💻 computer science

Cost-Sensitive Incremental Thermal Prediction with Adaptive Drift Detection for Predictive Maintenance on Single-Board Computers

This paper presents HT-CS, a cost-sensitive incremental learning framework combining a Hoeffding Tree regressor with ADWIN drift detection and sample-level weighting, which significantly improves critical temperature event recall and prediction accuracy on resource-constrained Raspberry Pi 4B devices while maintaining low latency and a small model footprint.

Original authors: Abd. Hallim¹, Maria Susan Anggreainy¹, Endra Oey², Widodo Budiharto

Published 2026-09-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Abd. Hallim¹, Maria Susan Anggreainy¹, Endra Oey², Widodo Budiharto

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a small computer, no larger than a credit card, working hard inside a factory or a smart home. These devices, known as single-board computers, are the quiet workhorses of the modern internet, running sensors and managing data without needing a connection to a massive cloud server. However, like any engine, they generate heat. If they get too hot, they slow down to protect themselves, much like a runner who has to stop and walk when their body temperature spikes. For these machines to keep working efficiently, they need to know when they are about to overheat before it happens. The challenge is that the conditions causing them to heat up are constantly changing, and the moments when they actually get dangerously hot are incredibly rare, happening in only a tiny fraction of their operating time.

This is the problem researchers at Binus University set out to solve. They wanted to build a system that could live directly on these small computers, learning in real-time as the device worked, to predict when the temperature would hit a critical point. The difficulty lies in the nature of the data: the computer's workload shifts from moment to moment, and the rare overheating events are so infrequent that a standard learning program would likely ignore them entirely, focusing only on the common, safe temperatures. Furthermore, the computer has very little memory, meaning it cannot store vast amounts of history or use complex, heavy-duty software to figure things out. It needs a lightweight, adaptable brain that can spot a dangerous trend immediately, even if it has only seen a few examples of that danger before.

The researchers developed a new method called HT-CS, which acts as a constantly learning observer. Instead of trying to memorize every past event, this system builds a simple decision tree, a flowchart of sorts, that updates itself with every new piece of data it receives. To handle the fact that the environment changes, the system includes a watchdog mechanism that notices when the data starts behaving differently. When this watchdog detects a shift, it resets the decision tree, clearing out old assumptions so the system can learn the new rules quickly. However, there was a catch: because overheating events are so rare, every time the system reset, it would forget the few examples of heat it had just learned, leaving it blind to the danger again.

To fix this, the researchers introduced a concept of "cost sensitivity." They taught the system that a single example of a hot temperature is far more important than a thousand examples of a cool one. By giving extra weight to those rare, critical moments, the system could hold onto the knowledge of what overheating looks like, even as it cleared away old data to make room for new patterns. They tested this approach on eight different Raspberry Pi computers, feeding them a massive stream of performance data containing over 400,000 records. The results showed that without this special weighting, the system would completely fail to predict the dangerous heat, missing nearly every instance. With the new method, it successfully identified more than half of the critical moments, all while keeping its own memory usage incredibly low—just 9.5 kilobytes, which is roughly the size of a single short text message.

The study also revealed that this balancing act is delicate. If the researchers made the system too eager to learn from the rare hot events, it would become unstable, resetting itself too often and losing its ability to make any accurate predictions at all. They found a specific setting where the system was just right: sensitive enough to catch the danger, but stable enough to keep learning. This configuration allowed the computer to predict temperature spikes with an average error of less than one degree Celsius, a level of precision that is vital for preventing damage. The system operates so fast that it can make a prediction in less than one-hundredth of a second, leaving plenty of time for the computer to take action before it gets too hot.

Ultimately, this work demonstrates that it is possible to give small, resource-limited devices the ability to learn from their own environment and protect themselves from failure. By combining a simple learning structure with a smart way of handling rare events, the researchers created a tool that can run continuously on the edge of the internet, watching for trouble without needing help from the cloud. This approach offers a practical path forward for keeping the growing network of smart machines running safely, ensuring they can adapt to changing conditions without ever needing to be reprogrammed or given more memory than they physically have.

Technical Summary: Cost-Sensitive Incremental Thermal Prediction with Adaptive Drift Detection for Predictive Maintenance on Single-Board Computers

Problem Statement
Predictive thermal management for edge IoT processors, specifically Single-Board Computers (SBCs) like the Raspberry Pi 4B, faces three converging challenges:

Non-Stationary Data Streams: Workloads, ambient conditions, and seasonal factors cause shifting data distributions (concept drift), requiring models that adapt continuously rather than relying on static batch training.
Severe Class Imbalance: Critical thermal events (temperatures $T \ge 60^\circ\text{C}$ , triggering firmware throttling) occur in only ~3% of operational records. Standard incremental learning algorithms often fail to learn these rare events, leading to "minority-class erasure."
Resource Constraints: Edge deployment requires models with minimal memory footprints (targeting <10 KB) and low latency, ruling out deep learning ensembles or large batch-processing approaches.

