A treatment unit attached to a wastewater treatment plant makes it possible to recover the biogas. This processing unit is young, operated recently, and the majority of the defects encountered in operation are not known. The real-time analysis of data from the processing unit is made complex by the very large number of sensors (>1500).
On the one hand, the customer wants to be able to understand the failures and automate their detection (compared to a manual analysis, required as it is), and on the other hand wants to anticipate the failures in order to be able to react upstream.
After an iterative phase on models, a selection of really relevant sensors was made, resulting in reducing the number of sensors to be used per model. In addition, a decomposition into three subsets was proposed to the operator, with the creation of as many models, allowing better analysis.
The use of DiagFit made it possible, in a short time, to create models to monitor the dynamics of a complex system.
Example of time series of measurements from different sensors, where the appearance of anomalies (areas in red) is not perceptible “to the eye” or with simple strategies for exceeding thresholds, because these are linked for the essential to anomalous complex correlations between sensors.
The results showed an ability to detect failures in the installation, with anticipations of a few hours to a week in relation to the alarms reported by the plant to the operating staff. These were in particular broken down or degraded equipment – detection of an H2S sensor out of service (anticipation of a day), pressure fault on a compressor which led to the change of a filter (anticipation of a few hours). Defects in the process chain were also identified – abnormal presence of activated carbon in the tank (two days in advance), or an excessively high H2S level (one week in advance).