In response to the significant variations in data distribution of industrial robot harmonic reducers under different operating conditions, the partial absence of data labels for certain conditions, and the incomplete information obtained from a single sensor, which together result in low diagnostic accuracy, a fault diagnosis method is proposed based on information fusion and subdomain adaptation for different operating conditions of harmonic reducers. Time-frequency graphs are constructed using wavelet transform on one-dimensional vibration data from source and target domains. Time-frequency information from multiple sensors is integrated using a wavelet transformbased image fusion method, and the fused image is created. To fully exploit the multi-representational features of the fused samples, an improved residual network with a multi-representation feature extraction structure is proposed. Simultaneously, in an unsupervised scenario, the multi-representation features of the fused samples from the source and target domains are subjected to subdomain adaptation, for reducing the distribution differences between subdomains of both domains. Transfer the knowledge from the label-rich source domain to the label-deficient target domain, and ultimately fault diagnosis of harmonic reducers can be achieved under different operating conditions. By establishing an experimental platform for the industrial robot harmonic reducers and conducting actual measurements, the proposed method can achieve an average accuracy of 98. 8% for all transfer tasks, and effectively enable fault diagnosis of harmonic reducers under different operating conditions in an unsupervised scenario.