Metal oxide semiconductor gas sensors are used in several applications, such as indoor air quality assessment, breath analysis, and industrial emission monitoring. In order to reach the full potential regarding the sensitivity and selectivity of those sensors for any application, temperature cycled operation, as well as careful calibration in the form of building a data-driven model is necessary. However, the calibration can be costly in terms of time and money as every sensor needs individual calibration. This work demonstrates that deep-learning based calibration transfer can be utilised to reuse the calibration model across different environmental scenarios for indoor air quality monitoring. Specifically, a sensor can be calibrated for new background conditions within six hours with the model achieving almost the same performance (RMSE: 16.4 ppb) as calibrated for 80 hours (RMSE: 14.8 ppb). This shows that it is possible to transfer a calibration between datasets with different gases and concentrations. The great reduction of the calibration time by 90 % or more would also allow in field calibration using reference analytical methods.

Predictive maintenance plays a critical role in ensuring the uninterrupted operation of industrial systems and mitigating the potential risks associated with system failures. This study focuses on sensor-based condition monitoring and explores the application of deep learning techniques using a hydraulic system testbed dataset. Our investigation involves comparing the performance of three models: a baseline model employing conventional methods, a single CNN model with early sensor fusion, and a two-lane CNN model (2L-CNN) with late sensor fusion. The baseline model achieves an impressive test error rate of 1% by employing late sensor fusion, where feature extraction is performed individually for each sensor. However, the CNN model encounters challenges due to the diverse sensor characteristics, resulting in an error rate of 20.5%. To further investigate this issue, we conduct separate training for each sensor and observe variations in accuracy. Additionally, we evaluate the performance of the 2L-CNN model, which demonstrates significant improvement by reducing the error rate by 33% when considering the combination of the least and most optimal sensors. This study underscores the importance of effectively addressing the complexities posed by multi-sensor systems in sensor-based condition monitoring.

Developing test-beds to test products and procedures and gain further insights is a common approach in science and industry. During the testing, data is often recorded to provide further insights to the analysts. Since the rise of machine learning and data-driven approaches, these test-beds often get over-instrumented to record as much data as possible. After that, the data is analyzed, and tailored algorithms are applied to achieve a machine learning model with high accuracy. However, many of these models later fail when applied in the real world because they lose their validity due to dataset or domain shift. This means that certain cross-influences were not, or not in their complete range, covered within the recorded data. In this contribution, a test-bed for cylindrical roller bearings has been developed where multiple cross-influences can be varied. It is designed for subsequent leave-one-group-out cross-validation to evaluate the robustness and transferability of machine learning models. Noteworthy features of the test-bed are the possibility of changing the position of the bearing in the test-bed without disassembling it from the shaft (perhaps causing unintentional damages) and that each bearing is measured in its undamaged condition as well before damaging it. In a first measurement campaign, three experiments had been carried out with an automated machine learning toolbox to evaluate the performance of the test-bed design.

Metal oxide semiconductor (MOS) gas sensors operated in temperature cycled operation (TCO) and calibrated with machine learning algorithms are increasingly promising for indoor air quality (IAQ) assessments. This can be attributed to the cost-efficient sensors, with a broad sensitivity spectrum and the possibility of continuous measurements. However, with the ever-increasing complexity of data-driven models used to calibrate the MOS gas sensors, understanding the connection between the raw input and the predicted gas concentration is especially important. In this work, two methods from the field of explainable AI are applied to our custom neural network (TCOCNN) and compared regarding their capability to identify essential parts of the raw input signal. For this purpose, a validation scheme is introduced to rate the explanation methods. Finally, it is shown that with only 7 % of the original raw input, root-mean-squared error (RMSE) values for formaldehyde that are only 22 % worse compared to the absolute best (15.8 ppb vs. 19.3 ppb) can be achieved. This more profound understanding of the sensor can then be used to show differences between sensors, allow more accessible models to be built, and optimize the temperature-cycled operation regarding the number of temperature steps.

