on both robots, and that transfer holds within a robot family but rarely across manufacturers. Fault signatures behave differently: an uncalibrated simulator gives no detectable fault signal on either robot, and aligning the feature distributions does not repair it. What separation there is comes from Effort: motor current distinguishes faults where measured tool motion is close to chance, and calibrating the simulator to reproduce that one channel recovers a payload fault on the real robot. We read this as simulation capturing the stable structure of a task while fault behaviour stays diverse and fault-type specific. We release FactoryNet as a growing, multi-embodiment substrate for industrial foundation models.

Abstract

Anomaly Detection

Explore similar work

FactoryNet: A Large-Scale Dataset toward Industrial Time-Series Foundation Models

Cross-Machine Anomaly Detection Leveraging Pre-trained Time-series Model

Do Time-Series Foundation Models Pay Off for Industrial Monitoring? A Cost-Aware Empirical Study