Proof-of-Concept:
How Semantic Folding can help monitor the operational status of municipal water pumps
Goal
Develop a predictive model that can anticipate future pump failures based on multiple sensor values
Data Set
Public data set hosted on Kaggle with 220k data points collected minute-by-minute over a span of 4 months from the 52 sensors installed on a pump
This short video compares the evolution of operational status fingerprints versus failure status fingerprints.
Challenges
- Out of the 220k data points available, only a couple of hundreds are anomaly records, depicting 7 events where the pump was not operational. This relatively small number poses a challenge for training deep learning models.
- No information about the sensor types or measurement units were available, making the parametrization of data filters very complex if not impossible.
- High performance computers are required to analyze multiple sensor data streams in real time.
Solution
First, a semantic space is created:
- A reference dictionary reflecting the water pump’s normal working status is built by including all records labeled as “normal” and excluding all records labeled as “broken” and “recovering”.
- The raw sensor data are converted into sparse distributed vectors (semantic fingerprints) where each active bit has a similar meaning across all fingerprints.
- The whole process is done by means of unsupervised training and takes about 15 minutes on a standard notebook.
During operations, 52 sensor values are recorded every minute. Each of these individual values is converted into a contextualized semantic fingerprint, and then, for each measurement record, the 52 fingerprints are aggregated and sparsified to build a representative status fingerprint at a given instant. Because of the fundamental properties of sparse vectors, no information is lost in this aggregation process.
Results
- When the entire time series of the pump sensors is converted into semantic fingerprints, a dynamic chronology (a “movie”) depicting the evolution of the system states emerges.
- One striking observation is the high degree of resemblance between consecutive status fingerprints, indicating that the coherence of the records is very high. In other words, Semantic Folding makes it possible to detect gradually-appearing anomalies at an early stage.
The Cortical.io Difference
- Semantic Folding enables to capture the hidden semantics of a system based on discrete metrics only.
- Representing sensor data with semantic fingerprints permits to discriminate between different states, for example between operational and failure states, without detailed knowledge of the associated sensor patterns.
- In the context of predictive maintenance, a database of fingerprints representing failure states could be used to measure in real time how closely the current operational status of the pump approximates a failure condition.
- The Semantic Folding method can be applied irrespectively of the complexity of the observed system, the number of sensors, or their types, as long as sufficient sensor metrics can be derived.
- Calculations based on semantic fingerprints are highly efficient because the measurement vectors are processed bit-wise and not by traditional vector multiplication (dot product), enabling their use in real-time and/or embedded applications.