Although metal oxide semiconductors are a promising candidate for accurate indoor air quality assessments, multiple drawbacks of the gas sensors prevent their widespread use. Examples include poor selectivity, instability over time, and sensor poisoning. Complex calibration methods and advanced operation modes can solve some of those drawbacks. However, this leads to long calibration times, which are unsuitable for mass production. In recent years, multiple attempts to solve calibration transfer have been made with the help of direct standardization, orthogonal signal correction, and many more methods. Besides those, a new promising approach is transfer learning from deep learning. This article will compare different calibration transfer methods, including direct standardization, piecewise direct standardization, transfer learning for deep learning models, and global model building. The machine learning methods to calibrate the initial models for calibration transfer are feature extraction, selection, and regression (established methods) and a custom convolutional neural network TCOCNN. It is shown that transfer learning can outperform the other calibration transfer methods regarding the root mean squared error, especially if the initial model is built with multiple sensors. It was possible to reduce the number of calibration samples by up to 99.3% (from 10 days to approximately 2 h) and still achieve an RMSE for acetone of around 18 ppb (15 ppb with extended individual calibration) if six different sensors were used for building the initial model. Furthermore, it was shown that the other calibration transfer methods (direct standardization and piecewise direct standardization) also work reasonably well for both machine learning approaches, primarily when multiple sensors are used for the initial model.

In this study, methods from the field of deep learning are used to calibrate a metal oxide semiconductor (MOS) gas sensor in a complex environment in order to be able to predict a specific gas concentration. Specifically, we want to tackle the problem of long calibration times and the problem of transferring calibrations between sensors, which is a severe challenge for the widespread use of MOS gas sensor systems. Therefore, this contribution aims to significantly diminish those problems by applying transfer learning from the field of deep learning. Within the field of deep learning, transfer learning has become more and more popular. Nowadays, building a model (calibrating a sensor) based on pre-trained models instead of training from scratch is a standard routine. This allows the model to train with inherent information and reach a suitable solution much faster or more accurately. For predicting the gas concentration with a MOS gas sensor operated dynamically using temperature cycling, the calibration time can be significantly reduced for all nine target gases at the ppb level (seven volatile organic compounds plus carbon monoxide and hydrogen). It was possible to reduce the calibration time by up to 93% and still obtain root-mean-squared error (RMSE) values only double the best achieved RMSEs. In order to obtain the best possible transferability, different transfer methods and the influence of different transfer data sets for training were investigated. Finally, transfer learning based on neural networks is compared to a global calibration model based on feature extraction, selection, and regression to place the results in the context of already existing work.

Metal oxide semiconductor (MOS) gas sensors used for indoor air quality monitoring require an intensive calibration to accurately quantify volatile organic compounds at ppb (parts per billion) level in complex gas mixtures. With the help of advances in the field of deep learning, particularly the use of convolutional neural networks together with neural architecture search, the noise of the quantification model can be reduced significantly with an uncertainty for xylene of 27 ppb. However, the calibration takes several days up to now. In this work, the concept of transfer learning is studied to reduce the required calibration time. It is shown that the calibration time of single sensors can be reduced by 96 %. The resulting uncertainty is only 21 ppb worse than the absolute best, i.e. the value for a complete individual calibration, which is sufficiently good. By slightly increasing the calibration time for transfer learning to 30 % of the initial time, an uncertainty value for xylene quantification of 36.3 ppb was achieved, and thus only 9.7 ppb worse than the best possible model.

In this work, a novel approach is presented to extract physically motivated features for damage detection of gears in single-stage gearboxes by an automated order analysis of the instantaneous angular speed. This extraction method was applied to measurement data of various magnetoresistive sensors installed in a gearbox test bed and the obtained characteristics were examined in validation scenarios for their susceptibility to external disturbances. The classification results were compared to results obtained with an automatic Machine Learning (ML) method both on the data of the magneto-resistive sensors and an accelerometer. The new method has major advantages, especially with respect to transferability to other rotational speeds.

This paper shows the opportunities of data preprocessing and how it influences the time required to record a sufficient amount of valid calibration data samples. Specifically, we approach the minimum needed time for calibration from two sides: on the one hand, repetitions are omitted for training one by another to define the lowest number of valid data that is needed for a model to achieve a reasonable accuracy. On the other hand we add samples, that are labeled as valid data points by steady-state detection to the dataset compared to a time-consuming manual annotation. The results will be demonstrated on a dataset of a metal oxide semiconductor gas sensor in temperature cycled operation measuring mixtures of artificial room air containing several volatile organic compounds and quantifying formaldehyde which is carcinogenic and therefore of high concern in indoor environments. The dataset is generated with an automated gas mixing system and then optimized with the help of data pre-processing methods based on steady-state detection, outlier detection and ResNet neural networks. The dataset can be reduced to only 50 % of the original data and is still able to train an artificial neural network with a root mean square error smaller than 25 % compared to the guideline value for formaldehyde concentration defined by the WHO.

