Predictive Maintenance (PdM) gains increasing attention for its use with power networks in order to improve the safety and resilience of power transmission. PdM seeks to offer proactive actions based on estimates of the health status of components or systems. As maintenance actions are made only when there are evidences of impending failures, applying PdM strategy will greatly reduce maintenance and operation costs while preventing disruptive downtime and energy blackouts. Data-driven methods have been popular in PdM to learn to predict potential failures using previous data records. Commonly, diagnostic/prognostic models are constructed with a supervised learning approach, which entails the availability of labelled training data prior to learning. Unfortunately, data labelling by humans is extremely expensive, particularly for big data sets. This presents a crucial challenge in the development of digital solutions for PdM in many practical scenarios. The goals of the project is to circumvent the costly work of data labelling by investigating new self-supervised learning schemes that rely on input data to produce supervisory signals such that labels of training data are no longer required. Second, to enable continual learning so that prediction models can be updated in terms of data streams that are continuously generated from a power network. The feature of continual (and life-long) learning is of high merit in fostering more informed and accurate maintenance decisions by handling evolving conditions of power networks such as aging effects of electronic components.