A soft sensor estimates a quantity that is hard or costly to measure by computing it from other, easier measurements with a model.
A soft sensor, or virtual sensor, is software that estimates a variable from other measurements instead of measuring it directly. It is used when the variable is expensive, slow, unreliable or impossible to measure in real time. It works by finding a relationship between the hard-to-measure target and easy online signals. The model can be physical (a mass balance, a hydraulic equation), statistical (regression, partial least squares) or machine learning (random forests, neural networks). It is trained on periods when both the target and the inputs were recorded, for example lab results matched to online sensor data, and then runs on live inputs. Treatment plants offer common examples. The standard laboratory test for biochemical oxygen demand takes five days, but a soft sensor can estimate effluent quality from online readings of flow, turbidity, ammonium, dissolved oxygen and conductivity. In drinking water plants, soft sensors can suggest a coagulant dose from raw water turbidity, pH and temperature. In supply networks, flow in an unmetered pipe can be estimated from pump power and pressure readings. In rivers, turbidity is widely used as a stand-in for suspended sediment concentration. What it needs: a good history of paired data covering the range of conditions, input sensors that are reliable and well maintained, and regular checks against lab or reference measurements. Strengths: continuous estimates at low cost, results faster than lab analysis, and a backup if a physical analyser fails. Limits: the estimate is valid only within the conditions seen in the training data. If the process changes, through a new treatment step, a change in source water or sensor drift, the soft sensor can drift away from the truth without any sign. Recalibrate it regularly and check its output against grab samples.