MEDICAL GAS SYSTEMS
signals as input, and fault prediction and detection as output including predicting the time when system is operating at normal, or first warning, second warning, and fault detected when or before it is happening. After fault prediction and detection, the machine learning system will also be trained to learn and diagnose what caused the faults and what actions can be taken to prevent the failure or reduce system damage.
Figure 2. Multi-sensor measurement signals in time domain.
management, and maintenance, by collecting data from all sources on the IoT networks. By real-time processing of
measurement data provided by dedicated sensors installed, autonomous decision making is enabled based on the online diagnosis of the system with right conditions, leading to increased reliability towards zero defects. Predictive maintenance based on machine condition monitoring can also be planned and scheduled, such that by deep learning, a knowledge-based system can be built to correct errors and improve efficiency. For MIMO systems of industrial
processes, such as the multi-sensor signals as input to predict and detect time of faults as output, prediction can be achieved by the nonlinear mapping or fitting that models the observed past data, and we aim to minimise the mean squared error (MSE) between the actual inputs and predicted outputs at the training process. The fault prediction is a complex nonlinear mapping where
deep wavelet neural networks (DWNN) with nonlinear fitting is developed by monitoring multi-sensor signals with feature extracted in time-frequency domain. To better map the inputs to outputs with highly nonlinear relationships, we construct deep wavelet neural networks (DWNN) as shown in Figure 1. In this study, we use Mexican hat wavelet as activation functions. It is the negative normalised second derivative of a Gaussian function and has admissibility condition and a symmetric structure. Its derivative is also a Gaussian function. Compared with commonly used deep
neural networks, using nonlinear Gaussian (Mexican hat) wavelet with sparsity as activation functions in the DWNN model can improve nonlinear fitting and convergence speed with adaptive learning rate, so that the power and capabilities of machine learning can be raised. To predict faults from multi-sensor
measurements, Figure 1 gives an example of a DWNN network to train four sensor
Experiments and discussions To evaluate the proposed deep learning based remote monitoring of MGS systems, a medical air plant to supply oxygen was employed. Based on a fixed-speed compressor, and with remote monitoring of multi-sensor signals over the built IoT networks and real-time machine learning developed for fault prediction and detection, the measurement signals of mains current (to measure status of compressor operating); flow rate (at pipeline output); oxygen line pressure (at pipeline output); and oxygen purity (to measure the quality of oxygen) were remotely collected and processed at a cloud server, where deep learning by DWNN was performed. The measurement data shown in
Figure 2 was remotely collected from four sensors in the plant and faults were detected following the rule that when oxygen purity is lower than a pre-set level that is not acceptable for medical use. Figure 3 shows the details of multi-sensor measurement signals in time domain, when the system was operating at normal conditions with no fault being detected. Under normal conditions, compressors
are operating at a loading and off-loading cycle, and that makes the sensor values of flow rate and oxygen line pressure vary at the same cycle. The collected data analysis in time domain (as time series: a sequence of data indexed by successive data points) can only provide limited information for fault prediction, giving no early warning. To predict faults at an early stage and
Figure 3. Details of multi-sensor measurement signals in time domain. IFHE DIGEST 2022
to optimise safety and efficiency, we propose to transform the raw multi-sensor data into the time-frequency domain to obtain a big feature vector containing all the relevant features from all of the sensors, where the change of operation condition features can be revealed by the time-frequency representation of the signals. The developed fast lifting wavelet transform is implemented to decompose the signals in both the time and frequency domain. Figure 4 gives an example of how the signals (as in Figure 2) can be represented in time-frequency domain, where each signal is decomposed by wavelets into low and high frequency parts. The features extracted in the time-frequency domain can then be used to feed the
73
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60 |
Page 61 |
Page 62 |
Page 63 |
Page 64 |
Page 65 |
Page 66 |
Page 67 |
Page 68 |
Page 69 |
Page 70 |
Page 71 |
Page 72 |
Page 73 |
Page 74 |
Page 75 |
Page 76 |
Page 77 |
Page 78 |
Page 79 |
Page 80 |
Page 81 |
Page 82 |
Page 83 |
Page 84 |
Page 85 |
Page 86 |
Page 87 |
Page 88 |
Page 89 |
Page 90 |
Page 91 |
Page 92 |
Page 93 |
Page 94 |
Page 95 |
Page 96 |
Page 97 |
Page 98 |
Page 99 |
Page 100 |
Page 101 |
Page 102 |
Page 103 |
Page 104 |
Page 105 |
Page 106 |
Page 107 |
Page 108 |
Page 109 |
Page 110 |
Page 111 |
Page 112 |
Page 113 |
Page 114 |
Page 115 |
Page 116
