Real Time NILM Based Power Monitoring of Identical Induction Motors Representing Cutting Machines in Textile Industry

Technical Summary: Real-Time NILM Based Power Monitoring of Identical Induction Motors in Textile Industry

Problem Statement
The textile industry in Bangladesh is a major economic contributor but remains highly energy-intensive with outdated monitoring practices. Current energy management in many facilities relies on manual tracking or Intrusive Load Monitoring (ILM), which requires extensive sensor installation, complex wiring, and high maintenance costs. While ILM provides detailed data, it is difficult to scale in large factories and suffers from calibration issues in harsh industrial environments. Non-Intrusive Load Monitoring (NILM) offers a cost-effective alternative by estimating individual appliance consumption from a single aggregate measurement point. However, existing NILM research is predominantly focused on residential environments with diverse loads. There is a significant gap in applying NILM to industrial settings, specifically for distinguishing between identical motor-driven loads (such as textile cutting machines) under real-time, low-frequency monitoring conditions.

Methodology
The authors developed a hardware and software framework to address these challenges:

• Experimental Setup: The system emulates a textile cutting section using three identical 50W induction motors (representing cutting machines) and two 15W bulb loads (representing auxiliary lighting).
• Hardware Architecture:
  • Sensing: ZMPT101B (voltage) and ZMCT103C (current) sensors capture data from the main supply line and individual load lines.
  • Processing: An Arduino Mega 2560 digitizes signals and computes real-time electrical parameters.
  • Transmission: An ESP8266 Wi-Fi module transmits data to cloud platforms.
  • Visualization: Data is stored in Google Sheets and visualized via a local 20×4 LCD and a remote Blynk mobile/web application.
• Dataset Creation: Over a 15-day period, the team collected 180,631 samples with an average sampling interval of 5 seconds. The dataset includes 16 different load state combinations, capturing both aggregate and individual load data.
• Model Selection: The study employs MATNILM (Multi-Appliance-Task Non-Intrusive Load Monitoring), a state-of-the-art model utilizing Multi-Head Self-Attention (2DMA) mechanisms. Unlike previous studies that used residential datasets (REDD, UK-DALE), this model was trained and evaluated on the newly created industrial dataset.
• Training Strategy: The dataset was split chronologically (81% training, 7.6% validation, 11.4% testing) to preserve temporal load patterns. The model was fine-tuned with specific hyperparameters (batch size 32, learning rate 0.001) and utilized data augmentation to handle the limited dataset size.
• Evaluation Metrics: Performance was assessed using Mean Absolute Error (MAE), Signal Aggregate Error (SAE), F1-score, and Normalized Disaggregation Error (NDE).

Key Contributions

1. Industrial Dataset Creation: The paper introduces a new dataset derived from experimentally tested identical induction motors, offering a benchmark specifically relevant to textile industry applications rather than relying on synthetic or residential data.
2. Integrated Real-Time Framework: The authors present a complete system that combines data acquisition, model deployment, and remote monitoring (via Blynk), moving beyond offline evaluation to a functional industrial prototype.
3. Evaluation on Identical Loads: The study specifically tests the MATNILM model's performance in a challenging scenario involving identical loads, a condition where most NILM algorithms struggle.
4. Identification of Industrial Limitations: The research explicitly highlights the difficulties NILM faces in industrial contexts, particularly regarding the distinction of identical machines and the impact of low-frequency sampling.

Results and Insights

• Aggregate vs. Disaggregated Performance: The system demonstrated reasonable accuracy in estimating total aggregate energy consumption (SAE average of 8.65%). However, per-appliance disaggregation faced significant challenges.
• Impact of Identical Loads: The model performed well when only a single machine was active. However, performance degraded sharply when two or more identical motors operated simultaneously. In these scenarios, the model frequently misallocated energy between the identical units, leading to higher MAE (ranging from 6.20W to 12.68W) and lower F1-scores (0.66–0.79).
• State Detection: The model struggled to distinguish ON/OFF transitions for the identical motors due to overlapping power signatures. Conversely, it tracked the non-motor load (bulbs) more effectively, though steady-state magnitudes were still subject to error.
• Real-Time Deployment: Upon deployment, the system remained functional with remote accessibility. While F1-scores improved in the real-time test (indicating better load presence detection), MAE and SAE increased, likely due to sensor noise and data shifts not present in the controlled training environment.

Significance and Claims
The paper concludes that while NILM holds potential for industrial energy monitoring, its application to identical motor-driven loads is currently limited. The authors modestly claim that their work highlights both the practical feasibility of a real-time, low-cost monitoring framework and the specific technical barriers that remain.

The study asserts that the primary limitations preventing high-accuracy disaggregation in this context are:

1. The inability of current models to distinguish between identical power signatures.
2. The reliance on low-frequency data, which fails to capture transient signatures necessary for differentiation.
3. The constraints of small-scale datasets on model generalization.

The authors suggest that future improvements require higher-frequency data collection, larger and more diverse industrial datasets, and advanced deep learning approaches specifically designed to handle identical loads. They do not claim to have solved the problem of identical load disaggregation but rather provide a critical evaluation of current capabilities in an industrial setting.