Data-driven soft-sensors for monitoring and fault diagnosis in wastewater treatment plants
- Autores: Pezhman Kazemi
- Directores de la Tesis: Jaume Giralt Marce (dir. tes.), Jean Philippe Andre Steyer (codir. tes.)
- Lectura: En la Universitat Rovira i Virgili ( España ) en 2020
- Idioma: español
- Tribunal Calificador de la Tesis: Christophe Bengoa (presid.), Lluís Corominas Tabares (secret.), Nicolas Bernet (voc.)
- Programa de doctorado: Programa de Doctorado en Nanociencia, Materiales e Ingeniería Química por la Universidad Rovira i Virgili
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
Failing to reach the specific effluent properties in wastewater treatment plants can adversely affect human health and environmental. Due to this, there are significant pressures on authorities for efficient design and operation of wastewater treatment plants (WWTPs). Therefore, to achieve regulatory standards for wastewater effluent in a cost-efficient way, the development of an advanced information framework for the control and supervision of the WWTPs is mandatory. For the implementation of this framework, the real-time measurements of crucial parameters (e.g., concentrations of nitrate and total nitrogen, phosphate and total phosphorus, suspended solids, biochemical oxygen demand (BOD) and chemical oxygen demand (COD), total volatile fatty acids (VFA)) are necessary. Measurement of such parameters is often associated with capital and maintenance costs, as well as the time delay. The focus of this thesis was to design soft-sensors that can be used besides conventional instrumentation to improve the process operation and safety. Due to the availability of the massive amount of process data in most modern WWTPs, data-driven methods have attracted significant attention. Therefore, in this thesis, we developed different data-driven soft-sensors for online prediction of a crucial parameter (VFA) and fault detection (FD) and diagnosis in WWTPs.
Firstly, we propose different data-driven soft-sensor for estimating total VFA concentration in anaerobic digestion (AD). We evaluated random forest (RF), artificial neural network (ANN), extreme learning machine (ELM), support vector machine (SVM) and genetic programming (GP) based on synthetic data obtained from the International Water Association (IWA) Benchmark Simulation Model No. 2 (BSM2). In addition, the model robustness was assessed to determine the performance of each soft-sensor under different process states.
Second, to prevent failures and serious consequences during the running of the AD plant, the VFA soft-sensors using different advanced techniques such as SVM, ELM and ensemble of neural network (ENN) are tested and compared in terms of accuracy and FD robustness for detecting process and instrument faults. To compare the proposed approaches with the traditional FD method, a principal component analysis (PCA) model was also developed. By applying soft-sensors, the residual signal, i.e., the difference between estimated and measured VFA values, can be generated. This residual signal was used in combination with univariate statistical control charts to detect the faults.
Third, we propose a complete adaptive process monitoring framework based on incremental principal component analysis (IPCA). This framework updates the eigenspace by incrementing new data to the PCA at a low computational cost. The contribution of variables is also recursively provided using a complete decomposition contribution (CDC). For the imputation of missing values, the empirical best linear unbiased prediction (EBLUP) method is incorporated into this framework.
Overall, this thesis presents the application of different data-driven soft-sensors for online prediction and FD in WWTP; it is also shown that they have strong potential for providing support to the operation of water treatment facilities.
The most important Conclusions Online prediction:
The VFA measurement in the AD process needs dedicated sensors, which are still currently expensive and very sensitive to harsh environmental conditions. Moreover, their measurement is always associated with a delay that is undesirable for real-time monitoring. Therefore, to overcome this issue, we investigated the application of data-driven techniques for online prediction of VFA. The prediction and generalization performances of different data-driven methods were estimated on a specific validation data set obtained from BSM2. Based on our results, it was found that not all data-driven methods are suitable for developing soft-sensors. For instance, the performance of the RF method for the prediction of VFA was very low. The low performance of RF is because it is a rule-based method, in which the data is categorized into different classes. Therefore, if applied to temporal data, weak results will be obtained due to the high number of classes. In contrast, the other methods such as ANN, ELM, SVM, and GP showed higher performance in the prediction of VFA. Among these methods, GP soft-sensor has a high potential for implementation in the control system due to its robustness and transparency compared to the other methods used in this thesis.
Fault detection and isolation:
The AD is considered as a complex process due to the biological steps that occur in it. Thus, precise monitoring of this complex process is mandatory to make the biogas production process more efficient, reliable, and profitable. The developed soft-sensors for online prediction of VFA were used for FD in AD by applying a simulated data set obtained from BSM2. By comparing different soft-sensors, it was found that although some soft-sensors have shown high accuracy in predicting VFA during normal operating conditions, they could not be effectively used for FD purposes because they are not robust enough. Combining the data-driven soft-sensors with SPC charts such as CUSUM chart showed significant improvement in the detection performance of small magnitude faults. A study on the data set with missing values suggests that CUSUM chart is very robust in FD during missing signals events.
The behavior of wastewater treatment processes is usually very non-stationary. It is often challenging to distinguishes the normal operation of the process from those caused by the varying influent conditions. In addition to the mentioned difficulties, the dynamic and nonlinear aspects involved in these processes must be considered. In this thesis, MSPC is proposed as a remedy for these difficulties. PCA is one of the MSPC methods used which used for FD. PCA accounts for collective effects, as it allows for the simultaneous analysis of all included variables. This would be very beneficial when there are many measurements available similar to the situation in WWTPs. A PCA model is trained using data from normal operation of the process and then, it is used to detect deviations from normal behavior. However, due to changing conditions, for instance, diurnal variations, seasonal changes and long term trends, the monitoring model must be updated. For this reason, we proposed the novel FD framework based on the IPCA. Our results, demonstrated with simulated faulty scenarios, show that IPCA is able to adapt time-varying process behavior while detecting and isolating faults. There are some delays in detecting the faults, which is acceptable since the simulated faults were small at the initial stage of their occurrence. IPCA can correctly isolate the fault, although in some cases, it was not possible to isolate the sensor directly and only the impact of the fault on the other measurements could be detected. Based on the proposed framework, very complex missing values pattern (e.g., more than one sensor signal failure at the same time) can be imputed in a real-time framework.
