Mechanical structures and civil infrastructure are systems prone to suffer damage afterlong duty times and varying environmental and operational conditions, which mightaffect their structural behaviour. Maintenance, in general terms, is evolving towardsCondition Based Maintenance and Predictive Maintenance, which requires a goodknowledge of the health status of the systems to be maintained. In the context ofmechanical and/or civil structures, several approaches have been proposed during theyears to tackle the Structural Health Monitoring issue and accurately estimate thestructure health state. Yet, it remains difficult to diagnose damages and estimate thestructural health in the presence of varying operating and environmental conditions.Particle Filters have already been proposed as a time-domain-based method in thefield of SHM, showing promising results as an estimator of hidden states. On theother hand, neural networks-based autoencoders have been used for structural damagedetection, extracting damage-related features from vibration measurements. In thisthesis work is proposed a combination of particle filters with autoencoders in orderto obtain an algorithm for structural damage identification and diagnosis, robust tochanging environmental conditions and both linear and non-linear damages. An adap-tive threshold is used to reduce human evaluation and create an automatic indicator.