Algorithms
This section provides a high-level description of the algorithms used in the analysis backend of Hexastate's app.
- Fault Trend Detection
- Anomaly Detection
- Threshold Alerting
- Health (Simple & Advanced)
Fault Trend Detection
This algorithm focuses on utilizing FFTs searching for symptoms of specific faults in the machines.
It uses a baseline of FFT data for the machine and compares new FFT data with the baseline. It then iterates through a collection of known symptoms of specific faults (e.g. Unbalance) and tests if it is present and in which degree in new FFT data when compared to the baseline.
The output of this model is a list of fault name and increase pairs.
Example
The fault name could be "Unbalance", and the increase could be 0.68%. This output would mean that symptoms of Unbalance has increased by 0.68% since the baseline was created.
Anomaly Detection
The analysis backend has more than 1 anomaly detection algorithm. The backend uses 2 different models checking for anomalies in the data.
One of the models utilizes a time-series of a specific data source, e.g. RMS data, and then calculates if each new RMS measurement is an anomaly or not.
The other model uses multiple data sources, e.g. temperature, RMS, and power consumption, to cross-analyze if any measurements can be considered an anomaly.
Threshold Alerting
The threshold alerting algorithm is simply an alarm system that checks if any data source, e.g. temperature, crosses a specific threshold like 35 degrees Celsius.
Health
Our most sophisticated models are the algorithms focusing on interpreting all the collected data of a machine to estimate the overall health of the machine.
These algorithms utilizes all data recorded by the sensors installed on a machine to estimate the current health of the machine.
The algorithms comes in 2 different versions: "simple" & "advanced". The "simple" algorithm is our first generation of this algorithm, while the "advanced" algorithm is the second generation. Advanced Health is built to be more flexible and customizable. Our experience is that no production environment is identical. Therefore, we saw a need to build a highly customizable algorithm that we can adjust to any production environment we encounter.
Why is the recipe of the Health algorithms not described?
As the Health algorithms is our most sophisticated models which we have spend countless hours developing, we don't want to provide details of its implementation.
The health models estimates the current machine health as one of the zones from the table below:
| Status | Color | Description |
|---|---|---|
| Good | Dark Green | The asset is in its best state. |
| Satisfactory | Light Green | Early signs of degradation or minor increases in vibration energy. |
| Unsatisfactory | Orange | Clear signs of degradation and excessive vibration energy. |
| Unacceptable | Red | Excessive vibration energy. Might require immediate action, either as component replacement, maintenance or installation optimizations. |