Cocaine by-product detection with metal oxide semiconductor sensor arrays

A range of n-type and p-type metal oxide semiconductor gas sensors based on SnO2 and Cr2O3 materials have been modified with zeolites H-ZSM-5, Na-A and H–Y to create a gas sensor array able to successfully detect a cocaine by-product, methyl benzoate, which is commonly targeted by detection dogs. Exposure to vapours was carried out with eleven sensors. Upon data analysis, four of these that offered promising qualities for detection were subsequently selected to understand whether machine learning methods would enable successful and accurate classification of gases. The capability of discrimination of the four sensor array was assessed against nine different vapours of interest; methyl benzoate, ethane, ethanol, nitrogen dioxide, ammonia, acetone, propane, butane, and toluene. When using the polykernel function (C = 200) in the Weka software – and just five seconds into the gas injection – the model was 94.1% accurate in successfully classifying the data. Although further work is necessary to bring the sensors to a standard of detection that is competitive with that of dogs, these results are very encouraging because they show the potential of metal oxide semiconductor sensors to rapidly detect a cocaine by-product in an inexpensive way.


XRD patterns of Cr2O3 modified with overlays of zeolite H-Y
. SEM images of a control Cr2O3 sensor (left) and a Cr2O3 sensor coated with three layers of zeolite H-Y (right). Figure S5. Sensor responses of three different SnO2 sensors to two pulses of 50 ppm ethanol at 400 ºC (A). Cr2O3 sensor responses to 5 and 10 ppm toluene, respectively, at 400 ºC (B). This test was performed to understand repeatability from one device to another.

Support Vector Machines
The sensors were selected because they displayed selective characteristics, they also provided distinct response patterns when the response magnitudes towards some gases were similar and because variability between repeat tests was generally found to be minimal. The dataset corresponded to tests carried out at 400 °C. The input data corresponds to that attained upon sensor exposure to 9 analytes: ethanol, ethane, acetone, toluene, propane, butane, methyl benzoate, ammonia and nitrogen dioxide.
Note that nitrogen dioxide was only used to compare whether the model was able to discriminate between ethane and an oxidising gas, given some sensors were previously found to provide resistive responses upon exposure to ethane.
The dataset contained information regarding the maximum conductive and maximum resistive responses of each sensor at different intervals following gas injection into the system. For instance, the response values after 5 seconds, 10 seconds, 50 seconds, 100 seconds, 200 seconds, 300 seconds, 400 seconds, 500 seconds were included initially to see how the models performed. Note that when a sensor was seen to provide high variability among repeat tests, the data was left out of the dataset.
An SVM SMO algorithm (with PolyKernel function) was used to train the dataset. The SMO algorithm offers computational speed, using a one-against-one approach. As suggested in the literature, the oneagainst-one approach is suitable for practical applications aimed at solving multi-class problems with a large number of training samples, as it speeds up the decision-making process and it would therefore be useful in the future as well with further data collection. (2) Kernels are used to solve multi-class problems and whilst some groups suggest that both polynomial and RBF kernel functions provide similar classification performance, other studies suggest that the performance of the polynomial kernel is consistently better when dealing with a large number of attributes. The classification performance of random forests has previously been reported as comparable to that of SVMs. (3) For this reason, the classification performance of random forests was also evaluated to understand the robustness of the SVMs in accurately classifying the data into gas type.  Table S2 -Summary table of all sensors used in this study. Two different concentrations and two different temperatures have been included for illustration purposes. Desorption was assessed according to the peak shapes. Peak tailing is directly related to baseline drift and sensor recovery. T90 and T10 values have been included for sensor heating temperatures of 400 °C, which was the temperature employed for the SVMs. Cells marks with a tick means 'Yes' and with a cross 'No'. Cells marked with '~' mean 'slight'.