Research On Domain Adaptive Machinery Fault Diagnosis Method Based On Broad Learning

Intelligent equipment maintenance and fault diagnosis are important technologies to guarantee the safety and reliability of mechanical equipment.In recent years,artificial in-telligence technology based on machine learning have achieved remarkable results in the field of mechanical fault diagnosis.In the classical machine learning theory,the training data from source domain and testing data from target domain need to satisfy the conditions of independent and identical distribution.However,due to the existence of data distribution discrepancies in dynamical working conditions,the independent and identical distribution conditions are difficult to meet,resulting in the phenomenon that the diagnostic performance of the model trained in the source domain degraded in the target domain.This phenomenon greatly threatens the safety and reliability of mechanical equipment.In recent years,aca-demics and industry have paid close attention to domain-adaptive mechanical fault diag-nostics based on deep transfer learning.As an extension of deep structured learning,broad learning is anticipated to further accelerate the diagnostic model’s training efficiency and improve its prediction performance.In this dissertation,based on the broad learning the-ory,the research on domain-adaptive mechanical fault diagnosis is investigated.The main contributions are as follows.Firstly,a label denoising algorithm based on broad learning is proposed for the problem of noisy labels in the domain adaptive data annotation tasks.The proposed algorithm extracts broad features through long short-term memory(LSTM)unit and broad learning network?weights domain samples through clustering strategy? calibrates prediction results through label correction layer? conducts domain adaptation through active confidence sampling and classifier fine-tuning.Eight variants are created by the combinations of weighting,calibra-tion,and adaptation.The effectiveness and superiority of the proposed algorithm have been validated in 30 cross-domain label denoising tasks from three datasets.The results demon-strate that the proposed label denoising algorithm has faster training speed,higher prediction accuracy,and can provide more reliable labels in the data annotation process.Subsequently,a sample synthesis algorithm based on broad learning is proposed for the class imbalance problem in domain-adaptive sample synthesis tasks.The proposed al-gorithm extracts broad features through short-time Fourier transform and broad learning network for model initialization? fuses the source domain and the target domain by active transfer sampling? synthesizes minority examples by random intra-class interpolation from fusion domain to balance inter-class ratio.The effectiveness and superiority of the proposed algorithm have been validated through three datasets containing 20 class-balanced and 27class-unbalanced cross-domain tasks.The results demonstrate that the proposed sample syn-thesis algorithm can synthesize enhanced samples with lower distribution discrepancy be-tween the fusion and target domains and produce higher prediction accuracy on the target domain.Then,a hyperparameter optimization algorithm based on broad learning is proposed for the problem of hyperparameter selection in domain-adaptive automated modeling tasks.The proposed algorithm consists of three components: a broad classifier,an active estimator,and a hyperparameter optimizer.In each iteration,the algorithm initializes a broad classifier through the hyperparameter optimizer and uses it to predict the target data? selects target data with reliable pseudo-labels through the active evaluator for cross-domain evaluation?updates the surrogate model through the optimizer and further optimizes the hyperparameter space of the broad classifier.The above steps are repeated until the number of iterations reaches a preset value.The effectiveness and superiority of the algorithm have been vali-dated in 14 cross-domain hyperparametric modeling tasks from three datasets.The results demonstrate that the proposed automated modeling algorithm has higher reliability in cross-domain evaluations and higher prediction accuracy in cross-domain tasks.Next,a federated optimization algorithm based on broad learning is proposed to address the data privacy problem in domain-adaptive collaborative optimization tasks.The proposed algorithm distributes a global model to different clients in the source domain through the central server for co-training? initializes the source domain of the global model through the federated averaging strategy? annotates new data in the target domain under different clients through the active transfer sampling strategy? updates the model through a dynamic asyn-chronous update strategy.The effectiveness and superiority of the proposed algorithm have been verified in 24 centralized and 24 decentralized cross-domain tasks from three datasets.The results demonstrate that the proposed federated optimization algorithm has better cross-domain adaptability and higher prediction accuracy.Finally,the key contributions of this thesis are summarized,and some future research directions are discussed.