The Application of "Deep Learning" Neural Networks for Sensor Failure Identification in the Kansas Mesonet

One of the many challenges in the collection of real-time meteorological data is the persistent threat of sensor malfunction. Real-time and automated quality control are a necessity for anyone who plans to collect this data. However, erroneous data can often avoid initial flagging and quality control/assurance procedures and demands more in depth analysis to identify malfunctions. “Deep Learning” neural networks, which can be trained based on historical weather data and previous station failures. This could potentially be deployed as a new method of real-time sensor failure identification. The Kansas Mesonet is a network of 80+ weather stations that collect data with 14 sensors and numerous measured and calculated variables. Decreasing response time to sensor failure with improved error detection would greatly increase the Mesonet’s capability to quality control live data, and repair faulty sensors.

A “neural network” consisting of four total layers of neurons was applied to previous data at the Kansas Mesonet’s Garden City station’s temperature sensor. The network consists of an input layer of 288 neurons representing a daily sum of five minute intervals. This data is then passed through two “hidden layers” of 512 neurons each continually adjusting the hidden weights and biases. These then “activate” the next layer of neurons. In this case, a piecewise linear-defined Rectified Linear Unit (ReLU) activation function is used to scale the weights and biases. During the network’s forward pass, this determines the neuron activation of subsequent layers. At the final layer, the “Softmax” activation function scales the network’s output as levels of “confidence” in each mode of sensor failure identification. Then, a “backward” pass is performed on the network. This optimizes the network’s weights and biases further for future training iterations. A combination of Stochastic Gradient Descent with momentum and Root-Mean-Square (RMS) error propagation, called an “Adam” optimizer, is utilized with a final data pass. This continuously tunes the network by changing the weights and biases in order to minimize loss (which is calculated by categorical cross-entropy).

This network was trained using historical data from the Kansas Mesonet’s Garden City station’s temperature sensor. This station was one of the original 13 of the Kansas Mesonet with data from 1986 to 2022 available for training the dataset. Different types of error were assigned a number, and each day of 2 meter 5-minute data classified as “Healthy” or with one of the following failure designations: “Unrealistic Measurement”, “Missing Data”, “Excessive Noise”, and “No Change in Data”. The training dataset was then passed forward and backward through the network in randomized batches of 64, optimizing the network. Then, Garden City’s “testing” current 2023 dataset to date was introduced in an attempt to categorize the different modes of sensor failure. In the future, the Kansas Mesonet hopes to utilize these “deep learning” networks at each station/sensor in order to rapidly increase the response time to different types of sensor failure, which can decrease the total amount of unrepresentative data found by Mesonet users.