MEDICAL GAS SYSTEMS
Sensor 1
Sensor 2 Sensor 3
Sensor 4
Normal First warning
Second warning Fault detected
Figure 1. Fault prediction and detection by deep wavelet neural networks. However, the characteristics of sensor
data are complex, involving high velocities, huge volumes, dynamic values and types, real-time updating, and interdependency between different data sources.4
These make it difficult to extract
relevant features directly from raw data to interpret and produce a representable form that is deliverable. Through continuously acquiring sensor data to perform specific tasks, MGS can carry out complex processes by learning from data, rather than following the pre-programmed rules, allowing a more compelling and robust control architecture with optimised performance to be built on.5 Automatic learning also vastly
improves product quality by fault detection and condition monitoring that can introduce predictive maintenance systems into production processes, replacing visual inspections. Indeed, deep learning technique is the key to complex applications such as fault prediction and predictive maintenance, and can be applied to coupled multiple-input multiple-output (MIMO) systems of complex industrial processes. However, due to the strong nonlinearity and its nearly instantaneous response to disturbances, fault prediction with optimised performance in such a complex MIMO system is difficult to achieve. To raise the power and capabilities of
current machine learning systems, where commonly-used deep neural networks require improvements in nonlinear fitting and convergence speed, we propose a novel machine learning system designed by building IoT networks to remotely collect data, and developing deep wavelet neural networks using nonlinear Gaussian (Mexican hat) wavelet with sparsity as activation functions, to process the real- time data with features extracted in time- frequency domain for fault prediction and detection.
72
Feature extraction in time-frequency domain Wavelet analysis that reveals time and frequency variant characteristics of faulty signals, can be used to zoom into a local region of interest in time-frequency domain, for fast and accurate monitoring and detection of any anomaly/fault operating condition. To extract features, a correlation coefficient between each signal and the healthy signal is assigned to the signal using wavelet analysis. With the property of time-frequency localisation, wavelets can detect the transient components related to large changes in signals.6
For any given signal,
wavelet transform (WT) can decompose this signal into many functions through translations (time index) and dilations (frequency index) of a single function called a mother wavelet. A discrete wavelet transform (DWT) maps the time domain signal into a real- valued time-frequency domain and the signals are described by the wavelet coefficients as detail coefficients (high frequency part) and approximation coefficients (low frequency part), which represent the smoothed part of the signal. Wavelet analysis provides a mapping that can trade off between time and frequency resolution. It is effectively a mathematical microscope, which allows the user to zoom on features of interest at different scales and locations. However, the need for improvement of wavelets comes from a shortcoming that is inherent because of its construction. Second generation wavelets were named when the concept of lifting was introduced,7 and opened a new direction to construct wavelets which are not necessarily translates and dilates of one fixed function. A construction using lifting is entirely spatial, and therefore ideally suited for building second generation wavelets when Fourier techniques are no longer available.
Lifting wavelets are more general in the sense that all the classical wavelets can be generated by the lifting scheme. The lifting scheme makes optimal use of similarities between the high and low pass filters so as to achieve a faster implementation of WT. To give more flexibility to design
wavelet filter with linear phase for fast implementation, we propose to use biorthogonal wavelet transform by lifting scheme, where signal is partitioned into even and odd components that are then mutually predicted (to zero signal in high pass part) and updated (to retain in low pass part signal moments). After normalization, the algorithm is recursively applied to the low pass part. By lifting wavelet transform, signals can
be decomposed into a series of subband signals with the use of a multi-resolution analytical property, where features related to system status and fault prediction and detection are extracted in time-frequency domain, and can be used to train and test the deep learning model.
Machine learning system and deep learning Real-time machine learning system is designed using PCBs (printed circuit boards) as machine controllers, signal collectors and transmitters, and the IoT communication networks are built for remote monitoring, predicting failures or degradation, taking actions for highly efficient operation of MGS. With the IoT networks built, real-time
fault prediction and detection based on multi-sensor measurements can be performed by deep learning, so that we can have an eagle-eye view of every event while or before it is happening. It also enables real-time big data analytics to optimise decision making and performance, creating an opportunity to apply deep learning for advanced control,
IFHE DIGEST 2022
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
