The project focuses on developing an advanced sensor-based sorting system for the detection and removal of hazardous and disruptive substances in recycling environments. The system integrates multiple sensor technologies, including VIS/NIR, Terahertz imaging, active thermography, and energy-dispersive X-ray scattering, combined with AI methods to improve the purity and quality of sorted materials. The goal is to overcome current limitations in detecting multi-layer packaging, contaminated objects, and black plastics, which are challenging to sort using existing technologies.

Contribution:

1.  Integrated VIS/NIR, Terahertz imaging, active thermography, and X-ray scattering sensors to enhance the detection and sorting of materials, especially black plastics and multi-layer packaging.
2.  Implemented machine learning algorithms to analyze sensor data, extract relevant features, and fuse data from multiple sensors for improved sorting accuracy.
3.  Designed and built a prototype system capable of operating in real-time sorting environments, meeting high-speed requirements (up to 3 m/sec).
4.  Developed simulation models for sensor data to validate the robustness of AI algorithms under various conditions, including noise and outliers.
5.  Created an evaluation strategy and annotated datasets for benchmarking the performance of different sorting modules, ensuring comparability and transparency in results.
6.  Implemented continuous self-optimization of AI models using feedback from the sorting process, leveraging AutoML and multi-criteria hyperparameter optimization (HPO) for efficient model tuning.

Developed a machine learning-based system to optimize electric vehicle (EV) charging schedules by integrating solar energy forecasts, weather data, and historical charging patterns. The goal was to minimize grid energy dependency while ensuring EVs are fully charged based on user needs and deadlines.

Contribution:

• Conducted exploratory data analysis (EDA) on historical EV charging data and weather forecasts to identify key patterns and features.
• Built and trained XGBoost models to predict energy consumption and charging sessions based on time-of-day, season, and weather conditions.
• Integrated solar energy forecasts to prioritize renewable energy usage during charging sessions.
• Developed an optimization algorithm using differential evolution to minimize grid energy usage while meeting EV charging deadlines.
• Created user-friendly recommendations for charging schedules, including optimal charging times and energy sources (solar vs. grid).
• Validated the model's accuracy and efficiency using metrics such as Mean Absolute Error (MAE) and R² Score for energy consumption predictions.

The project involved analyzing vibration data from a gearbox using sensors ID44 and ID58 to identify dominant frequencies, detect anomalies, and monitor changes in vibration characteristics over time. The analysis focused on understanding how vibration energy is distributed across different frequencies and how it changes with varying power levels. The goal was to ensure the system's reliability, identify potential issues, and recommend preventive maintenance strategies.

1.  Conducted a detailed analysis of amplitude differences and peak frequency alignment between sensors ID44 and ID58, identifying stronger vibrations at the location corresponding to ID58.
2.  Utilized logbook data to monitor changes in vibration characteristics over time, correlating vibration data with operational conditions such as speed, torque, and power levels.
3.  Performed spectrogram analysis using PSD (Welch's method) and FFT-based methods to study vibration patterns across different power transition phases, identifying dominant low-frequency vibrations and their intensity changes.
4.  Applied FFT and PSD techniques to detect anomalies and deviations in vibration data, ensuring system reliability and identifying potential issues.
5.  Provided recommendations for continuous monitoring and preventive maintenance based on the analysis of vibration data and identified anomalies.
6.  Explored the potential for automating the vibration analysis process for regular monitoring and early detection of issues.

This project involved integrating Tattile STARK cameras with Qognify VMS using the QAI Plugin. The focus was on configuring the cameras to send LPR event data to QVMS, ensuring secure and efficient data transfer, and enabling real-time event handling and alarm triggering. The project also included setting up Basic Authentication and HTTPS for secure communication.

Contribution:

• Configured Tattile STARK cameras to send LPR event data to QVMS using the QAI Plugin, including setting up hostnames, output actions, and event configurations.
• Implemented Basic Authentication and HTTPS to secure the data transfer between the cameras and QVMS.
• Developed and tested the plugin to ensure it could handle LPR events, including license plate recognition, vehicle type, and additional metadata.
• Created detailed documentation and test cases to validate the plugin's functionality, including event handling, error logging, and secure communication.
• Collaborated with the team to deliver the integration on schedule, ensuring it met all functional and security requirements.
• Provided training and support to end-users, ensuring they could effectively manage and configure the STARK camera integration.

The Schneeberger Energy Forecasting System was developed to predict and analyze energy consumption patterns across multiple machines in a manufacturing facility. The goal was to identify and mitigate energy consumption peaks by integrating machine data, weather data, and advanced forecasting models. The system provides actionable insights to optimize energy usage, reduce costs, and improve energy management.

Contribution:

• Conducted exploratory data analysis (EDA) on energy consumption data from the ECON system and weather data from the DWD (German Weather Service) to identify key patterns and features.
• Developed and implemented an XGBoost-based machine learning model to forecast energy consumption, achieving an R-squared value of 0.893 and an MAE of 15.856.
• Integrated time-based features (hour, day of the week, month) and statistical features (rolling mean, lagged values) into the model to improve prediction accuracy.
• Designed and implemented a user interface using Grafana to visualize energy consumption forecasts, historical data, and weather-related insights.
• Created an alerting system to notify stakeholders of potential energy peaks via email, with configurable thresholds.
• Conducted test cases to ensure data integrity and system resilience, including handling invalid data and server access failures.

The project aimed to develop a real-time classification model for forklift movements and wind turbine gearbox fault detection using sensor data. The focus was on creating a memory-efficient solution that could operate on microcontrollers with limited resources. The project involved data collection, pre-processing, feature engineering, and the implementation of machine learning models, including the Hoeffding Tree algorithm, for real-time classification and fault detection.

Contribution:

1. : Collected and pre-processed sensor data from forklifts and wind turbine gearboxes, including removing redundant features and calculating statistical metrics.
2. : Extracted meaningful features such as mean, standard deviation, kurtosis, and skewness from sensor data to improve classification accuracy.
3. : Implemented and evaluated baseline machine learning models (Random Forest, Logistic Regression, Decision Tree, etc.) for forklift movement classification.
4. : Developed and optimized the Hoeffding Tree algorithm for real-time classification, ensuring compatibility with memory-constrained microcontrollers.
5. : Adapted the Hoeffding Tree model for embedded systems using C, focusing on memory and computational efficiency.
6. : Conducted prequential evaluation to assess model performance, achieving an accuracy of 92.78% for forklift movement classification and high accuracy for gearbox fault detection.