Predictive Maintenance (PdM) has been a challenge due to insufficient labeled data and dynamic industrial environments. These problems require a data driven solution for machine degradation analysis which is capable of on-device learning and adapting to the changing industrial environments. Therefore, we propose a new algorithm TEEMSC -Trainable Energy Efficient Machine Diagnosis using Singular Values and Canonical Crosscorrelation for machine degradation analysis which can be trained completely on a sensor node and the learned parameters are adaptive to dynamic industrial scenarios. TEEMSC learns from unlabeled data in mainly unsupervised manner. In comparison to TEEMSC, the existing data driven unsupervised methods mostly do anomaly detection and the neural network based solutions either rely upon a cumbersome data acquisition and labeling process or suffer from concept drift caused due to changing environmental scenarios. Considering bearing degradation as a PdM scenario we have tested the algorithm on PRONOSTIA and XJTU bearing dataset and compared performance of TEEMSC with three existing degradation trend analysis methods namely Principal Component Analysis, Power Spectral Density and Kurtosis on the wavelet decomposition of the original vibration signal. Our results can clearly outperform existing unsupervised degradation trend analysis and anomaly detection methods for on-device learning.