Jürgen Herp

State transition of wind turbines based on empirical inference on model residuals

In this study, we propose a method to monitor state transitions for wind turbines based on the online inference on model residuals. Slowly developing faults in wind turbine can, when not detected and fixed on time, cause severe damage and downtime. Early state transition detection attempts to reduce the risk of sever damage and downtime by recognizing changes in the data and adapted predictive models appropriately. As fault detection studies often deal with hard thresholds, the Bayesian analysis comes with the advantage of probability measures. We propose a Bayesian approach to state transition based on hidden variables relevant for the online predictor, namely the time since the last state transition. It is of great interest to see that exact online inference can be performed at every time step, given an underlying predictive model based on a hazard function. Here the hazard function describes how likely it is to undergo a transition given the data since the last state transition. It is imperative that the hyper-parameters are known before-hand in order to perform the inference on the model. We show that Bayesian inference on state transition can be performed for assumed fixed and known hyper-parameters, and we emphasize that the selection of the hyper-parameters can be treated as a machine learning problem and trained given a data set. Comparing fixed to learned hyper-parameters points out the impact they have on the predictive performance.

Dependency in State Transitions of Wind Turbines—Inference on Model Residuals for State Abstractions

Abstracting turbine states and predicting the transition into failure states ahead of time is important in operation and maintenance of wind turbines. This study presents a method to monitor state transitions of a wind turbine based on the online inference on residuals. In a Bayesian framework, the state transitions are based on a hidden variable relevant for the predictor, namely the information of the current state. Given an underlying predictive model based on a students t-distribution for the samples and a conditional prior on the state transition, it is shown that state transitions can be abstracted from generated data. Two models are presented: 1) assuming independence and 2) assuming dependence between states. In order to select the right models, machine learning is utilized to update hyperparameters on the conditional probabilities. Comparing fixed to learned hyperparameters points out the impact machine learning concepts have on the predictive performance of the presented models. In conclusion, a study on model residuals is performed to highlight the contribution to wind turbine monitoring. The presented algorithm can consistently detect the state transition under various configurations. Comparing to heuristic interpretations of the residuals, both models can qualitatively inform about the time when a state transition occurs.

Dimensionality reduction by bayesian eigenvalue-analysis for state prediction in large sensor systems: with application in wind turbines

The potential of the theory of random matrices are presented and evaluated as a statistical tool to represent the empirical correlations in a study of multivariate time series. A new sub space state prediction framework is proposed, consisting of the combination of a Bayesian state prediction algorithm and the eigenvalues of the empirical correlation matrix. In an industrial use-case of wind turbines, remarkable agreement between the theoretical prediction (based on the assumption that the correlation matrix is random) and empirical data, concerning the density of eigenvalues associated with the time series of different sensors, are found. Finally, the proposed framework outperforms the existing Bayesian state prediction algorithm and is computationally more feasible than feeding unprocessed data.

Assessment of bowel cleansing quality in colon capsule endoscopy using machine learning: a pilot study

Background and study aims  The aim of this study was to develop a machine learning-based model to classify bowel cleansing quality and to test this model in comparison to a pixel analysis model and assessments by four colon capsule endoscopy (CCE) readers. Methods  A pixel analysis and a machine learning-based model with four cleanliness classes (unacceptable, poor, fair and good) were developed to classify CCE videos. Cleansing assessments by four CCE readers in 41 videos from a previous study were compared to the results both models yielded in this pilot study. Results  The machine learning-based model classified 47 % of the videos in agreement with the averaged classification by CCE readers, as compared to 32 % by the pixel analysis model. A difference of more than one class was detected in 12 % of the videos by the machine learning-based model and in 32 % by the pixel analysis model, as the latter tended to overestimate cleansing quality. A specific analysis of unacceptable videos found that the pixel analysis model classified almost all of them as fair or good, whereas the machine learning-based model identified five out of 11 videos in agreement with at least one CCE reader as unacceptable. Conclusions  The machine learning-based model was superior to the pixel analysis in classifying bowel cleansing quality, due to a higher sensitivity to unacceptable and poor cleansing quality. The machine learning-based model can be further improved by coming to a consensus on how to classify cleanliness of a complete CCE video, by means of an expert panel.

Texture classification from single uncalibrated images: Random matrix theory approach

We studied the problem of classifying textured-materials from their single-imaged appearance, under general viewing and illumination conditions, using the theory of random matrices. To evaluate the performance of our algorithm, two distinct databases of images were used: The CUReT database and our database of colorectal polyp images collected from patients undergoing colon capsule endoscopy for early cancer detection. During the learning stage, our classifier algorithm established the universality laws for the empirical spectral density of the largest singular value and normalized largest singular value of the image intensity matrix adapted to the eigenvalues of the information-plus-noise model. We showed that these two densities converge to the generalized extreme value (GEV-Frechet) and Gaussian G1 distribution with rate O(N1/2), respectively. To validate the algorithm, we introduced a set of unseen images to the algorithm. Misclassification rate of approximately 1%–6%, depending on the database, was obtained, which is superior to the reported values of 5%–45% in previous research studies.

Statistical meandering wake model and its application to yaw-angle optimisation of wind farms

The wake produced by a wind turbine is dynamically meandering and of rather narrow nature. Only when looking at large time averages, the wake appears to be static and rather broad, and is then well described by simple engineering models like the Jensen wake model (JWM). We generalise the latter deterministic models to a statistical meandering wake model (SMWM), where a random directional deflection is assigned to a narrow wake in such a way that on average it resembles a broad Jensen wake. In a second step, the model is further generalised to wind-farm level, where the deflections of the multiple wakes are treated as independently and identically distributed random variables. When carefully calibrated to the Nysted wind farm, the ensemble average of the statistical model produces the same wind-direction dependence of the power efficiency as obtained from the standard Jensen model. Upon using the JWM to perform a yaw-angle optimisation of wind-farm power output, we find an optimisation gain of 6.7% for the Nysted wind farm when compared to zero yaw angles and averaged over all wind directions. When applying the obtained JWM-based optimised yaw angles to the SMWM, the ensemble-averaged gain is calculated to be 7.5%. This outcome indicates the possible operational robustness of an optimised yaw control for real-life wind farms.

Semi-definite programming-based localization algorithm in networks with inhomogeneous transmission medium

We study the asymptotic properties of a semi-definite programming (SDP) based localization algorithm in a network of wireless sensors with an inhomogeneous RF transmission interface given incomplete and inaccurate pairwise distance measurements between sensor-sensor and sensor-anchors. A novel relaxed SDP approach based on a graph realization problem with noisy time-of-arrival (TOA) measurements with additive Gaussian noise and inaccurate transmission permittivity and permeability coefficients both with additive standard Gaussian noise and a varying dielectric constant is proposed. Denoting the true distance between a pair of sensors i and j by [EQUATION] and the set of known pair-wise distances between sensors-sensors and sensors-anchors by Λ, an upper bound for the expected value of the optimal objective relaxed SDP problem is obtained, showing that its asymptotic properties can potentially grow as fast as [EQUATION].