Feature Engineering-Based Operational State Recognition of Rotating Machines

Such an assumption is being proved wrong by the engineers who work hard on finding solutions to convert these fossil fuel-consuming machines to also operate on renewable fuels. ICE-based power plants have a crucial role in the green transition as a balancing element for the fluctuating nature of wind and solar energy production.

Vibration analysis and machine learning methods

The future goals impose new requirements and raise uncertainties considering the whole lifecycle of the power plants. They generate a need for the development of new tools and methods in a wide range. Within the operational phase of the lifecycle, particularly in the domain of structural condition monitoring, vibration analysis techniques have long been the cornerstone of getting precise insights into the health of rotating machinery and along with the operational data estimating their remaining useful life. On the other hand, increased computing power, and the emergence of the industrial internet of things (IIoT) have created a foundation for continuous operational monitoring in real-time, or at least in near real-time. In this context, vibration analysis (VA) and machine learning (ML) methods can be used to build precise and efficient state recognition models for rotating machines as shown in this case.

Operational state recognition of a generating set

This article introduces simple and computationally light models for the operational state recognition of a generating set (genset). The models were developed in a research project (Digibuzz-VTT) forming part of a joint research effort called DigiBuzz financed by Business Finland and are thoroughly described in a master’s thesis [1]. DigiBuzz was led by LUT University between 10/2019 and 01/2022. One of the partner companies in DigiBuzz, Wärtsilä Finland Oy, provided the dataset for building the operational state recognition models. The data consists of accelerations acquired from a Wärtsilä 20V31SG genset measured at various constant power output levels, as well as during some occasional fault situations. Gensets combine an ICE and an electric generator. They are typically used for producing power to the electric grid. While the electric grids have constant frequency, the power demand fluctuates. As a result, the gensets operate at constant speeds but with variable power output. The grids may encounter occasional disturbances which cause abnormal operation of a genset. Thus, the dataset effectively covers the acceleration response of a genset within its typical operational range.

Inertia forces and gas forces

The operational state recognition models discussed in this article are built around the cyclic nature of the operation of ICEs. The general assumption is that the dynamic behaviour, at steady load and constant rotational velocity across engine cycles, repeats itself and that load variation can be seen as a notable change in the dynamic response. Thanks to Newton, most of us know that acceleration and vibration is caused by force, and think that the relation between them is linear. Considering ICEs, the principal forces exciting vibrations can be divided into inertia and gas forces. The origin of the inertia forces are the moving parts of the engine, namely the crank and piston mechanisms. Thus, at constant rotation speed the inertia forces remain periodically stationary. However, due to the virtual linearity between force and acceleration, the gas forces, provoked by the cylinder pressure, do vary in sync with load variations, even though the rotation speed remains constant, since they are responsible of making the engine run and they must adjust to the load demand. Normalized tangential forces at crank pin for different loads during one engine cycle (four-stroke) are presented in Figure 1.

Therefore, if the detection of variations in the load is of interest, it is crucial to extract only the effect of the gas forces on the vibration response. Unlike the gas forces, the inertia forces have an analytic solution which happens to be periodic. It states that the inertia forces have cyclic components only at the frequency of rotation and its second multiple, which then leads to all the other frequency components of the vibration response to depend only on the gas forces. The harmonic frequency components of a signal can be efficiently computed using fast Fourier transform (FFT). The harmonic coefficients of the torque of a four-stroke gasoline engine at full load and at idle presented in Figure 2 were determined by Porter as early as in 1943 [2]. For a four-stroke engine one engine cycle equals two rotations of the crankshaft. In Figure 2 order 1.0 equals the rotation frequency and hence order 0.5 the engine cycle frequency.

Accurate ML models are seldom trained using raw data. The training of ML models often needs features that are sensitive to changes in the quantity being predicted by the model. Considering ICEs (and all the previous explanations), a feature sensitive to power output variation is extracted from the vibration response by simply computing the harmonic coefficient at order 1.5. This feature extracted from the three signals of only one suitably placed triaxial accelerometer can be used for training an accurate classifier of different power output levels of a Wärtsilä 20V31SG genset. The accuracy of the classifier model can be increased by adding the signal power of the three signals to the features of the model.

Smoothing out cyclic variations is possible

However, the operation of an ICE in practice is never perfectly constant between engine cycles even at steady load and therefore there is always cyclic variation in the acceleration response as well. This is typical especially considering spark ignited engines, such as the Wärtsilä 20V31SG, for which the peak cylinder pressure between consecutive cycles varies significantly. Considering the presented state recognition models the effect of the cyclic variation can be smoothened by extracting the feature values from signal segments that are multiple engine cycles long. By extending the length of the signal segment the prediction given by the model gets further away from real-time. Therefore, the right balance between the accuracy and timeliness of the model must be sought depending on the application and needs. In this case the accuracy is very high even when using signal segment length of two engine cycles. At the nominal operation speed of the genset, that is at 750 rpm, one engine cycle lasts 0.16 seconds.

The confusion matrix of a classifier trained with features extracted from two engine cycles long signal segments is presented in Figure 3. Logistic regression was used as the classifier algorithm and the features were the acceleration amplitude at order 1.5 and the signal power extracted from the signals of one triaxial accelerometer. The classes are different power output levels givens as percentages of the rated power of the genset: 0 %, 50 %, 75 %, 90 %, 95 %, and 100%.

Novelty detection can recognise abnormal operation

The recognition of abnormal operation can be done using novelty detection. Novelty detection is a subtype of binary classification in which a trained model predicts if a data sample belongs to the same class of the data it was trained with or not. The same features that were used for training the classifier model can be used for training the novelty detection models as well. Separate novelty detection models can be built for each power output level. The result of two novelty detectors trained using different algorithms, One-class support vector machine (OC SVM) and local outlier factor (LOF), are presented in Figure 4.

Features extracted from continuous one-minute-long signals of one triaxial accelerometer were given as input for the novelty detectors. The novelty detector value 0 represents normal operation and value 1 abnormal operation. The result is given as a moving average taken over a window of one engine cycle and step size of one. The abnormal operation of the genset took place at around 30 seconds which is clearly detected by both novelty detectors. [1]

Further development of the recognition models

An ambitious future goal is not only the timely detection of abnormal operation but also the recognition and classification of different types of faults. The scarcity of data measured during fault situations hinders the development of such models. However, one possible solution could be the production of data through simulations of fault situations. In fact, the first steps in that direction have already been taken using finite element method simulations of a genset (Figure 5) [3]. Further development of the operational state recognition models and their deployment in industrial settings has been planned to take place soon as part of new joint development projects between Wärtsilä, VTT, and (hopefully a long list of) other interested parties.

 

Photo: Wärtsilä corporation

References
[1] Junttila, J., 2021, Operational State Recognition of a Rotating Machine Based on Measured Mechanical Vibration Data. Master’s thesis, Arcada University of Applied Sciences (2021)
[2] Porter, F.P., 1943, Harmonic Coefficients of Engine Torque Curves. In: ASME, Journal of Applied Mecchanics, 10(1): A33-A48. DOI: https://doi.org/10.1115/1.4009248
[3] Junttila, J., Sillanpää, A. Lämsä, V.S., 2022, Validation of Simulated Mechanical Vibration Data for Operational State Recognition System, 2022 IEEE 23rd International Conference on Information Reuse and Integration for Data Science (IRI), San Diego, CA, USA, 2022, pp. 138-143, doi: 10.1109/IRI54793.2022.00040.