The alarm data contain abundant fault information related to almost all components of the wind turbine. Reasonable analysis and utilization of alarm data can assist wind farm maintenance personnel in quickly identifying the types of turbine faults, reducing operation and maintenance costs. This paper proposes a clustering analysis method that groups similar alarm sequences with the same fault type. Firstly, the alarm data are preprocessed, where alarm sequences are segmented, and redundant alarms are removed. Then, a domain knowledge-fused Word2vec (DK-Wrod2vec) method is introduced to transform non-numeric alarm codes into numeric vector representations. Finally, new distance metrics are incorporated into the K-means clustering algorithm to improve clustering performance. The performance of the proposed clustering method is assessed by applying it to labeled alarm sequences. The results demonstrate that the clustering performance is the best when using DK-Word2vec and the word rotator’s distance compared with other methods. Additionally, with the optimal parameter combination, the fault types of unlabeled alarm sequences are also analyzed.

Alarms generated by a wind turbine alarm system indicate the need for emergency action by operators to protect the turbine from running into risky conditions. However, it can be challenging for operators to identify the fault types that trigger alarms, particularly with few labeled fault samples. This paper proposes a novel fault diagnosis method for wind turbines with alarms that collaboratively uses labeled and unlabeled alarms to improve diagnosis accuracy. First, the proposed method distinguishes different alarm sequences using a designed Siamese convolutional neural network with an embedding layer (S-ECNN) model. Then, the fault category of an unknown alarm sequence is diagnosed based on similarity scores. Specifically, the Skip-gram model is used to mine potential relationships among alarms in unlabeled alarm sequences, and pretrained alarm vectors are obtained. In the S-ECNN model, the pretrained alarm vectors are further optimized and trained using labeled alarm sequences. The similarity scores are calculated based on the distance between the extracted discriminative features of alarm sequences. The effectiveness of the proposed method is validated using actual alarm data from a wind farm.

Tunneling magnetoresistive (TMR) sensors have broad application prospects because of their high sensitivity and small volume. However, the inherent hysteresis characteristics of TMR affect its applications in high accuracy scenarios. It is essential to build a model to describe the attributes of hysteresis of TMR accurately. Preisach model is one of the popular models to describe the behavior of inherent hysteresis for TMR, whereas it presents low accuracy in high-order hysteresis reversal curves. Furthermore, the traditional Preisach model has strict congruence constraints, and the amount of data seriously affects the accuracy. This paper proposes a hysteresis model from a probability perspective. This model has the same computational complexity as the classic Preisach model while presenting higher accuracy, especially in high-order hysteresis reversal curves. When measuring a small amount of data, the error of this method is significantly reduced compared with the classical Preisach model. Besides, the proposed model’s congruence in this paper only needs equal vertical chords.

Hysteresis is one of the main factors affecting the measurement accuracy of TMR sensors, especially in dynamic measurements. The commonly used Preisach hysteresis compensation model has some problems, such as complex modeling and difficulty in accurately measuring the step time, resulting in low accuracy in dynamic measurements. In this paper, considering the distribution characteristics of the conversion time of the hysteresis operator in dynamic measurements, a dynamic hysteresis model based on the probability estimation of the hysteresis operator and its conversion time is proposed. Compared with the existing methods, this method only needs to calculate the distribution of the sensor hysteresis operator to realize the calculation of hysteresis characteristics without a physical model or fitting algorithm. It has good generalization performance and a corresponding fast speed. In the test of two typical TMR sensors, compared with the transmission Preisach model, the maximum error of this method is reduced by 46.7%, the variance can be reduced by 90.2%, and the average value can be reduced by 65.1%.