A novel detection method combining empirical wavelet transform (EWT) with long short-term memory (LSTM) network was proposed in order to address the challenges of joint servo parameter anomaly detection in industrial robot factory inspection scenarios, particularly under the requirements of sensor-free implementation, high accuracy and real-time performance. An integrated PLC-intelligent terminal-cloud server platform was developed for real-time joint data acquisition, enabling sensorless collection, storage and transmission of joint current and speed signals while supporting cloud-based condition monitoring and anomaly detection. EWT was employed to decompose current signals into inherently smooth characteristic components, which were directly fed into the LSTM network as input features to enhance prediction accuracy. A bidirectional LSTM architecture enhanced by convolutional neural networks (CNNs) and attention mechanisms was implemented to reconstruct complete periodic signals in order to resolve incomplete signal cycles during robotic motion. Then these reconstructed signals were compared with standard signal components through difference metrics to achieve real-time anomaly detection. Experimental validation using 1 set of standard servo parameters and 24 sets of abnormal parameters demonstrated that the proposed method achieved precise fault joint localization, maintained strong consistency with injected anomaly severity levels, and reached a detection accuracy exceeding 90%.