The increasing significance of text data in power system intelligence has highlighted the out-of-distribution(OOD)problem as a critical challenge,hindering the deployment of artificial intelligence(AI)models.In a clos...The increasing significance of text data in power system intelligence has highlighted the out-of-distribution(OOD)problem as a critical challenge,hindering the deployment of artificial intelligence(AI)models.In a closed-world setting,most AI models cannot detect and reject unexpected data,which exacerbates the harmful impact of the OOD problem.The high similarity between OOD and indistribution(IND)samples in the power system presents challenges for existing OOD detection methods in achieving effective results.This study aims to elucidate and address the OOD problem in power systems through a text classification task.First,the underlying causes of OOD sample generation are analyzed,highlighting the inherent nature of the OOD problem in the power system.Second,a novel method integrating the enhanced Mahalanobis distance with calibration strategies is introduced to improve OOD detection for text data in power system applications.Finally,the case study utilizing the actual text data from power system field operation(PSFO)is conducted,demonstrating the effectiveness of the proposed OOD detection method.Experimental results indicate that the proposed method outperformed existing methods in text OOD detection tasks within the power system,achieving a remarkable 21.03%enhancement of metric in the false positive rate at 95%true positive recall(FPR95)and a 12.97%enhancement in classi-fication accuracy for the mixed IND-OOD scenarios.展开更多
Recently,intelligent fault diagnosis based on deep learning has been extensively investigated,exhibiting state-of-the-art performance.However,the deep learning model is often not truly trusted by users due to the lack...Recently,intelligent fault diagnosis based on deep learning has been extensively investigated,exhibiting state-of-the-art performance.However,the deep learning model is often not truly trusted by users due to the lack of interpretability of“black box”,which limits its deployment in safety-critical applications.A trusted fault diagnosis system requires that the faults can be accurately diagnosed in most cases,and the human in the deci-sion-making loop can be found to deal with the abnormal situa-tion when the models fail.In this paper,we explore a simplified method for quantifying both aleatoric and epistemic uncertainty in deterministic networks,called SAEU.In SAEU,Multivariate Gaussian distribution is employed in the deep architecture to compensate for the shortcomings of complexity and applicability of Bayesian neural networks.Based on the SAEU,we propose a unified uncertainty-aware deep learning framework(UU-DLF)to realize the grand vision of trustworthy fault diagnosis.Moreover,our UU-DLF effectively embodies the idea of“humans in the loop”,which not only allows for manual intervention in abnor-mal situations of diagnostic models,but also makes correspond-ing improvements on existing models based on traceability analy-sis.Finally,two experiments conducted on the gearbox and aero-engine bevel gears are used to demonstrate the effectiveness of UU-DLF and explore the effective reasons behind.展开更多
In order to address the issue of overly conservative offline reinforcement learning(RL) methods that limit the generalization of policy in the out-of-distribution(OOD) region,this article designs a surrogate target fo...In order to address the issue of overly conservative offline reinforcement learning(RL) methods that limit the generalization of policy in the out-of-distribution(OOD) region,this article designs a surrogate target for OOD value function based on dataset distance and proposes a novel generalized Q-learning mechanism with distance regularization(GQDR).In theory,we not only prove the convergence of GQDR,but also ensure that the difference between the Q-value learned by GQDR and its true value is bounded.Furthermore,an offline generalized actor-critic method with distance regularization(OGACDR) is proposed by combining GQDR with actor-critic learning framework.Two implementations of OGACDR,OGACDR-EXP and OGACDRSQR,are introduced according to exponential(EXP) and opensquare(SQR) distance weight functions,and it has been theoretically proved that OGACDR provides a safe policy improvement.Experimental results on Gym-MuJoCo continuous control tasks show that OGACDR can not only alleviate the overestimation and overconservatism of Q-value function,but also outperform conservative offline RL baselines.展开更多
Classifying patterns of known classes and rejecting ambiguous and novel(also called as out-of-distribution(OOD))inputs are involved in open world pattern recognition.Deep neural network models usually excel in closed-...Classifying patterns of known classes and rejecting ambiguous and novel(also called as out-of-distribution(OOD))inputs are involved in open world pattern recognition.Deep neural network models usually excel in closed-set classification while perform poorly in rejecting OOD inputs.To tackle this problem,numerous methods have been designed to perform open set recognition(OSR)or OOD rejection/detection tasks.Previous methods mostly take post-training score transformation or hybrid models to ensure low scores on OOD inputs while separating known classes.In this paper,we attempt to build a unified framework for building open set classifiers for both classification and OOD rejection.We formulate the open set recognition of K-known-class as a(K+1)-class classification problem with model trained on known-class samples only.By decomposing the K-class problem into K one-versus-all(OVA)binary classification tasks and binding some parameters,we show that combining the scores of OVA classifiers can give(K+1)-class posterior probabilities,which enables classification and OOD rejection in a unified framework.To maintain the closed-set classification accuracy of the OVA trained classifier,we propose a hybrid training strategy combining OVA loss and multi-class cross-entropy loss.We implement the OVA framework and hybrid training strategy on the recently proposed convolutional prototype network and prototype classifier on vision transformer(ViT)backbone.Experiments on popular OSR and OOD detection datasets demonstrate that the proposed framework,using a single multi-class classifier,yields competitive performance in closed-set classification,OOD detection,and misclassification detection.The code is available at https://github.com/zhen-cheng121/CPN_OVA_unified.展开更多
基金supported in part by the Science and Technology Project of the State Grid East China Branch(No.520800230008).
