To　enhance　the　diagnostic　capabilities　of　defective　epoxy　coating　structures　using　terahertz　non-destructive　testing　technology,　this　paper　proposed　a　multiple　damage　identification　algorithm　based　on　multi-domain　feature　fusion　and　machine　learning.　Three　typical　defect　types　of　coating　structure　with　three　severity　levels　were　investigated.　Firstly,　Finite-Difference　Time-Domain　(FDTD)　modelling　generated　terahertz　pulsed　imaging　(TPI)　signals　for　various　defective　structures,　incorporating　ageing-induced　property　variations.　Secondly,　time　domain,　frequency　domain　and　wavelet　packet　energy　parameters　were　extracted,　then　filtered　via　the　importance　analysis　using　an　improved　random　forest　method　to　obtain　the　critical　features　sensitive　to　the　defective　structures.　Subsequently,　the　selected　features　were　reconstructed　into　optimised　eigenvectors　and　processed　through　a　cascading　Support　Vector　Machines　(SVM)　classifier　with　Particle　Swarm　Optimisation　(PSO)　algorithm.　Defect　types　classification　and　defect　severity　assessment　were　implemented　through　the　cascading　classifier.　Compared　to　the　traditional　single-stage　multiclass　classifier,　the　proposed　feature　filtering　algorithm　significantly　enhanced　discriminative　feature　precision　and　better　captured　critical　coating　structural　characteristics.　This　optimisation　ultimately　improved　defect　classification　accuracy　and　reliability.　The　results　indicated　that　the　method　was　effective　and　could　be　recommended　for　the　actual　application.