Sensor Analysis FrameworkLink
The following section details the framework used for post processing the data in the iScape project.
The Sensor Analysis Framework is mainly used to handle sensor data acquisition and apply sensor models for actual pollutant concentration calculations. In the following sections, we detail the principles of low cost sensor calibration that we follow, as well as the usage within the framework.
All the files and notebooks related to this document can be found online in our Github repository.
How we use itLink
Generally, the sensor data is ingested in the Sensor Analysis Framework directly from the API, or by inserting it via csv. As detailed in the Low Cost Sensor Calibration Section, each of the sensors is treated differently, and some of them require more analysis than others, e.g.: the Metal Oxyde sensors vs the electrochemical sensors.
The framework can be used to process the Living Labs Station's data, loading the data and applying tailor made algorithms for the sensors within:
Additionally, the Framework can be used to load in data from the Smart Citizen Kit's sensors and used to generate models and analysis based on other references, following the diagram below:
Similarly, this approach can be used to improve the Living Labs Station's data, where known effects from temperature or humidity can be input as corrections for the models generated.
Finally, all the models generated are stored and can be easily loaded from the Github Repository.
The structure of this framework can be split between two main areas: one dedicated to sensor data analysis and model generation; and an automated sensor correction script that recovers sensor data from the API, processes it, and posts it back to the API with the use of the results obtained in the data analysis. This structure is decribed in the following diagram:
Furthermore, within the sensor data analysis, a notebook and several scripts are included, intended for modeling and data visualisation, including exploratory data analysis and a testing environment for sensor model calibration. This includes interfacing with the SmartCitizen API in order to download available sensors from the platform, as well as local csv analysis. Further functionalities are explained in the following sections.
A deeper lookLink
The main component of the Framework is a Jupyter Notebook with the following features:
- An interface to either retrieve data from the Smart Citizen's API in a simple way or to load them from local sources (in csv format, compatible with the SCK SD card data)
- A data handling framework based on the well known Pandas package
- An exploratory data analysis interface to study sensor behaviour and correlations
- A flexible sensor calibration model interface with classical statistical methods such as linear regression, ARIMA, SARIMA-X among others, as well as more modern Machine Learning techniques with the use of LSTM (Long short term memory) networks and MLP (Multi Layer Perceptron) models for sequential data prediction and forecasting
- An interface to statistically validate and study the performance of these models, export and store them
- As a bonus, an interface to convert the python objects into the statistical analysis language R
The framework also provides several functionalities within signal analysis field using numpy and scipy frameworks.
An example of the workflow can be seen below:
Step by step guides
Loading in the dataLink
As mentioned, data can be downloaded from the SmartCitizen API with the KIT IDs or using local csv. In order to tidy up the data, the recordings are organised around the concept of test, an entity containing all the kits' references, sensors and general information regarding the conditions at which the measurements were carried out:
- Test Location, date and author
- Kit type and reference
- Sensor calibration data or reference
- Availability of reference equipment measurement and type
A brief schema of the test structure is specified below:
All this structure is filled up at the test creation with a dedicated script, saving future time to understand mismatching reading units, timestamps formats and so on.
Finally, the devices' data is stored as Time Series data, with DateTime index in ISO8601 format. This is an important consideration for data visualisation and posterious modelling.
Exploratory data analysisLink
The device's data can be quickly analysed using a simple interface in order to quickly select the desired channels and timeframes. Some of the functionalities already included are:
- Time Series visualisation
- Correlation plot and pairs plot
- Heatmaps for geospatial data
The data models section includes an easy to use and full interface to select data from different devices within a test in order to calibrate different models. Since the data is mainly based on Time Series analysis, it interfaces with common statistics and machine learning frameworks such as sci-kit learn, tensorflow, keras, and stats models. These frameworks provide tools to perform:
- Outliers detection with Holt-Winters methods (triple exponential smoothing)
- Data study and analysis for multicollinearity and autocorrelation in order to determine significant variables and avoid model overfit with non-significant exogenous variables
- Trend decomposition and seasonality analysis
- Data split between training and test dataset, with the possibility of rolling prediction
- Baseline model estimations in order to assess minimum targets for model quality (using naive regression models)
- Ordinary Linear Regression techniques for univariate and multivariate linear and non-linear independent variables
- ARIMA-X (Autorregresive, Integrated, Moving Average) models with exogenous variables using Box-Jenkis parameter selection methods
- More advanced machine learning techiques with RNN (Recurrent Neural Networks) for sequence predictions:
- MLP (MultiLayer Perceptron) model with basic RNN cells
- Single and multiple layers LSTM (Long-Thort Term Memory) networks with the possibility of including several shifted variables in the model training and prediction
An example of the model is shown below for the estimation of the SGX4514 CO with the use of the rest of the Kit's available sensor, using a single layer LSTM network only two weeks of training:
Depending on the model selected, different validation techniques are implemented, in order to verify models' assumptions and avoid data misinterpretation (i.e. Durbin Watson or Jacque Bera test for linear regression). Finally, it is important to follow carefully the instructions as stated in the notebook, in order to avoid low model quality.
Model import/export and storageLink
Once the model is analysed and validated, it can be saved and exported. This allows using the model in the future with the same variables in other sensor studies. The model objects are serialised with joblib and can be uploaded to the Models Github Repository for later use.