<p>The proliferation of efficient edge computing has enabled a paradigm shift of how we monitor and interpret urban air quality. Coupled with the dense spatiotemporal resolution realized from large-scale wireless sensor networks, we can achieve highly accurate realtime local inference of airborne pollutants. In this paper, we introduce a novel Deep Neural Network architecture targeted at latent time-series regression tasks from continuous, exogenous sensor measurements, based on the Transformer encoder scheme and designed for deployment on low-cost power-efficient edge processors.


Sequential skeleton and average foot pressure data for normal and five pathological gaits (i.e., antalgic, lurching, steppage, stiff-legged, and Trendelenburg) were simultaneously collected. The skeleton data were collected by using Azure Kinect (Microsoft Corp. Redmond, WA, USA). The average foot pressure data were collected by GW1100 (GHIWell, Korea). 12 healthy subjects participated in data collection. They simulated the pathological gaits under strict supervision. A total of 1,440 data instances (12 people x 6 gait types x 20 walkings) were collected.


Arbitrarily falling dices were photographed individually and monochromatically inside an Ulbricht sphere from two fixed perspectives. Overall, 11 dices with edge size 16 mm were used for 2133 falling experiments repeatedly. 5 of these dices were modified manually to have the following anomalies: drilled holes, missing dots, sawing gaps and scratches. All pictures in the uploaded pickle containers have a resolution of 400 times 400 pixels with normalized grey scale floating point values of 0 (black) through 1 (white).


The datasets contain files for training (“x_training.pickle”, w/o anomalies) and testing (“x_test.pickle”, w/ and w/o anomalies). Labels were saved in “y_test.pickle” whereas label zero correspond to non-anomalous data. Because the pose of the falling dice was not constrained the two fixed perspectives had the chance to see anomalies at all in 60 out of 100 experiments. Hence the test dataset contains 60 anomalous samples. Furthermore, data is augmented w.r.t. erased patches, changes in image constituents like brightness, and altered geometry like flipping and rotating.The shapes of the pickles are

  • w/o augmentation, x_train.pickle: (2000, 2, 400, 400)
  • w/o augmentation, x_test.pickle: (133, 2, 400, 400)
  • w/o augmentation, y_test.pickle: (133,)
  • w/ augmentation, x_train.pickle: (4000, 2, 400, 400)
  • w/ augmentation, x_test.pickle: (133, 2, 400, 400)
  • w/ augmentation, y_test.pickle: (133,)