1. Executive Summary: The Story

Our goal was to teach an AI model to understand the "healthy" behavior of a critical Air Handling Unit (AHU) by predicting its normal airflow based on fan speed. This is a cornerstone of predictive maintenance, as deviations from this predicted norm are often the first sign of a developing mechanical fault, such as a clogged filter or a slipping fan belt.

We followed a structured, three-phase process: deeply analyzing the data, building a simple "baseline" model to prove the concept, and finally training advanced AI models to find the optimal solution.

Key Takeaway: The project was a definitive success. We developed an advanced AI model (XGBoost) that predicts the AHU's normal airflow with over 96% accuracy. This model is significantly more accurate and reliable than the original baseline approach.

2. The Business Advantage: Baseline vs. Champion Model

Baseline Model (Polynomial Regression)

The initial baseline model served a critical purpose. It provided a fast and simple way to prove that a predictable relationship exists in the data. This validated the project's core concept and set a clear, quantitative performance benchmark that any advanced model would need to surpass.

• Value: Good for catching major operational deviations and confirming the viability of a data-driven approach.
• Limitation: Lacks the sensitivity to detect subtle, early-stage faults, leading to a higher risk of missed detections and false alarms.

Champion Model (XGBoost)

The superior performance of the XGBoost model translates directly into tangible business value. By upgrading from the baseline to this champion model, we unlock the true potential of predictive maintenance.

• Earlier Fault Detection: The XGBoost model's average error is 24.5% lower than the baseline. This allows us to set a much tighter threshold for what is considered "normal." Instead of waiting for a large, obvious 500 l/s drop in airflow, we can now confidently flag a more subtle 300 l/s deviation. This means we can detect issues like a filter beginning to clog days or even weeks earlier.
• Reduced Wasted Labor & Increased Trust: The model's reliability (RMSE) is 25.7% better, meaning it is far less likely to make large, incorrect predictions. This drastically reduces false alarms, ensuring that when an alert is triggered, it is credible. Maintenance teams can trust the system and avoid wasting valuable time investigating non-existent issues.
• Proactive vs. Reactive Maintenance: Early warnings transform maintenance from an emergency-driven, reactive process into a planned, proactive one. Repairs can be scheduled during planned downtime, which is cheaper, less disruptive to operations, and allows for better allocation of labor and parts.

3. Our Approach: A Step-by-Step Journey

Phase 1: Understanding the Data

Before building any models, we performed an Exploratory Data Analysis (EDA) to visualize the AHU's behavior. This confirmed a strong, predictable, but distinctly non-linear relationship between fan speed and airflow, validating the need for sophisticated modeling techniques.

Figure 1: Exploratory Data Analysis. The scatter plot confirmed the non-linear relationship guiding our strategy.

Phase 2: Building a Baseline Model

We established a benchmark using a Polynomial Regression model. This approach captured the general fan curve with over 93% accuracy but showed consistent errors in certain operating ranges, proving it was a solid but beatable baseline.

Baseline Model Parameters (Polynomial Coefficients)

For technical reference, the baseline model is a 3rd-degree polynomial defined by the following learned coefficients, which create the fan performance curve:

• Constant 3 (for Speed³): 1154.4
• Constant 2 (for Speed²): -62.31
• Constant 1 (for Speed): 1.0331
• Bias (Intercept): 33.574

Figure 2: Baseline Model Performance. The model captures the main trend, but the error plot showed room for improvement.

Phase 3: Developing Advanced Models & Selecting a Winner

We tested more powerful AI models—Random Forest and XGBoost. The XGBoost model was the clear winner, with its predictions clustering tightly along the ideal "perfect fit" line, demonstrating superior accuracy and reliability.

Figure 3: Final Model Performance. The "Actual vs. Predicted" plot shows near-perfect alignment, proving the model's high accuracy.

4. Final Performance Comparison

5. Conclusion & Final Recommendation

Conclusion: This project successfully developed and validated a high-performance AI model that can serve as the "brain" for a predictive maintenance system for our AHUs.

Recommendation: We strongly recommend proceeding to the next stage: pilot deployment. This involves integrating the champion XGBoost model with our live data streams to begin monitoring the AHU's health in real-time. This is the critical next step in transitioning to a more intelligent, proactive, and data-driven maintenance culture.