Research On Self-validating Soft Sensor Modeling And Its Application In Wastewater Treatment Process

With the ever-increasing development, industry system operation and maintenance arebecoming more complicated with the features of coupled variables, significant nonlinearities,parameters shift and time delay, thereby resulting in more and more rigrid requirements onprocess modeling and control. However, traditional sensors are unavailable and criticalvariables can not detect efficiently in a real-time way, leading to imposibilty of efficientprocess optimization and diagnosis. Soft sensors are a good alternative in a complex processin response to lack of efficient sensors to detect critical parameters. Against the background ofwastewater treatment, this paper focuses on soft sensors modelling together with theknowledge of chemical and biological reactions, further presents self-validating soft sensorsconcept by combining data reconstruction and hardware sensor self-validating techniquewhich addresses the issues of soft sensor model on-line identification, on-line self-validation,on-line self-reconstruction and complex process modelling. In addition, this paper is the firstattempt to control chemical dosage control in sewer networks. Main results have beenobtained as follows:1. To address the problems exsting in wastewater process and inefficiency of globalmodel, two local model learning methods, i.e., the RBF neural network and the LWPRalgorithm, are presented. PCA and Jolliffe parameters are combining as data pre-treatmentmothod, not only addressing the issues existing in pure Jolliffe parameters but also makingfull use of PCA to descrease the dimentions of input data. Due to significant time delay whenusing tradictional sensors in WWTP, in particular, BOD5dectection requires5-10daysanalysis, the simulation results showed that both local models performe well with a excellentaccuracy, provding more potential to use local modeling methods to soft sensor modelling.2. To improve ability of JIT learning, JIT-PLS and JIT-RVC algorithms are proposed inthis section. Their data selection method is enhanced by robust nearest algorithm and biascompensation algorithm. By doing so, the abilities of nonlinear tracking, dynamicalprocessing and resisting noises are improved dramatically. The simulation shows that theenhanced JIT algorithems outperform conventional JIT and RPLS, it not only improves theprediction accuracy but also enhances the adaptability and robustness of soft sensors.3. To overcome the reliable problems of soft sensors, dissimilar to traditional soft sensors,we propose an integrated framework known as SEVA soft sensors. This will perform asfollows: Validate the input sensors before making a prediction. A PCA model proposed byDunia et al. is obtained to detect, identify and reconstruct a faulty input sensor. Meanwhile, this PCA model can be further utilized in the sequential data selection for JIT learning; Builda JIT-ENS prediction model for output variables on the basis of the validated inputs, ratherthan the raw input variables, thus ensuring the prediction model is well conditioned; Validatethe prediction values of the JIT-ENS model by utilizing confidence intervals. Theseconfidence intervals, obtained from ICP algorithm, can characterize the uncertainty of theprediction model and provide further useful information about prediction quality; Generateanother three types of outputs for a soft sensor: Input Sensor Status (ISS), Output sensorstatus (OSS) and Validated Measurement (VM), beyond uncertainty values (UV) andPrediction Values (PV). The usefulness of the proposed SEVA soft sensors is demonstratedthrough a case study of a wastewater treatment process.4. To optimize chemical dosage in the sewer network, ARMA-based soft sensor isproposed in this section. It is not only capable of making6hours ahead prediction, but also isfurther used to improve feedforward control system. Therefore, the chemical dosage cost issaved dramatically. The methology is implemented in both SewerX model platform andBellambi pump station, further showing the efficiency of ARMA model-based feedforwardcontrol strategy.Finally, the summaries are obtained and pay the way for further research.