Continuous accurate quantification of hazardous VOCs for monitoring Indoor Air Quality (IAQ) with low-cost sensor systems would allow demand-controlled ventilation to significantly reduce health effects but is still an elusive goal. Especially monitoring of formaldehyde is in high demand due to its extensive emission from a wide range of sources, especially building materials, furniture, textiles and cleaning products [1]. While detection of formaldehyde is possible at concentrations well below the WHO recommended short-term (30-minute) guideline value of 0.1 mg/m3 (81 ppb) [1] using typical MOS sensors [2] achieving accurate quantification in complex indoor environments also requires a high level of selectivity. In this contribution we demonstrate formaldehyde quantification with an uncertainty of 11.3 ppb, corresponding to 15% of the WHO recommended guideline value, in a complex mixture of VOCs and other interfering gases typical for indoor environments. Our approach is based on a MOS sensor with four gas sensitive layers integrated on one micro hotplate (SGP30, Sensirion, CH) [3]. With extended access (possible with a non-disclosure agreement) it is possible to set the temperature of the heater and read out the resistance of the individual layers. Therefore, we can combine physical and virtual multisensor methods for the generation of multiple signals by operating the four different gas sensitive layers in temperature cycled operation (TCO). For selective measurement of VOCs in indoor environments, we have designed a complex temperature cycle comprising 12 different low temperatures ranging from 100 to 375°C interlaced with high temperature phases of 425°C. The total duration of the temperature cycle (T-cycle) is 120 s, which is suitable for IAQ applications where the gas composition changes slowly. Calibration is based on an automatic gas mixing apparatus (GMA) [4] using random mixtures [5]. The calibration scheme is based on mixtures of four VOCs (formaldehyde, acetone, benzene and toluene) as well as carbon monoxide (CO) and hydrogen (H2) as typical interfering gases. In addition, relative humidity is also varied. RH and all six gas concentrations are randomly selected from predefined distributions for each variable to reflect typical variations of these gases in indoor environments. Each specific mixture is then offered in the GMA for 20 min, i.e. 10 T-cycles, and the calibration model is built using only those cycles with stable signal patterns, i.e. cycles are excluded when the gas concentrations are not in steady state. The results presented here are based on a calibration run with a total of 495 gas exposures. The resistance values of the four gas sensitive layers are recorded every 25 ms resulting in 4 x 4800 raw data samples for every T-cycle. Features are extracted using adaptive linear approximation (ALA), i.e. the individual cycles are approximated by linear segments and the mean and slope of each segment are calculated for further evaluation. The achieve high performance quantification for formaldehyde we use a 10-layer deep modified ResNet [6]. The performance of the prediction is validated thoroughly using 10-fold cross validation. Specifically, the overall dataset is split into 10 subsets based on the gas exposures and each subset is used a test data for a model trained with the remaining 9 subsets. The prediction performance is therefore given as Root Mean Square Error for Validation (RMSEV) and the variation of the RMSEV for the different splits is also checked to ensure stable predictions. Note that this approach allows determination of prediction models for each of the gases included in the test. For formaldehyde, tested over a range from 0 to 700 ppb we achieve an RMSEV of 11.3 ppb, which is fairly stable over the full range, see Fig. 1. This corresponds to an uncertainty of less than 15% of the WHO guideline value independent of the concentrations of acetone (range 0 .. 1.850 ppb), benzene (0 .. 1180 ppb), toluene (0 .. 250 ppb), CO (100 .. 2000 ppb), H2 (300 .. 2500 ppb) and rel. humidity (25% .. 75% RH). The variance of the RMSEV for the different splits is 3 ppb, indicating that a reliable prediction model is achieved. The overall duration of the calibration run can be reduced considerably by ML-based optimized selection of the T-cycles with stable patterns while still achieving an RMSEV for formaldehyde of 20 ppb, corresponding to an uncertainty of 25% of the WHO guideline value [7]. The resulting prediction models for different VOCs were further validated in realistic indoor field tests using standard VOC analysis (e.g. sampling on Tenax followed by GC/MS analysis) and performing release tests, i.e. evaporation of a certain amount of VOC in a room and comparing the measurement predictions with the theoretical values expected for uniform distribution. [1] D.A. Kaden et al.: Formaldehyde, in: WHO Guidelines for Indoor Air Quality: Selected Pollutants, Geneva: World Health Organization; 2010, ISBN: 978-92-890-0213-4. [2] M. Leidinger et al., J. Sens. Sens. Syst., 3 (2014), pp. 253-263, doi:10.5194/jsss-3-253-2014. [3] D. Rüffer, F. Hoehne, J. Bühler, Sensors, 18/4 (2018), 1052, doi:10.3390/s18041052. [4] N. Helwig et al., Meas. Sci. Technol. 25 (2014) 055903, doi:10.1088/0957-0233/25/5/055903 [5] T. Baur et al., J. Sens. Sens. Syst. (2020) 9, 411-424, doi:10.5194/jsss-9-411-2020. [6] K. He et al.: Deep residual learning for image recognition, 2015, [Online] last access: 30 November 2020, available: http //arxiv.org/pdf/1512.03385v1:PDF [7] Y. Robin et al.: Machine Learning based calibration time reduction for Gas Sensors in Temperature Cycled Operation, IEEE I2MTC – International Instrumentation and Measurement technology Conference 2021, submitted. Figure 1

