Optimizing air handling unit operations in clean rooms using artificial intelligence

2.1. Cleanroom requirements for lambda sensor assembly

Our sensors need to be manufactured in a controlled environment. [9] The temperature should be between 23 and 25 degrees Celsius, and the humidity should be approximately 50%. It was difficult to keep it that way. When we changed the way, we worked or the number of people working the temperature and humidity changed too.

2.2. Sensor installation and data collection

To monitor and regulate the cleanroom environment, temperature and RH sensors were strategically installed and integrated into Bosch’s Deep sight monitoring system. As shown in Figure 2.

Figure 2. Temp & RH trends taken from online Bosch internal monitoring system before the model implementation Decision-making process for ai implementation Figure 3. Internal Bosch Online monitoring system showing the RH and Temperature values

Over two months we collected 40,000 environmental data points across a range of operational scenarios. That provided us with useful training and pattern-recognition materials.

2.4. Root Cause Analysis using 5Why for AHU/Assembly

The data revealed cyclic variations in temperature even though the blower was run at a constant speed, indicating that the system was not adjusting.The high humidity levels were uncomfortable for the workers and raised concerns about corrosion.[10] Overcooling during calm nights increased the likelihood of condensation. Real-time monitoring clearly indicated that the AHU configuration was not responsive enough. implementation

2.5. Methodology

We began by collecting the data, identifying operational scenarios (classes), building and evaluating the CART model and deploying the model.

A. Collection of data Figure 4. Temperature trends taken from online Bosch internal monitoring system before the model implementation

2.4. Root cause analysis

Our analysis was additionally extended by 5-Why analysis and cause and effect diagrams [11]. The root cause analysis identified three issues, irregular use of the Air Handling Unit (AHU), sometimes excessive and sometimes insufficient, variation in internal heat load due to dynamic occupancy, [12] and absence of automated feedback to control blower speed. Together,these factors resulted in unstable environmental conditions in the cleanroom. To idetify the root causes, a 5-Why analysis and a cause-and-effect diagram were created.

Figure 5. Ishikawa diagram done by the primary author as part of internal root cause identification

This curve represents the rate of temperature change and shows the real-time dynamics of heat-load. A scatter plot of temperature vs slope, revealed nine distinct patterns corresponded to various environmental circumstances.

Figure 6. Scatter diagram showing predicted class values(generated using MATLAB)

3.2. Slope-based classification for control logic

Nine classifications have been created from the data: slope ranges: (negative, neutral, and positive) and temperature zones (low, medium, and high).

3.1. Data cleaning and feature extraction

Our dataset was cleaned and relevant features were extracted. The slope represented the rate at which heat loads varied. Using a scatter diagram of temperature versus slope, We created groups corresponding to nine distinct patterns. The CART model had the following characteristics : • Temperature (°C) • percentage of relative humidity • Temperature Slope (°C/min): calculated over a period of five minutes • Former Class State (used to maintain stability across changes) • Rate of Sensor Sampling:

Each class was associated with a particular action for using the VFD to adjust the blower speed. Validation of Class Definitions: The identified patterns were grouped into nine distinct classes, each class corresponds to a specific VFD blower-speed control action. [13] These classes were analysed and validated using MATLAB plots, and input from domain experts, particularly for boundary cases in which classification was ambiguous. Proper care was taken to ensure correct labelling under such conditions. These validated classes were used to develop a rule-based AI system to generate control actions based on real-time operating conditions [14]. This helps the system respond to actual environmental variations observed inside the cleanroom....

One sample per minute from the temperature sensor One sample per minute from the relative humidity sensor Slope Calculation Formula: Slope =

Figure 7. Classification diagram classifying the heat load class of the heat load conditions over the observation period from primary author

3.3. Model workflow and architecture

The current environmental condition is determined using nine-class logic. The frequency of VFD was changed from 30 to 50 Hz.

Only three misclassifications (mainly during suddenheat-load changes). The confusion matshowed

All of this was done via an Edge Gateway connected to the Bosch network [15], and decisions were made locally without the need for a central server.

Figure 10. Identifying of leaves in a decision tree for the given training data set generated in MATLAB

Figure 8. Installation diagram for Edge AI based optimisation control of Air handling units

Figure 11. Confusion Matrix Results for Decision Tree Model Training generated model performance

3.5. Edge integration and system parameters

AI was implemented on an edge device(ECU-1251-2LAN 4COM Gateway)to manage real-time control independently of the cloud System specifications: VFD frequency range: 30–50 Hz. Figure 9. Sequence of Process Flow of the Air Handling Unit

3.4. CART-based testing and validation

Data were split 70% for training and 30% for testing. Performance summary: • Accuracy: 99.6

These specifications ensured environmentally friendly and energy efficient AHU operation that [16].

ECU-1251-2LAN 4COM Modbus/BACnet/101/104/DNP3/PLC/ Azure/AWS IoT Gateway

High-accuracy sensors were used for continuous real-time monitoring of environmental parameters [17].

Figure 12. Edge Device used in the AI model used for Heat & humidity Calculation

Table 3. Data of Parameters post-deployment of High accuracy sensors • Sensors were calibrated before deployment to ensure accurate measurements. • Data were collected at 1-minute intervals and logged into the Bosch internal monitoring system.

