Keywords

2.1 Introduction

Wind turbines are complex systems that include mechanical, electrical, hydraulic and instrumentation sub-systems. Off-shore wind turbines include further complexities due to their distance from the shore and the cost of access and maintenance. Hence there is a need to develop Condition-Based Maintenance strategy in order to plan the required maintenance without reduction in the availability of the wind turbine. Corrective maintenance (operate to failure) is not an option as it could be costly process and given the sea and weather conditions, timely access for maintenance might not be possible. Particularly that off-shore wind turbines need specialised equipment and training. Preventive or scheduled maintenance means to schedule maintenance regardless of the conditions of the wind turbine; in most cases this will include significant additional costs and possibly unnecessary maintenance procedures and spare parts. For further generic descriptions of the three maintenance strategies, please refer to Graisa and Al-Habaibeh [1]. Therefore, The use of Condition-Based Maintenance is critical for the effective and commercially viable operation of wind turbines. When considering the key components and sub-systems of an off-shore wind turbine, see Fig. 2.1, it becomes apparent the complexity of the wind turbine and the need for condition-based maintenance to enhance productivity and reduce cost.

Fig. 2.1
A photograph and a schematic diagram. The photograph represents offshore wind turbines. The diagram of the offshore wind turbine highlights the key parts, namely the wind vane, nacelle, yaw system, low-speed shaft, rotor blade, main bearings, generator, tower of wind turbine, and others.

a Off-shore wind turbines, and b a simplified structure and key components of off-shore wind turbine

Full size image

Numerous research work has been done to present sensor fusion for off-shore wind turbines. For example [2] has presented data fusion-based damage identification for offshore wind turbines; where a damage sensitivity index is suggested that is a function of the energy ratio between the acceleration and angular velocity. However, the proposed work had some limitations in the validation process and the limited number of signals. In reference [3] a fault diagnosis method for wind turbine gearbox bearings with fusion of vibration and current signals is presented. The results show that the proposed method can learn enhanced fault-related features particularly on compound faults. Reference [4] presents the use of ASPS approach for condition monitoring of helical gears using automated selection of features and sensors. The results show that the methodology can help sensor fusion by selecting the most sensitive sensors and signal processing methods to the detection of faults. A comprehensive literature review [5] has presented the challenges and opportunities for using artificial intelligence in the offshore wind sector. The survey concludes that the identification of technological priorities for the integration in off-shore wind turbines is still needed in order to deliver a better offshore wind farm lifecycle management. Based on the above, there is still a limitation of how to develop a suitable sensor fusion system for off-shore wind turbines due to the complex nature of the signals and the variability of operational parameters.

2.2 The Wind Turbine Under Consideration

The data considered for this paper is captured from the operation of a 7 MW wind turbine with a standard hub height of about 110 m above mean sea level and rotor diameter of about 171 m. It has a minimum rotor speed of 5.9 rpm and rated average rotor speed of about 10.6 rpm. The minimum wind speed of operation is 3.5 m/s and the maximum wind speed of operation is 25 m/s; with rated average wind speed of 13 m/s. The data provided by the SCADA system for this paper includes a wide range of sensory data, with a sampling rate of 1 sample per 10 min. Figure 2.2 presents the complexity of the of the sensory data captured from the SCADA system of the off-shore wind turbine between May 2017 and October 2018. The data includes 287 alarm signals, 388 electrical signals, 105 control signals, 321 temperature signals and 143 pressure signals. This is in addition to other signals. As presented in Fig. 2.2, mainly 6 groups of signals are captured via the SCADA system as shown in Fig. 2.2. The data in Fig. 2.2 is grouped as temperature (a), control (b), alarm (c), electrical (d), pressure (e) and other signals (f). Due to the variation in the data type and values, Eq. (2.1) is used to present the data as a normalised way to compare between the signals.

$$ S = log\left| x \right| $$
(2.1)
Fig. 2.2
6 contour speckled plots, a to f. They illustrate the temperature, control, alarm, electrical, pressure, and other signals data, respectively, captured from the S C A D A system of the offshore wind turbine.

Examples of the sensory data captured from the SCADA system of the off-shore wind turbine between May 2017 and October 2018

Full size image

where \(S\) is the processed signal and \(x\) is the raw signal from the SCADA system.