Existing approaches struggle with the "stability-plasticity dilemma" in streaming imbalanced data: drift detectors reset model statistics to adapt to new data, but this reset inadvertently erases the sparse statistics required to recognize rare critical events.

Methodology: The HT-CS Framework
The authors propose HT-CS, a cost-sensitive incremental learning framework designed to run on SBCs. The architecture integrates three core components:

Hoeffding Tree (HT) Regressor: A "Very Fast Decision Tree" capable of learning from data streams in a single pass. It uses the Hoeffding bound to determine split points with finite data, ensuring asymptotic equivalence to batch learning while maintaining a compact memory footprint.
ADWIN (ADaptive WINdowing) Drift Detection: A statistical change detector that monitors the stream of prediction errors. When a distribution shift is detected, ADWIN triggers a reset of the Hoeffding Tree (reducing it to a single leaf) to prevent catastrophic forgetting of new concepts.
Cost-Sensitive Learning (CSL): To counteract the erasure of minority class statistics during ADWIN resets, the framework applies a sample-level cost weight ( $c$ ). Records where the target temperature $y_t \ge 60^\circ\text{C}$ are assigned a weight $w_t = c$ , while others receive $w_t = 1.0$ . This amplifies the contribution of critical-zone records to the split statistics, ensuring the tree converges on rare events even within short inter-reset windows.

Experimental Setup

Dataset: The LwHBench benchmark, comprising 404,611 records from eight Raspberry Pi 4B units under four workloads (CPU, GPU, Storage, Memory).
Split: A strict temporal split was used: 399,611 records for training (stream) and the final 5,000 records for testing.
Evaluation: The study employed a systematic parameter sweep of the cost weight $c \in \{1, 2, 5, 10, 15, 20, 34\}$ with and without ADWIN. A $2\times2$ ablation study isolated the effects of ADWIN and CSL.
Metrics: Primary metrics included Critical-Zone Recall (detection of $T \ge 60^\circ\text{C}$ ), Mean Absolute Error (MAE), and $R^2$ . Deployment constraints were measured via inference latency and model size.

Key Results
The parameter sweep revealed three distinct behavioral zones:

Safe Zone: Low cost weights ( $c=2$ ) with ADWIN.
Degradation Zone: Moderate to high weights ( $c=5$ to $15$) with ADWIN, where MAE increases significantly.
Collapse Zone: High weights ( $c \ge 20$ ) or $c=1$ with ADWIN, resulting in 0% recall due to tree destabilization or minority erasure.

Optimal Configuration ( $c=2$ , ADWIN Enabled):

Critical-Zone Recall: 54.30% (a 23.30 percentage point improvement over the unweighted streaming baseline).
Mean Absolute Error (MAE): 0.760°C.
Model Footprint: 9.5 KB (serialized size).
Inference Latency: 0.061 ms per record on a Raspberry Pi 4B, providing a ~16,000x margin over the 1 Hz sensor sampling rate.
Precision: 49.3% (with a 12.00% false-alarm rate).

Ablation Study Findings:
The $2\times2$ factorial analysis quantified the interaction between components:

ADWIN alone ( $c=1$ ): Caused a -31.00 pp drop in recall due to the erasure of minority class statistics before they could stabilize.
CSL alone ( $c=2$ , no ADWIN): Provided a +7.92 pp improvement in recall but lacked drift adaptability.
Synergy: The combination (HT-CS) yielded a +46.38 pp interaction effect, demonstrating that CSL is essential to counteract the negative impact of ADWIN resets on rare-event learning.

Significance and Claims
The paper claims that HT-CS successfully addresses the gap in lightweight, on-device predictive maintenance for SBCs by:

Enabling Real-Time Adaptation: It achieves drift adaptation without cloud dependency or offline retraining, meeting the <10 ms latency and <10 KB footprint requirements for edge Digital Twins.
Solving Minority Erasure: It demonstrates that cost-sensitive weighting is a necessary mechanism to preserve the learnability of rare critical events in streaming environments subject to concept drift.
Defining Operational Boundaries: The study identifies specific "safe," "degradation," and "collapse" zones for cost weights, providing a practical guide for deploying incremental learners on imbalanced streams.

The authors position HT-CS not as a replacement for high-fidelity batch models, but as a viable, resource-constrained solution for continuous thermal monitoring where latency and memory are binding constraints. The work serves as a foundation for future self-referencing digital twin architectures that can autonomously adjust operating parameters within defined safety boundaries.

Technical Summary: Cost-Sensitive Incremental Thermal Prediction with Adaptive Drift Detection for Predictive Maintenance on Single-Board Computers

More like this