TWI844374B - Feature extraction of metal impact vibration signals and mechanical-learning identification method thereof - Google Patents

Abstract

The present invention extracts features of vibration signals of metal impact. Furthermore, a mechanical-learning method for identifying the features is applied by acquiring the time-domain waveform data of the vibration signals. At first, the spectrum distribution at the instant moment of impact is locked for time-frequency analysis. Then, filtering and spectral feature extraction are performed by the self-developed equal-bandwidth spectrum-energy-ratio method. Finally, the modified mechanical-learning method is used to discriminate real and fake metal impact signals. Hence, the present invention identifies metal impact vibration signals, and diagnoses the status of a mechanical equipment without stopping it, so as to improve its mechanical operation stability, and to avoid malfunctions or accidents that may cause unintended power outages.

Description

Feature extraction and mechanical learning identification method of metal impact vibration signal

The present invention relates to a feature extraction and mechanical learning identification method of metal impact vibration signals, especially to a non-invasive vibration signal diagnosis technology for power plant equipment impacted by foreign objects, and in particular to a technology that can be applied to the identification of foreign object impacts on various metal material machines.

The loose component monitoring system is one of the important instrumentation systems in nuclear power plants. It is used to detect whether there are loose components in the reactor coolant system (RCS) and provide alarms, signal displays and data to power plant operators in real time, so that power plant operators can take appropriate measures to deal with the threats caused by loose components. In the international experience of nuclear power plant operation, nuts, bolts, pins, pipe components and hand tools used for maintenance have been found in the reactor coolant system. Such components may damage steam generator tubes, reactor internal components and reactor coolant pumps, and require at least tens of millions of NT dollars to repair. Loose parts can also jam control rods or block passages within flow reactors or steam generators, causing plant operation problems and posing safety hazards.

In this regard, most existing knowledge technologies require complex processing operations on vibration signals, or require feature analysis through spectrum image data, and use existing machine learning methods for pattern recognition, without any improvement or innovation in vibration signal analysis and machine learning computing technology. In addition, although another knowledge technology proposes a simple method for extracting vibration signal spectrum features, the result of directly capturing time domain signals will cause Gibbs phenomenon in the spectrum, interfering with high-frequency signals, and it does not propose a pattern recognition computing method.

Although the use of vibration signals for non-destructive fault diagnosis of mechanical equipment has been widely studied and applied, the way in which mechanical equipment generates vibration varies greatly, such as movement, rotation, or impact by external objects, and the characteristics of the corresponding vibration signals are also very different. In addition, considering that many countries have listed nuclear power plants as environmentally friendly power generation equipment and have restarted the construction of nuclear power plants, the safety of nuclear power plants does not allow any mistakes. Therefore, in view of the shortcomings and defects in the knowledge and technology, there is an urgent need to improve them. It is necessary to improve the existing shortcomings and develop an invention that can be applied to other metal material machinery to accurately identify foreign body impacts to ensure the safe and stable operation of nuclear power plants.

The main purpose of the present invention is to overcome the above problems encountered by the prior art and to provide a general-purpose method for feature extraction and mechanical learning identification of metal impact vibration signals that can quickly and correctly determine whether it is a true metal impact event and does not cause excessive burden of program calculations due to the simplicity of the method.

Another purpose of the present invention is to provide a method for assisting domestic nuclear power plants in distinguishing true and false impact signals of metal loose components detection systems. In addition to being able to correctly distinguish true and false impact signals in a short time to ensure the safe and stable operation of nuclear power plants, it can also be applied to the feature extraction and mechanical learning identification method of metal impact vibration signals for abnormal detection of foreign objects intrusion in other mechanical equipment.

