Today I am going to introduce you to the data-science we are doing in Tyde. Currently we are using Artificial Intelligence to do anomaly detection for assets, such as transformers, stators, rotors. What we used is a state-of-the-art technique called Deep Learning.
The human brain consists of many neurons. The neurons are connected to each other to process the received information. Deep learning uses mathematical functions to simulate the function of neurons.
Similar to the brain’s neurons, the mathematical neurons are connected to each other to form neural networks. Depending on the topology of the connections, various types of neural networks can be formed, each with different abilities and purposes.
In Tyde, the deep learning technique used to detect anomalies is called autoencoder.
Autoencoder is a type of neural network that attemps to reconstruct its inputs. An autoencoder has two parts, encoder and decoder. The encoder part turns the original high dimensional input into a compressed representation with lower dimension. By compressing the original data, we can extract the most important information from the original data. After the compression, we use the decoder part to reconstruct the original input, based on the compressed representation.
In Tyde, we collect high volume and high throughput data about the normal operation of assets in the hydropower plants. We use the normal data to train the autoencoder, and use the reconstruction error, that is, the difference between autoencoder input and output, as an indicator of anomalies. If the reconstruction error is large, it indicates that the autoencoder has never see this kind of data before, so the input data has a higher probability of being in an non-normal state. It’s just like with an experienced engineer with long experience in the hydropower industry which has seen lots of data about the normal operation of the hydropower plants; if one day the engineer sees some data with values that has never appeared before, it indicates that something has happened and the engineer will react to it and move forward with further investigation.
Assets in the hydropower plants has a hierarchical architecture; for example, a generator has several sub-assets such as rotor and stator. We use the encoder part of the autoencoder, to get the most important information from lower level assets, then feed the compressed data into the higher-level asset, by doing so, we can use as much the information of an asset as possible.
By clicking an asset in the asset diagram, you can get access to the user interface for the anomaly detection. The red color indicates which month and which days the asset has anomalies. By clicking a specific day, we can find the sensor readings with the corresponding anomaly score in the user interface. The anomaly score is based on the reconstruction error. As we can see, if the model feels the sensors have some anomaly behavior, the model will generate a high anomaly score, and mark a red bar, then Tyde will send some notifications to the operator.
In Tyde, the training of models is very easy because we have a special service which we call the Orchestrator. The orchestrator is fully automatic; it determines what training needs to be done, which assets are lacking models, which model must be trained first, and so on. The Orchestrator also allows advanced users to train, delete and tuning the models using his/her own hyperparameters, directly through the user interface. Finally, users can select and activate different versions of models, according to the model’s performance, and it all happens with one click.
Thanks for reading.
Duo Zhang, Data Scientist at Broentech Solutions