Academia.eduAcademia.edu

Figure 15 – uploaded by Paweł Pławiak

See full PDFdownloadDownload figure
Figure 14: Results of the quantitative analysis of neural networks from Table - accuracy-complexity criterion. A higher percentage score means a higher rating - a better result.

Figure 14 Results of the quantitative analysis of neural networks from Table - accuracy-complexity criterion. A higher percentage score means a higher rating - a better result.

Related Figures (19)

www-.elsevier.com/locate/neucom $0925-2312(14)00611-0 http://dx.doi.org/10.1016/j.neucom.2014.04.026 NEUCOM 14183
www-.elsevier.com/locate/neucom $0925-2312(14)00611-0 http://dx.doi.org/10.1016/j.neucom.2014.04.026 NEUCOM 14183
Figure 1: Pump WPTO2-10 (cross-section); | - lid, 2 - bearings, 3 - body, 4 - side dish, 5 - outer race of the bearing rotor, 6 - shaft, 7 - rotor, 8 - pistons, 9 - separator, 10 - lid 11 - shield with a constant inclination angle, 12 - shield of timing gear [18].
Figure 1: Pump WPTO2-10 (cross-section); | - lid, 2 - bearings, 3 - body, 4 - side dish, 5 - outer race of the bearing rotor, 6 - shaft, 7 - rotor, 8 - pistons, 9 - separator, 10 - lid 11 - shield with a constant inclination angle, 12 - shield of timing gear [18].
Figure 2: Schema of axial piston pump with hinged shield; 1 - piston, 2 - shaft and rotor, 3 - shield 4 - distributor plate, D - distance between the pistons, d - diameter of the piston, h - the maximum displacement of the piston [18].
Figure 2: Schema of axial piston pump with hinged shield; 1 - piston, 2 - shaft and rotor, 3 - shield 4 - distributor plate, D - distance between the pistons, d - diameter of the piston, h - the maximum displacement of the piston [18].
Table 3: Successive preprocessing stages of measured data, for all pump consumptions.
Table 3: Successive preprocessing stages of measured data, for all pump consumptions.
Figure 3: The visualization of preprocessing stages for the dynamic pressure in the time (A,B,C,D,E,F) and frequency (A,B,G,H,LJ) domain. Each color corresponds to specific consumptions of the pump. A — raw data; B— example fragment of signal after removing at the beginning and at the end thirty samples and the constant component; C — signal after standardization; D — piece of signal generated by dividing the whole signal (for one consumption) into 25 fragments consisting of 2616 samples and imposing the resulting fragments on each other; E — piece of the signal taken for the final analysis; F — signal after normalization; G - signal after normalization; H — piece of signal generated by dividing the whole signal (for one consumption) into 25 fragments consisting of 7128 samples and performing fast Fourier transform (with a Hamming window with a length of 256 samples) and imposing the resulting fragments on each other; I — piece of signal taken for final analysis; J — signal after normalization.
Figure 3: The visualization of preprocessing stages for the dynamic pressure in the time (A,B,C,D,E,F) and frequency (A,B,G,H,LJ) domain. Each color corresponds to specific consumptions of the pump. A — raw data; B— example fragment of signal after removing at the beginning and at the end thirty samples and the constant component; C — signal after standardization; D — piece of signal generated by dividing the whole signal (for one consumption) into 25 fragments consisting of 2616 samples and imposing the resulting fragments on each other; E — piece of the signal taken for the final analysis; F — signal after normalization; G - signal after normalization; H — piece of signal generated by dividing the whole signal (for one consumption) into 25 fragments consisting of 7128 samples and performing fast Fourier transform (with a Hamming window with a length of 256 samples) and imposing the resulting fragments on each other; I — piece of signal taken for final analysis; J — signal after normalization.
Figure 4: The visualization of preprocessing stages for the vibrations in the time (A,B,C,D,E) and frequency (A,B,F,G,H) domain. Each color corresponds tc specific consumptions of the pump. A — raw data; B-— example fragment of signal after removing at the beginning and at the end thirty samples and the constan component; C — piece of of signal generated by dividing the whole signal (for one consumption) into 25 fragments consisting of 2616 samples and imposing the resulting fragments on each other; D — piece of signal taken for final analysis; E — signal after normalization; F — piece of signal generated by dividing the whole signal (for one consumption) into 25 fragments consisting of 7128 samples and performing fast Fourier transform (with a Hamming window with a length of 128 samples) and imposing the resulting fragments to each other; G — piece of signal taken for final analysis; H — signal after normalization.
