Deep Learning-based Industrial Soft-sensing Method Driven By Hybrid Data And Knowledge

The strict requirements of modern industrial manufacturing for process control quality and product quality promote the continuous development of soft-sensing methods and technologies.As an effective means to solve the problems of online measurement of difficult-to-measure variables in complex industrial processes,the real-time detection of product defects,and other problems,the deep learning-based soft sensor has become one of the research focuses in the field of soft sensing with its strong data feature extraction ability.However,in the context of complex industrial big data,data complexity characteristics such as irregular sampling,highdimensional redundancy,spatio-temporal sequence characteristic,noise mixing characteristic,imbalance characteristic,and other complex characteristics significantly affect the measurement performance of deep learning-based soft sensors.The data-driven attribute of deep learningbased soft sensors causes reliability problems such as poor interpretability and vulnerability to adversarial attacks.Based on these problems,this paper uses deep learning technology as the core foundation.It introduces process knowledge as the auxiliary to study the hybrid knowledge-and data-driven soft-sensing method and address four problems:(1)data imbalance characteristics affect soft-sensing performance,(2)low-quality dynamic data affect soft-sensing performance,(3)the deep learning’s black-box characteristics make it challenging to explain the soft sensor,and(4)data vulnerability to adversarial attacks results in an unreliable soft sensor model.The research is described as follows.1.First,the hybrid-driven soft-sensing method for imbalanced data is studied.With the imbalanced X-ray image of weld defects in oil pipes as the research object,a contrast enhancement conditional generative adversarial network(CECGAN)deep-learning model is proposed to solve the problem of uneven contrast of the X-ray image and resample the imbalanced data globally.Then,data knowledge on weld defects is introduced to evaluate the generation quality of CECGAN to ensure the effect of data enhancement.This knowledgecombined hybrid-driven soft sensor data preprocessing scheme ensures that the soft sensing of defects is conducted based on obtaining high-quality,balanced datasets.Furthermore,a transfer learning method based on the Xception deep-learning model containing ImageNet’s data knowledge is proposed to help train a high-precision soft-sensor target network for defect detection and recognition.Compared with other soft-sensing methods,the experimental results demonstrate that the hybrid-driven soft-sensing method ensures the rationality and diversity of the generated samples in the data preprocessing stage and achieves the highest accuracy in the defects’ soft-sensing stage:an F1-SCORE of 0.909 and defect-recognition accuracy of 92.5%.2.Second,the soft-sensing method for low-quality dynamic data is studied.Given that the noise,redundancy,and dynamic characteristics of complex industrial process data significantly affect the accuracy of soft sensing,a main low-pass filter based on complementary ensemble empirical mode decomposition(CEEMD)results is proposed,and an auxiliary filter based on the energy value of the decomposition results is designed based on empirical knowledge.The two filters are combined to reduce noise.Based on the noise reduction results,the proposed two-level isometric mapping(TLIsomap)method is used for nonlinear feature transformation to remove pseudo-component and cross-correlation redundancy.After the noise and redundancy reduction of low-quality data are completed,the data are serialized.A deep gated recurrent unit(DGRU)network is established using sequence data to conduct semi-supervised dynamic modeling to extract the complex dynamic characteristics of the data.The proposed soft sensor is verified in the industrial case of the sulfur recovery unit.The experimental results demonstrate that CEEMD-TLIsomap-DGRU can overcome the low-quality dynamic problem of data and improve soft-sensor prediction accuracy.3.Third,the interpretable soft-sensing method driven by hybrid data and knowledge is studied.Given that the deep learning-based soft sensor developed by complex industrial data lacks interpretability and results in unreliable sensing results,two hybrid-driven interpretable soft sensor methods are studied and proposed.Their effectiveness is verified in the industrial case of thermal deformation sensing of the air preheater rotor in power plant boilers.An interpretable hybrid-driven soft-sensing method based on the generative adversarial network(GAN)framework is proposed,which explores the feasibility of building a hybrid-driven model based on the GAN structure and achieves satisfactory sensing accuracy and post hoc interpretability.Accordingly,this paper analyzes the interpretation requirements of the soft sensor and proposes a deep multiple attention soft sensor(DMASS)sequential deep-neural network.The structure of this network is composed only of various attention mechanisms.These attention mechanisms ensure the self-interpretability of data selection and sensor modeling during soft sensor development.Data selection is achieved using variable-attention and time-lag-attention mechanisms constructed by introducing mechanism knowledge.The obtained attention weight produces self-interpretable data selection results.The mathematical expression for the extracted feature,the self-attention activation structures’attention matrix,the information path diagram for DMASS’s training,and the uncertainty-aware interval prediction demonstrate the self-interpretability of sensor modeling.4.Fourth,the soft sensor’s adversarial attack and defense for vulnerable data are studied.Given that complex industrial data are easily attacked,producing unreliable results for the deep learning-based soft sensor,this paper studies the adversarial attack and defense of those sensors.An adversarial attack framework for deep learning-based soft sensors is established from three aspects of information and timing and optimization objectives.The basic requirements(rationality,imperceptibility,and stability)for the practicability of such adversarial attacks are proposed.Accordingly,a knowledge-guided adversarial attack(KGAA)method is proposed to conduct black-box attacks on proxy deep learning-based soft sensors.This method adds mechanism knowledge to the objective function and adds new constraints.By defining obstacle functions,the optimization problem is reconstructed to solve the optimization’s ill-posed problem when attacking the original regression model.Furthermore,based on KGAA,the corresponding defense methods of adversarial training are proposed.Finally,the attack and defense methods were verified for thermal deformation sensing of the air preheater rotor.Compared with other adversarial attack methods,the KGAA is more practicable.The adversarial training of KGAA can be implemented to enhance the adversarial robustness of the deep learning-based soft sensor and ensure the safe,stable operation of the air preheater.