The adhesive strip of a glass window tends to age over long-term use, undermining the sealing integrity of the structure and posing safety risks. Such damage is often highly concealed, and traditional manual inspections fail to detect it in time, leading to an imbalanced distribution of healthy and faulty data. To address this issue, this study proposes a nonintrusive fault diagnosis method that integrates piezoelectric impedance technology with a Transformerbased deep learning model. To tackle the core challenges of scarce fault samples and imbalanced data distributions in practical applications, we innovatively propose a dataaugmentation model—CBAMVAE, which combines a convolutional block attention mechanism with a variational autoencoder. By learning the distribution of real fault data, the model generates synthetic samples to expand the dataset and enhance the Transformer’s generalization capability. Furthermore, to improve the quality of generated data and increase diagnostic accuracy, we incorporate the PPIR technique, combining it with CBAM-VAE to form the PPIR-CBAM-VAE collaborative optimization method. PPIR retains critical resonance peak features, removes nonpeak points, and applies linear interpolation to restore nonpeak regions, thereby enriching samples diversity while significantly enhancing dataset stability. Experimental results show that under a highly challenging healthytofault ratio of 20∶3, the PPIRCBAMVAE method achieves a diagnostic accuracy of 92.13%. When the imbalance ratio is 4∶1, the accuracy improves markedly from 92.27% (baseline) to 96.45%, greatly boosting the recognition of minorityclass faults samples. This study establishes a systematic fault diagnosis framework that integrates EMI technology, a Transformer model, and the innovative PPIRCBAMVAE data augmentation method, providing a highly sensitive and broadly applicable solution for health monitoring of building sealing systems. The adhesive strip of a glass window tends to age over long-term use, undermining the sealing integrity of the structure and posing safety risks. Such damage is often highly concealed, and traditional manual inspections fail to detect it in time, leading to an imbalanced distribution of healthy and faulty data. To address this issue, this study proposes a nonintrusive fault diagnosis method that integrates piezoelectric impedance technology with a Transformerbased deep learning model. To tackle the core challenges of scarce fault samples and imbalanced data distributions in practical applications, we innovatively propose a dataaugmentation model—CBAMVAE, which combines a convolutional block attention mechanism with a variational autoencoder. By learning the distribution of real fault data, the model generates synthetic samples to expand the dataset and enhance the Transformer’s generalization capability. Furthermore, to improve the quality of generated data and increase diagnostic accuracy, we incorporate the PPIR technique, combining it with CBAM-VAE to form the PPIR-CBAM-VAE collaborative optimization method. PPIR retains critical resonance peak features, removes nonpeak points, and applies linear interpolation to restore nonpeak regions, thereby enriching samples diversity while significantly enhancing dataset stability. Experimental results show that under a highly challenging healthytofault ratio of 20∶3, the PPIRCBAMVAE method achieves a diagnostic accuracy of 92.13%. When the imbalance ratio is 4∶1, the accuracy improves markedly from 92.27% (baseline) to 96.45%, greatly boosting the recognition of minorityclass faults samples. This study establishes a systematic fault diagnosis framework that integrates EMI technology, a Transformer model, and the innovative PPIRCBAMVAE data augmentation method, providing a highly sensitive and broadly applicable solution for health monitoring of building sealing systems.