Figure 4: The visualization of preprocessing stages for the vibrations in the time (A,B,C,D,E) and frequency (A,B,F,G,H) domain. Each color corresponds tc specific consumptions of the pump. A — raw data; B-— example fragment of signal after removing at the beginning and at the end thirty samples and the constan component; C — piece of of signal generated by dividing the whole signal (for one consumption) into 25 fragments consisting of 2616 samples and imposing the resulting fragments on each other; D — piece of signal taken for final analysis; E — signal after normalization; F — piece of signal generated by dividing the whole signal (for one consumption) into 25 fragments consisting of 7128 samples and performing fast Fourier transform (with a Hamming window with a length of 128 samples) and imposing the resulting fragments to each other; G — piece of signal taken for final analysis; H — signal after normalization.
Figure 5: Comparison of the separation of states consumption of the pump, on the basis of application parameters, used for pre-processing the signal, for th dynamic pressure in the time domain. Considered parameters are: pre-processing method and length of the cut fragment: A — without constant component 2616 samples; B — without constant component, 3840 samples; C — without constant component, 7128 samples; D — without constant component and afte normalization, 2616 samples; E — without constant component and after normalization, 3840 samples; F — without constant component and after normalization 7128 samples G — without constant component and after standardization, 2616 samples H — without constant component and after standardization, 3840 sample: I — without constant component and after standardization, 7128 samples.
Figure 5: Comparison of the separation of states consumption of the pump, on the basis of application parameters, used for pre-processing the signal, for th dynamic pressure in the time domain. Considered parameters are: pre-processing method and length of the cut fragment: A — without constant component 2616 samples; B — without constant component, 3840 samples; C — without constant component, 7128 samples; D — without constant component and afte normalization, 2616 samples; E — without constant component and after normalization, 3840 samples; F — without constant component and after normalization 7128 samples G — without constant component and after standardization, 2616 samples H — without constant component and after standardization, 3840 sample: I — without constant component and after standardization, 7128 samples.
Figure 6: Comparison of the separation of states consumption of the pump, on the basis of application parameters, used for pre-processing the signal, for the vibrations in the time domain. Considered parameters are: pre-processing method and length of the cut fragment: A — without constant component, 2616 samples; B — without constant component, 3840 samples; C — without constant component, 7128 samples; D — without constant component and after normalization, 2616 samples; E — without constant component and after normalization, 3840 samples; F — without constant component and after normalization, 7128 samples G — without constant component and after standardization, 2616 samples H — without constant component and after standardization, 3840 samples I — without constant component and after standardization, 7128 samples.
Figure 6: Comparison of the separation of states consumption of the pump, on the basis of application parameters, used for pre-processing the signal, for the vibrations in the time domain. Considered parameters are: pre-processing method and length of the cut fragment: A — without constant component, 2616 samples; B — without constant component, 3840 samples; C — without constant component, 7128 samples; D — without constant component and after normalization, 2616 samples; E — without constant component and after normalization, 3840 samples; F — without constant component and after normalization, 7128 samples G — without constant component and after standardization, 2616 samples H — without constant component and after standardization, 3840 samples I — without constant component and after standardization, 7128 samples.
Figure 7: Comparison of the separation of states consumption of the pump, on the basis of application parameters, used for pre-processing the signal, for th dynamic pressure in the frequency domain, for the signal without constant component and after normalization. Considered parameters are: length of th cut fragment and Hamming window length: A — 2616 samples, 512 samples; B — 3840 samples, 512 samples; C — 7128 samples, 512 samples; D — 261 samples, 256 samples; E — 3840 samples, 256 samples; F — 7128 samples, 256 samples G = 2616 samples, 128 samples H — 3840 samples, 128 samples I- 7128 samples, 128 samples.
