Application of Signal Processing and Machine Learning Algorithm to Determine Operation Conditions of Crane Facilities

초록

This study presents the process of obtaining and analyzing sensor data for crane facilities used in shipyard assembly processes and diagnosing normal and fault conditions by applying machine learning. First, a small test equipment was built to analyze the vibration signal and fault type of the crane. The target equipment has mechanical failure characteristics of the driveshaft bearing. Therefore, vibration and temperature sensors are attached to bearings of the target equipment to collect data according to the operation conditions. The operation conditions were set to normal and abnormal conditions (Guide broken, No Grease), and the data collection rate was 1.6 kHz, which was measured for 10 minutes depending on the operation conditions. Based on the obtained data, factors in the time and frequency domain were analyzed in a statistical methods and features were extracted through correlation analysis. Correlation analysis referred to Pearson correlation coefficients, extracted 30 factors according to operating conditions from vibration data, and finally selected 15 factors through correlation analysis with correlation coefficients of 0.6 or higher. The extracted features were performed by applying PCA (Principal Component Analysis), a representative dimensional reduction method, to visualize the operation status of the target equipment and apply machine learning algorithms. Applied machine learning algorithms use representative supervised learning algorithms, k-NN(K-nearest neighbors), SVM(Support Vector Machine), Naive Bayes and MLP(Multi-Layer Perceptron) to evaluate accuracy. The accuracy(%) of the learning models was calculated at 92.38, 91.12, 90.49 and 91.26, respectively. To validate the performance of the generated learning model (algorithm), we use the ROC(Receiver Operating Characteristic) curve to calculate the AUC(Area Under the Curve). We also applied arbitrary data to machine learning models to verify accuracy