文摘The increasing significance of text data in power system intelligence has highlighted the out-of-distribution(OOD)problem as a critical challenge,hindering the deployment of artificial intelligence(AI)models.In a closed-world setting,most AI models cannot detect and reject unexpected data,which exacerbates the harmful impact of the OOD problem.The high similarity between OOD and indistribution(IND)samples in the power system presents challenges for existing OOD detection methods in achieving effective results.This study aims to elucidate and address the OOD problem in power systems through a text classification task.First,the underlying causes of OOD sample generation are analyzed,highlighting the inherent nature of the OOD problem in the power system.Second,a novel method integrating the enhanced Mahalanobis distance with calibration strategies is introduced to improve OOD detection for text data in power system applications.Finally,the case study utilizing the actual text data from power system field operation(PSFO)is conducted,demonstrating the effectiveness of the proposed OOD detection method.Experimental results indicate that the proposed method outperformed existing methods in text OOD detection tasks within the power system,achieving a remarkable 21.03%enhancement of metric in the false positive rate at 95%true positive recall(FPR95)and a 12.97%enhancement in classi-fication accuracy for the mixed IND-OOD scenarios.
基金supported in part by the National Natural Science Foundation of China(52105116)Science Center for gas turbine project(P2022-DC-I-003-001)the Royal Society award(IEC\NSFC\223294)to Professor Asoke K.Nandi.
文摘Recently,intelligent fault diagnosis based on deep learning has been extensively investigated,exhibiting state-of-the-art performance.However,the deep learning model is often not truly trusted by users due to the lack of interpretability of“black box”,which limits its deployment in safety-critical applications.A trusted fault diagnosis system requires that the faults can be accurately diagnosed in most cases,and the human in the deci-sion-making loop can be found to deal with the abnormal situa-tion when the models fail.In this paper,we explore a simplified method for quantifying both aleatoric and epistemic uncertainty in deterministic networks,called SAEU.In SAEU,Multivariate Gaussian distribution is employed in the deep architecture to compensate for the shortcomings of complexity and applicability of Bayesian neural networks.Based on the SAEU,we propose a unified uncertainty-aware deep learning framework(UU-DLF)to realize the grand vision of trustworthy fault diagnosis.Moreover,our UU-DLF effectively embodies the idea of“humans in the loop”,which not only allows for manual intervention in abnor-mal situations of diagnostic models,but also makes correspond-ing improvements on existing models based on traceability analy-sis.Finally,two experiments conducted on the gearbox and aero-engine bevel gears are used to demonstrate the effectiveness of UU-DLF and explore the effective reasons behind.
基金supported by the National Natural Science Foundation of China(62373364,62176259)the Key Research and Development Program of Jiangsu Province(BE2022095)。
文摘In order to address the issue of overly conservative offline reinforcement learning(RL) methods that limit the generalization of policy in the out-of-distribution(OOD) region,this article designs a surrogate target for OOD value function based on dataset distance and proposes a novel generalized Q-learning mechanism with distance regularization(GQDR).In theory,we not only prove the convergence of GQDR,but also ensure that the difference between the Q-value learned by GQDR and its true value is bounded.Furthermore,an offline generalized actor-critic method with distance regularization(OGACDR) is proposed by combining GQDR with actor-critic learning framework.Two implementations of OGACDR,OGACDR-EXP and OGACDRSQR,are introduced according to exponential(EXP) and opensquare(SQR) distance weight functions,and it has been theoretically proved that OGACDR provides a safe policy improvement.Experimental results on Gym-MuJoCo continuous control tasks show that OGACDR can not only alleviate the overestimation and overconservatism of Q-value function,but also outperform conservative offline RL baselines.
基金supported by the National Key Research and Development Program,China(No.2018 AAA0100400)National Natural Science Foundation of China(Nos.U20A20223,62222609 and 62076236).
文摘Classifying patterns of known classes and rejecting ambiguous and novel(also called as out-of-distribution(OOD))inputs are involved in open world pattern recognition.Deep neural network models usually excel in closed-set classification while perform poorly in rejecting OOD inputs.To tackle this problem,numerous methods have been designed to perform open set recognition(OSR)or OOD rejection/detection tasks.Previous methods mostly take post-training score transformation or hybrid models to ensure low scores on OOD inputs while separating known classes.In this paper,we attempt to build a unified framework for building open set classifiers for both classification and OOD rejection.We formulate the open set recognition of K-known-class as a(K+1)-class classification problem with model trained on known-class samples only.By decomposing the K-class problem into K one-versus-all(OVA)binary classification tasks and binding some parameters,we show that combining the scores of OVA classifiers can give(K+1)-class posterior probabilities,which enables classification and OOD rejection in a unified framework.To maintain the closed-set classification accuracy of the OVA trained classifier,we propose a hybrid training strategy combining OVA loss and multi-class cross-entropy loss.We implement the OVA framework and hybrid training strategy on the recently proposed convolutional prototype network and prototype classifier on vision transformer(ViT)backbone.Experiments on popular OSR and OOD detection datasets demonstrate that the proposed framework,using a single multi-class classifier,yields competitive performance in closed-set classification,OOD detection,and misclassification detection.The code is available at https://github.com/zhen-cheng121/CPN_OVA_unified.