Due to the complexity of the data, it is difficult to implement an artificial intelligence system without the use of a simplified methodology to detect the most suitable signals that have high sensitivity to a fault or group of faults. The ASPS approach suggested in [4] is implemented with some suitable modification. The implemented signal processing methods are: maximum, minimum, mean, standard deviation (std), Root mean square (RMS), maximum of absolute value max(abs), coefficient of variation, Skewness, Kurtosis, Crest, Clearance, RSS, Covariance, Interquartile, Range. For further details, see [6].

As presented in Fig. 2.3, the signal data is divided as ‘before’ and ‘after fault’, where after fault is a period after removing or dealing with an existing fault. A wide range of normalised features are extracted from the signals (before and after the fault) and the difference is compared to create a Δ variable which is then used to populate an association matrix to present the change in a sensory characteristic feature (Δ of a given sensor and signal processing method). In this case, the method attempts to recognise the most sensitive features to detect a fault uses a ‘black-box’ concept, which is based on the relationship between the inputs and outputs, without the need to fully analyse each signal or feature in relation to the fault.

Fig. 2.3
A schematic illustrates a signal before and after faults, followed by a sensor feature extraction of both that leads to a normalized difference in value using a color gradient scale. There is an association matrix of signal processing with respect to the sensor.

The adaptation of the ASPS approach to detect the difference in feature values due to a fault, or group of faults

Full size image

2.3 Results and Discussion

In this case study of the off-shore wind turbine, some maintenance and faults were detected between the 6th of November and the 9th of November, therefore, we have selected two periods before and after the faults to see which signals have changed before and after those faults. This paper presents the concept and methodology in detail; and future work will include its relationship to the success of the artificial intelligent CBM system. Figure 2.4 presents the association matrix for the main 5 sensors categories. Each pixel in the association matrix presents the Δ value of the sensitivity to a sensory signal and signal processing method (sensory characteristic feature) to the fault under consideration (before and after the faults). The data is normalised so that the maximum sensitivity is 1 and the minimum sensitivity is 0; with a colour map to reflect the values as suitable.

Fig. 2.4
5 association matrices of feature versus sensor. It illustrates the temperature data, control data, alarm data, electricity data, and pressure data from October 25 to November 04 versus November 16 to 26, respectively. All have cells with low to high values using a color gradient scale.

The association matrix for the main 5 groups of sensory signals

Full size image

When considering examples of the signals that have been identified to have the most sensitivity, Fig. 2.5 presents a pressure sensor which clearly shows a different behaviour or levels before and after the fault period. Similarly, Fig. 2.6 presents an example of one of temperature sensors which is found to be sensitive to the change in the conditions.

Fig. 2.5
A line graph of sensor value versus the sample number plots the pressure sensor data from October 25 to November 4 versus November 16 to 26. It has a severe fluctuating trend up to sample number 1600 approximately and follows a lesser fluctuating trend throughout.

One of the pressure sensors that was detected to have high sensitivity and its values before and after the fault generation

Full size image
Fig. 2.6
A line graph of sensor value versus the sample number plots the temperature data from October 25 to November 4 versus November 16 to 26. It has a severe fluctuating trend up to sample number 1550 approximately and follows a lesser fluctuating trend throughout.

A Temperature sensor which has been found to have high sensitivity and its values before and after the fault generation

Full size image

When the data in Fig. 2.6 is processed by the 16 signal processing methods presented above, Fig. 2.7 presents the comparison between ‘before’ and ‘after’ values and how the sensitivity of the sensory characteristic features is the calculated and normalised to crate the association matrix.

Fig. 2.7
3 bar graphs plot temperature data. A of before and after values versus feature number has a maximum after value at 13. B of difference value versus feature number has a maximum at 13. C of the normalized value and sensitivity versus feature number has a maximum value at 8 and 14.

The way the sensitivity of the sensory characteristic features are calculated and introduced to the association matrix

Full size image

2.4 Conclusion and Future Work

When an off-shore wind turbine presents over 1254 sensory signals, it is become difficult to select the most sensitive ones for a specific fault or group of faults, particularly with the complex operational conditions of wind turbines. Hence, the implementation of artificial intelligence system becomes a challenging task. To enhance the design and reliability of artificial intelligence system, the selection of the most sensitive sensory signals and signal processing methods is needed. In this paper, the authors use a modified ASPS approach [4] for the selectin of the most sensitive sensory characteristic features before and after a fault, or group of faults. The results show that the suggested methodology is able to detect the sensitive sensors to enable in the future the design and implementation of a reliable Condition-Based Maintenance strategy using artificial intelligence.