محمد رحمتي انداني

The proper an‎d timely maintenance of railway turnouts an‎d crossings is a key factor in ensuring the safety an‎d stability of the railway transportation network. In this study, machine learning techniques were employed to eva‎luate the geometric condition of railway turnouts an‎d to determine their health status an‎d level of deterioration. To this end, the required database was first developed through systematic field data collection an‎d preprocessing from a series of railway turnouts. In this regard, 20 parameters representing key structural an‎d geometric defects of the turnouts in different parts of the switch blade an‎d crossing were identified in accordance with recognized stan‎dards, an‎d periodic data collection for turnout maintenance was carried out accordingly. During the field measurements, a failure database for 11 turnouts along the Yazd railway line was compiled over 30 time periods. Subsequently, the defect parameters were labeled at different levels according to the permissible tolerances defined by international stan‎dards. After data preparation, the geometric maintenance condition of these turnouts was assessed using various machine learning algorithms, including clustering an‎d classification methods. In the first stage, to cluster the turnouts, the Hierarchical Agglomerative an‎d K-Means++ algorithms were employed. Accordingly, using unsupervised eva‎luation criteria, the optimal number of clusters was determined, an‎d the turnouts were classified into four distinct clusters. The clustering results from the Hierarchical Agglomerative an‎d K-Means++ algorithms categorized the turnouts into four maintenance-condition limits: Safety, al‎e‎rt, Intervention, an‎d Immediate Action. According to these results, the clustering an‎d labeling of the turnouts demonstrate a strong consistency between the obtained outcomes an‎d the qualitative maintenance condition levels defined by the EN stan‎dard. After labeling the turnouts, the next stage focused on analyzing an‎d eva‎luating the performance of various machine learning classification methods in identifying the maintenance condition of the turnouts. In this regard, the data were classified an‎d eva‎luated using four machine learning algorithms: Decision Tree (DT), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), an‎d Logistic Regression (LR), under two approaches: direct four-class classification an‎d hybrid hierarchical four-class classification. The results indicated that, under the hybrid hierarchical approach, the algorithms achieved an average improvement of 10-12 percent in classification accuracy compared to the direct four-class approach. This improvement resulted from the gradual decomposition of the problem into two decision-making levels (a binary classification of healthy/faulty an‎d a three-class classification of fault types), which reduced class overlap an‎d enhanced detection accuracy. This approach applies the Divide-an‎d-Conquer (D&C) concept in machine learning. The eva‎luation results revealed that the proposed hybrid hierarchical DT-KNN algorithm achieved a prediction accuracy of 82%, outperforming the best-performing algorithm in the direct classification approach (Decision Tree, with an accuracy of 66%), representing a notable 16% improvement an‎d identifying it as the most effective algorithm in this study. Overall, the findings indicate that by leveraging the high potential of both supervised an‎d unsupervised machine learning models, an efficient framework can be developed for intelligent monitoring, early fault detection, an‎d preventive maintenance planning for railway turnouts.

نگهداشت سوزن¬ها , يادگيري ماشين , الگوريتم¬هاي خوشه¬بندي و دسته¬بندي , اصل تقسيم و غلبه

Turnouts Maintenance , Machine Learning , Clustering an‎d Classification Algorithms , Divide an‎d Conquer Principle

Mohammad Rahmati Andani

Dr.Jabbar Ali Zakeri - Dr. Hamidreza Heydari