Discrimination of VOCs molecules via extracting concealed features from a temperature-modulated p-type NiO sensor

Gas sensor fabrication and sensing testFirstly, the as-prepared NiO nanoparticles were dispersed in ethanol and sonicated for ˜20 min to form a paste. Then, the paste was brushed onto a ceramic substrate (1.5 × 1.5 × 0.25 mm3, with predefined gold testing and heating electrodes) as a sensing layer to make a four-legged sensor. After the natural drying of the paste in ambient temperature, the NiO sensor was aged at ˜ 250 °C (5 V heating voltage) for 7 days to ensure a good stability. In order to eliminate the interference of external thermal couple on the accurate assessment of sensor temperature, a non-contact infrared thermometer (MS-IR, Telops Inc.) was used to calibrate the real temperature of the sensing layer. The relationship between heating voltage and sensor temperature can be found in the supporting information (Fig. S2).The gas response properties were examined in dynamic flowing system, the schematic diagram is shown in Fig. 1. 100–1000 ppm of VOCs analyte was introduced into a mini-chamber, the concentration of VOCs analyte was varied by adjusting the flow rate of the standard gas (1000 ppm in dry air) and dry air, controlled by mass flow controllers (CS-200, Sevenstar Electronics Co., Ltd.). The total flow rate was fiexed to be 500 sccm to eliminate the fluctuation of sensor temperature during gas switching. The electrical responses were meaured by Keithley 4200A-SCS parameter analyzer. A staircase heating waveform [35,37] with voltage range of 3–5 V (programmed by Keithley 4200) is applied to the sensor heater. A period of 20 s of heating waveform represented a compromise between the needs for quick response time and the sharpening of characteristic features of the resistance versus time curve. The temporal variation profile of the sensor resistance R(t) is considered as temperature-modulated response to the testing atmosphere. Before measuring sensor resistance Rgas(t) under VOCs atmosphere, the tests started with the recording of resistance Rair(t) in reference gas (dry air). Then, ethanol, formaldehyde, toluene, benzene and chlorobenzene, each at 9 different concentrations levels (100–900 ppm), 45 analytes in total, were tested. Throughout the experiments, the sampling rate of Keithley 4200 was fixed to be 100 samples/s, and the ambient temperature was monitored, but not controlled, ranging from 23 to 25 ℃.Data processingIn order to extract distinguishable features of VOCs molecules from the raw temperature-modulated response, a signal preprocessing method was carried out through three steps. Step one, the variations of the sensor resistance were converted into the sensitivity responses S(t) under the thermal modulation to get rid of TCR interferencesS(t) = Rgas(t) / Rair(t)then, using the following formula, the responses were normalized to cover the 0–1 magnitude range and to removal concentration related information of VOCsy(t) = (S(t) - Smin) / (Smax - Smin)where Smin and Smax are the minimum and maximum values of sensitivity under thermal modulation, respectively, y(t) denotes the data of response after normalization and t is the time. In the third step, the background noise information was suppressed through the db2 wavelet transformation, and the 1–7 t h level detail coefficients of each response pattern were removed. The low frequency 7th level approximation coefficients were used as the feature vector for the main classification process. The db2 wavelet transformation (a member of DWT) is based on the use of recurrence relations to generate progressively finer discrete samplings of an implicit mother wavelet function. During the transformation, the signal is repeatedly decomposed into low and high frequencies at each level to further increase the frequency resolution. Therefore, the noise information can be sampled and removed via this transform at a sub-space from a time-frequency localization.