Figure 7: Comparison of the separation of states consumption of the pump, on the basis of application parameters, used for pre-processing the signal, for th dynamic pressure in the frequency domain, for the signal without constant component and after normalization. Considered parameters are: length of th cut fragment and Hamming window length: A — 2616 samples, 512 samples; B — 3840 samples, 512 samples; C — 7128 samples, 512 samples; D — 261 samples, 256 samples; E — 3840 samples, 256 samples; F — 7128 samples, 256 samples G = 2616 samples, 128 samples H — 3840 samples, 128 samples I- 7128 samples, 128 samples.
Figure 8: Comparison of the separation of states consumption of the pump, on the basis of application parameters, used for pre-processing the signal, for the vibrations in the frequency domain, for the signal without constant component. Considered parameters are: length of the cut fragment and Hamming window length: A — 2616 samples, 512 samples; B— 3840 samples, 512 samples; C— 7128 samples, 512 samples; D- 2616 samples, 256 samples; E — 3840 samples, 256 samples; F — 7128 samples, 256 samples G — 2616 samples, 128 samples H — 3840 samples, 128 samples I-—7128 samples, 128 samples.
Figure 8: Comparison of the separation of states consumption of the pump, on the basis of application parameters, used for pre-processing the signal, for the vibrations in the frequency domain, for the signal without constant component. Considered parameters are: length of the cut fragment and Hamming window length: A — 2616 samples, 512 samples; B— 3840 samples, 512 samples; C— 7128 samples, 512 samples; D- 2616 samples, 256 samples; E — 3840 samples, 256 samples; F — 7128 samples, 256 samples G — 2616 samples, 128 samples H — 3840 samples, 128 samples I-—7128 samples, 128 samples.
Table 4: The parameters of the neural networks.
Table 4: The parameters of the neural networks.
Figure 11: Example perfect matching of the MLP neural network response to the response model for a test set, for frequency domain for dynamic pres- sure.
Figure 11: Example perfect matching of the MLP neural network response to the response model for a test set, for frequency domain for dynamic pres- sure.
Figure 9: Matching of the GRNN response to the response model for a test set, for time domain for dynamic pressure.
Figure 9: Matching of the GRNN response to the response model for a test set, for time domain for dynamic pressure.
Figure 12: Example perfect matching of the MLP neural network response to the response model for a test set, for frequency domain for vibrations. Figure 10: Matching of the GRNN response to the response model for a test set, for time domain for vibrations.
Figure 12: Example perfect matching of the MLP neural network response to the response model for a test set, for frequency domain for vibrations. Figure 10: Matching of the GRNN response to the response model for a test set, for time domain for vibrations.
The results of quantitative analysis of neural networks Dynamic pressure - time domain
The results of quantitative analysis of neural networks Dynamic pressure - time domain
of the pump on the basis of vibrations than in the dynamic pres- sure, as evidenced by a lower average percentage error for all networks. This is in line with expectations, reached on the ba- sis of the visualization of input signal waveforms for dynamic pressure and vibration. Figure 14 shows the results of neural networks which have obtained in criterion accuracy-complexity (higher percentage score means a better result). On this basis, we can conclude that all the networks achieved better results for the frequency do- main - which results from a smaller error (Figure 13). With re-
of the pump on the basis of vibrations than in the dynamic pres- sure, as evidenced by a lower average percentage error for all networks. This is in line with expectations, reached on the ba- sis of the visualization of input signal waveforms for dynamic pressure and vibration. Figure 14 shows the results of neural networks which have obtained in criterion accuracy-complexity (higher percentage score means a better result). On this basis, we can conclude that all the networks achieved better results for the frequency do- main - which results from a smaller error (Figure 13). With re-
Figure 13: Results of the quantitative analysis of the neural networks from Table - average error. A higher percentage score means a greater average error - worse result.
Figure 13: Results of the quantitative analysis of the neural networks from Table - average error. A higher percentage score means a greater average error - worse result.
Related topics:
Signal ProcessingPositive Displacement PumpComputational Intelligence Methods
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved