Cross-domain fault diagnosis approach for ball screws integrating optimized double decomposition and deep transfer learning

• Innovation Strategy Inspired by TRIZ Theory: By utilizing analytical tools such as component function dissection, Ideal Final Result (IFR) exploration, and contradiction analysis, a sophisticated fault diagnosis technical framework has been established. • Parameter Optimization & Fault Feature Enhancement Method: MICOA was introduced to improve VMD . The parameters of VMD were fine - tuned by MICOA. A secondary decomposition integrating ICEEMDAN , MICOA and VMD was developed. On this basis, IMVS was created. Features from secondary decomposition were transformed into SDP images by IMVS. Fault feature distinguishability and diagnosis efficiency were enhanced. This method is consistent with the TRIZ “increasing ideality” principle. • Deep Transfer Learning Network: SE - Resformer - Transfer was developed by the researchers to address domain adaptation and limited labeled data. This model adeptly integrates a feature extractor, fault predictor, and domain discriminator to facilitate the transfer of features between source domains and target domains. The challenges posed by diverse working conditions and varying data distributions significantly impede the application of Transformer models in diagnosing faults in CNC machine tool ball screws. This study employed TRIZ theory to systematically examine these challenges and to identify critical contradictions. Thereupon, a cross-domain fault diagnosis approach for ball screws predicated on the optimized dual decomposition technique and deep transfer learning was devised. Initially, the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) is employed to preprocess the original signals, effectively mitigating noise interference. Following this, the Multi-Strategy Improved Coati Optimization Algorithm (MICOA) is utilized to optimize the Variational Mode Decomposition (VMD). Additionally, by transforming the Intrinsic Mode Functions (IMF) procured from the decomposition into a Symmetric Dot Pattern (SDP), the distinctiveness of features was further augmented. To further enhance cross-domain fault diagnosis, a deep transfer learning model supported by a domain adversarial mechanism, designated as SE-Resformer-Transfer, was conceived. This mechanism enables efficient feature transfer between domains, thus facilitating adaptation to varying working conditions even when faced with limited labeled data. Validation through the bearing-fault dataset from Jiangnan University and the ball-screw working-condition simulation dataset indicates that the accuracy of this model reaches as high as 99.47 %. Notably, even at a signal-to-noise ratio of −5 dB, the model retains an accuracy of 90.97 %. The examination of various transfer tasks further substantiates the stability and robust predictive performance of the proposed method in diagnosing faults across different working conditions.