Validation We compared live AHU responses with model predictions across all nine classes and found that the VFD’s behaviour matched the model outputs exactly

Figure 13. Model workflow for Validation of the model implementation

3.6. Experimental setup / cleanroom specifications

Cleanroom Environment Testing was performed in a Class 1000 Bosch cleanroom, maintaining strict temp and RH to ensure product quality. Specs: 1. Room Volume-120m3 2. temp-23± 2degree C 3. RH-50± 10% 4. Changes-25 to 30 per Hour VFD control logic transformed class outputs into blower-speed adjustments, maintaining settings where they needed to be. Sensors and Instrumentation

VFD Control Logic The Variable Frequency Drive(VFD)controlling the AHU blower was linked to the AI model outputs through a class-to-frequency mapping as summarized below:

Table 4. Data of the VFD controling the Blower • The AI model categorized real-time temperature slopes and environmental measurements into these control classes. • The VFD was used to dynamically control blower speed based on the mapped frequency to minimize energy consumption while maintaining prescribed temperature and relative humidity limits. Airflow and Ventilation Balance CFD models were used to analyze airflow patterns, ensuring laminar flow in critical areas, maintaining balance between supply and exhaust, and minimizing the formation of hot and cold spots. Heat Load Estimation The internal heat load in the clean room was estimated from environmental sensor data and internal equipment. The contribution of individual pieces of equipment was calculated by converting their

power ratings into equivalent thermal loads (in watts). In addition, the rate of change of temperature (temperature slope) was used as an indicator of environmental heat variations.

Statistical validation showed that temperature deviation decreased from 2.8°C to 1.9°C and that RH variance decreased; both changes were confirmed by paired t-tests.

To evaluate the impact of the proposed system, energy consumption was monitored using chilled water BTU meters before and after AI implementation, providing a direct measurement of energy usage.

3.7. Classification and regression tree (cart)

The Classification and Regression Tree (CART) model was chosen for the AI-based cleanroom control system for a number of technical and practical reasons. The main reason is that it has a transparent and rule-based structure, which makes it easier for operators to understand the model’s decision-making process. Moreover, compared with models such as neural networks and support vector machines. it has shorter training times and requires relatively low computational resources. It also performs well on structured datasets commonly encountered in cleanroom settings. Hyperparameter tuning was performed using a grid search, considering parameters such as maximum depth, minimum samples and splitting criteria. Pruning was used to reduce overfitting, and k-fold cross-validation was used to evaluate model performance. These decisions are supported by previous studies indicating decision-tree based methods are reliable for fast and interpretable HVAC control applications.

3.8. Data preprocessing

The dataset, consisting of around 40,000 data points, was prepared as follows: • Data were collected directly from the cleanroom using sensors measuring temperature and relative humidity to ensure quality and reliability, as this work is data driven process [18]. • Missing values (less than 2%) were filled using linear interpolation. • Class imbalance was addressed using random over-sampling. • Feature values were normalizing to the interval [0,1]. • Outliers were removed using the Interquartile Range (IQR) method. • Noise reduction was performed using a moving average filter with a window size of five.

4.2. Energy consumption

The proposed AI-based approach dynamically modulates blower speed, [20] yielding approximately a 60% reduction in blower energy consumption. This enhancement is clearly underlined by an analysis comparison between the baseline system and the AI-based model [21].

Table 6. Energy Consumption Comparison (Baseline vs AI Model) Sample calculation for % reduction: %Reduction =

4.3. Operational autonomy and blower speed distribution

The AI model processed real-time environmental data, calculated slopes, mapped readings to the appropriate control class, and operated continuously without human intervention. It provides evidence that an interpretable, edge-ready AI can achieve precise climate control in advanced manufacturing cleanrooms, [22] thereby improving quality and efficiency. (Gao et al.2024).

Table 7. AHU Blower Speed Distribution (Baseline vs AI Control) The AI model concentrated the blower’s operation in the 50-60 hz range, energy consumption

Figure 15. Temp & RH trends taken from online Bosch internal monitoring system After the model implemented

4.4. Limitations and future work

Despite the observed improvements, certain limitations remain. The current model not capture the seasonal variations and external disturbances such as door openings and maintenance works [23]. The models robustness and adaptability could be increasing the training periods . In addition, system scalability can be enhanced through integration with a centralized building management system. In the future, this model might be improved to investigate heat produced by spinning shafts at bearings [24]; and heat produced by smart beams in use [25]providing an air heating system. In buildings equipped with air heating systems, indoor climate is commonly controlled based on a reference zone, which leads to unsatisfactory thermal condition in the other zones. In addition, this strategy is not energy-efficient for buildings due to overheating that can occur in some zones. This study presents control logic for a novel-designed demand-controlled ventilation and air heating system and develops a variable air volume (VAV.

Figure 16. Energy Reduction taken from online Bosch internal monitoring system after the model implementation

Figure 17. Energy savings comparison for AHU motor with & without AI model, energy data taken from Bosch internal EMS

Figure 14. Controlled Temperature & Humidity & within the Cleanroom Range Figure 19. VFD modulation trends as per the commands from AI model to preadjust according to the predicted heat load taken from Bosch internal monitoring system