This article proposes a deep learning (DL) approach based on the time-frequency transformation of audio signals, classified using a convolutional neural network (CNN). The aim is to develop an intelligent and automated measurement system to detect adhesion defects between layers of ancient architectural plaster. The method emulates the traditional conservator’s procedure based on acoustic disturbance, auscultation, detection and classification. The system uses a hardware device known in the literature as PICUS for the generation and acquisition of acoustic signals. The acquired signals are then processed through an automated pipeline, which incorporates time-frequency transformation and data analysis within a custom-designed DL architecture. In this work, the PICUS system and the acoustic data acquisition procedure are first presented, with particular emphasis on the processing system and the overall architecture of the CNN. The proposed system improved multiclass classification accuracy compared to previous works, reaching 93% accuracy. Notably, the results demonstrate significant accuracy in identifying areas requiring stabilization. This accuracy, enhanced by the benefits of time-frequency transformation, enables new analytical scenarios guided by explainability techniques.