Predictive maintenance of Wind Turbine Systems

CASE STUDY SNAPSHOT

Customer : System Integrator in renewable energy domain
Size : 51-200
Project vertical : Renewables
Challenge : To avoid unexpected downtimes of windmills and energy loss
Solution : Perform Predictive maintenance of Wind Turbine Systems using embedded ML technologies
Services Availed :  Embedded ML development, Edge Analytics, Electronics product development services
Tools and Technologies :

• MCU: ST Micro STM32L4R5
• Frameworks: TinyML, One-Class Support Vector Machine
• Programming Language: C
• Compiler: GCC

INTRODUCTION

One of our earlier acquaintances, a system integrator, is in the field of wind turbines. They deal with complete installation of the windmills, setting up the generation ecosystem, connecting to the grid and maintaining the same. As down times results in financial and reputational losses, the customer wanted to implement a predictive maintenance mechanism by which the wind turbines can be monitored continuously, anomalies detected, and preventive measures taken to avoid complete breakage.

CHALLENGE

The wind turbine had to be monitored at multiple locations for anomalies continuously. But as the systems are being installed in remote areas with intermittent connectivity to the internet, it ruled out cloud-based processing. Hence, the overall predictive maintenance operation must be done locally and only the results are shared to the central cloud server. Embien with its capabilities in Edge AI algorithms and designs, took up this challenge, proposed and successfully developed the solution using vibration analysis mechanism.

SOLUTION

Wind Turbine System

Wind turbines are composed of several key parts that work together to harness wind energy and convert it into electricity. Some of the main components of a typical horizontal-axis wind turbine are Tower, Nacelle, Rotor Blades, Hub, Main Shaft, Gearbox, Generator, Yaw System, Pitch System, Anemometer and Wind Vane and finally the Brakes and Safety Systems. These forms a complex web of electro-mechanical systems that help maximize the efficiency of the system and generate huge power to the tune of Mega watts.

Anomaly detection Techniques

There are various mechanisms to detect anomalies in a wind turbine, that are used in tandem to identify issues at different places.

Vibration Sensors:

Accelerometers, placed strategically in different parts of the turbine, monitor vibrations and changes in vibration patterns.

Temperature Sensors:

Monitor the temperature of critical components, such as the gearbox and bearings. Sudden increases or irregular temperature variations can signify potential issues.

Oil Condition Sensors:

Sensors measuring lubricating oil quality and condition can detect contaminants, degradation, or excessive wear in the lubricant, indicating potential gearbox issues.

Power and Load Sensors:

Power output and load variations in the turbine can be attributed to abnormalities in generated power or excessive loads on the components, indicating potential mechanical or electrical issues.

Wind Speed and Direction Sensors:

Sudden changes or extremes in wind conditions may affect turbine operation, and anomalies in these measurements might indicate potential turbine stress or inefficiency.

Current and Voltage Sensors:

Sensors monitoring electrical parameters, such as current and voltage in the generator and electrical systems, can detect fluctuations or irregularities that may point towards electrical issues or faults.

Ultrasonic and Infrared Sensors:

Ultrasonic sensors can detect structural defects or cracks in turbine components by sending and receiving ultrasonic waves. Infrared sensors can identify hotspots or temperature anomalies in the structure.

SCADA Systems and Data Analytics

Supervisory Control and Data Acquisition (SCADA) systems collect and analyze sensor data in real-time. Data analytics and machine learning techniques applied to the vast sensor data can help identify patterns, trends, or anomalies that are indicative of potential faults or failures.

Vibration Analysis

Our product engineering team identified vibration analysis as the detection method of choice. Vibration signals captured by sensors contain valuable data about the structural resonance and various components within the machine. Under normal operating conditions, a machine generates specific vibrations that reflect its typical behavior and characteristics. However, as degradation sets in, there is a noticeable alteration in the vibration signals, indicating changes in the machine's behavior and performance. Vibration measurement is commonly done in gearboxes, turbines, bearings, and shafts.

Vibration data based predictive maintenance offer several advantages:

Early Detection:

They can often detect mechanical issues or abnormalities in rotating machinery earlier than other methods, as changes in vibration patterns can indicate faults before they escalate.

Comprehensive Insights:

Vibration analysis provides comprehensive insights into the condition of various components (like bearings, gears, shafts) by analyzing different vibration frequencies and patterns.

Proven Reliability:

Vibration monitoring has a proven track record in predictive maintenance, making it a reliable method for many industrial applications.

System Design

As the system cost must be competitive and vibration analysis has to be done in multiple places in a single turbine, it was proposed to go for a low cost MCU based system. It was decided to go for a STM32L4R5 based system. While the low-speed rotor shaft and bearing runs at about 0.3 Hz, the high-speed shafts operate at about 53 Hz. It is established to have vibration sensors that perform in the realm of up to 10 kHz and up. Hence accelerometer with 10KHz sampling period was selected for measuring the vibrations. Connectivity to the SCADA system was provided on an RS485 interface over Modbus protocol.

Embedded ML Training

Embien’s Embedded ML development engineers worked on the algorithms to achieve the necessary functionality. Among the array of machine learning algorithms considered, the One-Class Support Vector Machine (SVM) stood out as particularly accurate in predicting anomalous conditions. It is specifically employed to forecast abnormal machine conditions. Employing an unsupervised machine learning approach, this method doesn't require labelled data for training or relies on correct patterns to train the model.

Wind Turbine Vibration Data training​

Our engineers collected a large amount of data from multiple wind turbines and used them to train the model offline in the PC. The vibration data from the motor lacked identifiable patterns when visualized in the time-domain, exhibiting non-harmonic characteristics. Due to this absence of discernible patterns, RMS (Root Mean Square) based feature extraction techniques were applied to derive meaningful data attributes. These features were then used to train the model. The output was made available in the PKL format.

Edge Analytics

Considering the resource constrained environment on the STM32 MCU, TinyML framework was employed to enable machine learning inference directly. The trained model (in PKL format) was converted to a C header and loaded onto the TinyML framework.

Wind Turbine Vibration Data analysis​

During runtime, the vibration data was continuously read from the accelerometer and RMS based feature extraction done on the target device. These features were then fed to the trained model and the result was compared with a pre-calculated threshold. The comparison result was indicative of a fault or a normal condition. This detection process was further refined and filtered to ensure a significant threshold is reached in terms of time to prevent false negative. The results were then communicated to the Supervisory Control and Data Acquisition (SCADA) system via the Modbus interface.

Apart from the predictive maintenance of Wind Turbine Systems, Embien also provided mechanisms to update the system firmware along with the machine learning model on field when necessary.

BENEFITS

Our predictive maintenance capabilities enabled customers to proactively maintain the system along with the following benefits.

• Predictive maintenance seamlessly integrated on the edge, enabling real-time data collection, analysis, and decision-making.
• Low-cost system development enabling higher Return on Investment (RoI).
• With system accuracy over 95%, downtimes reduced significantly.
• Performing Predictive maintenance on STM32 locally minimizes the amount of data that needs to be transmitted over the network, reducing bandwidth usage and associated costs.
• The results seamlessly integrated with the existing SCADA avoiding the need for additional resources.

CONCLUSION

The predictive maintenance of Wind Turbine Systems design was deployed to a few locations and studied for performance. It was a success as it could identify real problems in the field helping the customer from unintended shutdowns. They could send their technicians as and when anomalies detected and avoided major failures. Embien was once again part of that customer’s success story with our predictive analysis expertise.