Process sensor data allow for not only the control of industrial processes but also an assessment of plant conditions to detect fault conditions and wear by using sensor fusion and machine learning (ML). A fundamental problem is the data quality, which is limited, inter alia, by time synchronization problems. To examine the influence of time synchronization within a distributed sensor system on the prediction performance, a test bed for end-of-line tests, lifetime prediction, and condition monitoring of electromechanical cylinders is considered. The test bed drives the cylinder in a periodic cycle at maximum load, a 1 s period at constant drive speed is used to predict the remaining useful lifetime (RUL). The various sensors for vibration, force, etc. integrated into the test bed are sampled at rates between 10 kHz and 1 MHz. The sensor data are used to train a classification ML model to predict the RUL with a resolution of 1 % based on feature extraction, feature selection, and linear discriminant analysis (LDA) projection. In this contribution, artificial time shifts of up to 50 ms between individual sensors' cycles are introduced, and their influence on the performance of the RUL prediction is investigated. While the ML model achieves good results if no time shifts are introduced, we observed that applying the model trained with unmodified data only to data sets with time shifts results in very poor performance of the RUL prediction even for small time shifts of 0.1 ms. To achieve an acceptable performance also for time-shifted data and thus achieve a more robust model for application, different approaches were investigated. One approach is based on a modified feature extraction approach excluding the phase values after Fourier transformation; a second is based on extending the training data set by including artificially time-shifted data. This latter approach is thus similar to data augmentation used to improve training of neural networks.

With air quality being one target in the sustainable development goals set by the United Nations, accurate monitoring also of indoor air quality is more important than ever. Chemiresistive gas sensors are an inexpensive and promising solution for the monitoring of volatile organic compounds, which are of high concern indoors. To fully exploit the potential of these sensors, advanced operating modes, calibration, and data evaluation methods are required. This contribution outlines a systematic approach based on dynamic operation (temperature-cycled operation), randomized calibration (Latin hypercube sampling), and the use of advances in deep neural networks originally developed for natural language processing and computer vision, applying this approach to volatile organic compound measurements for indoor air quality monitoring. This paper discusses the pros and cons of deep neural networks for volatile organic compound monitoring in a laboratory environment by comparing the quantification accuracy of state-of-the-art data evaluation methods with a 10-layer deep convolutional neural network (TCOCNN). The overall performance of both methods was compared for complex gas mixtures with several volatile organic compounds, as well as interfering gases and changing ambient humidity in a comprehensive lab evaluation. Furthermore, both were tested under realistic conditions in the field with additional release tests of volatile organic compounds. The results obtained during field testing were compared with analytical measurements, namely the gold standard gas chromatography mass spectrometry analysis based on Tenax sampling, as well as two mobile systems, a gas chromatograph with photo-ionization detection for volatile organic compound monitoring and a gas chromatograph with a reducing compound photometer for the monitoring of hydrogen. The results showed that the TCOCNN outperforms state-of-the-art data evaluation methods, for example for critical pollutants such as formaldehyde, achieving an uncertainty of around 11 ppb even in complex mixtures, and offers a more robust volatile organic compound quantification in a laboratory environment, as well as in real ambient air for most targets.