To achieve the above purpose, the present invention is a method for extracting and identifying the characteristics of metal impact vibration signals by mechanical learning, which at least includes the following steps: Step 1: Measure and receive the vibration signal using a vibration sensing element, and process the time domain waveform of the vibration signal into time-frequency domain spectrum information having both time domain and frequency domain by time-frequency analysis conversion, and extract the spectrum at the moment of impact; Step 2: The spectrum uses the equal-bandwidth spectrum energy ratio method to simultaneously perform filtering and spectrum feature extraction to obtain spectrum feature coefficients that can distinguish high and low frequency bands; and step 3: the spectrum feature coefficients are used to distinguish true and false metal impact signals using the improved machine learning method and multi-dimensional space distance search algorithm; if it is determined to be a true impact signal, the warning light will be turned on to notify personnel to handle it, if not, continue monitoring.

In the above-mentioned embodiments of the present invention, the time-frequency analysis transform is a short-time Fourier transform (STFT), a Hilbert-Huang transform, or a wavelet transform.

In the above-mentioned embodiment of the present invention, the horizontal axis of the time-frequency domain spectrum is frequency, and the vertical axis is time. The time segment with the largest energy in the time-frequency domain spectrum is the spectrum at the moment of impact.

In the above-mentioned embodiments of the present invention, the broadband spectrum energy ratio method is selected from a self-developed spectrum feature extraction method.

Please refer to "Figures 1 to 4", which are respectively the overall process diagram of the present invention, the spectrum diagram of the present invention using time-frequency analysis conversion to capture the impact moment, the schematic diagram of the present invention's equal-bandwidth spectrum energy ratio method, and the process diagram of the present invention's improved multi-dimensional space distance mechanical learning. As shown in the figure: The present invention is a feature extraction and mechanical learning identification method for metal impact vibration signals, which at least includes the following steps: Step s1: Use a vibration sensing element to measure and receive a vibration signal, and process the time domain waveform of the vibration signal into time-frequency domain spectrum information with both time domain and frequency domain by time-frequency analysis conversion, and capture the spectrum of the impact moment. Step s2: The spectrum at the moment of the impact is filtered using the equal-bandwidth spectrum energy ratio method and spectrum feature extraction to obtain spectrum feature coefficients that can distinguish high and low frequency bands. Step s3: The spectrum feature coefficients are used to distinguish true and false metal impact signals using the improved machine learning method and multi-dimensional space distance search algorithm; if it is determined to be a true impact signal, the warning light is turned on to notify the personnel to handle it, if not, the monitoring continues. In this way, a new feature extraction and mechanical learning identification method of metal impact vibration signals is constructed through the above-disclosed process.

In a preferred embodiment of the present invention, the time-frequency analysis transform is a short-time Fourier transform (STFT), a Hilbert-Huang transform, or a wavelet transform.

In a preferred specific embodiment of the present invention, when the sensing element receives a vibration signal, its time domain waveform is shown in Figure 2 (a). After time-frequency analysis and conversion, a time-frequency domain spectrum as shown in Figure 2 (b) can be obtained, wherein the horizontal axis of the time-frequency domain spectrum is frequency and the vertical axis is time. The energy of the time segment of the black solid line frame in the spectrum is the largest, which is the spectrum at the moment of the impact, as shown in Figure 2 (c).

In a preferred embodiment of the present invention, the bandwidth spectrum energy ratio method is selected from a self-developed spectrum feature extraction method. The implementation method is to divide the spectrum of the impact moment of Figure 2 (c) into N intervals, as shown in Figure 3, and the recommended N value is ²ⁿ , where n is a positive integer; the bandwidth of each interval is Δf, and the area between each interval curve and the horizontal axis is calculated respectively. , then the ratio of each interval area to the total area , which can be used as the characteristic coefficient of the spectrum, as shown in formula (1): (1)

In a preferred embodiment of the present invention, the detailed calculation process of the improved machine learning method using the multi-dimensional space distance search algorithm is shown in FIG. 4, which at least includes the following steps: Step s31: Spectral characteristic coefficients of the vibration signal to be determined , and the spectral eigenvalues of all sample points , perform Euclidean distance in N-dimensional space Calculate, as shown in formula (2): (2) Step s32: From all the sample points, select the first k points with the smallest Euclidean distance as the judgment sample points. Step s33: Calculate the inverse of the distance of the selected k judgment sample points as the weight value of the category judgment , as shown in formula (3): (3) Step s34: Calculate the category values of the k judgment sample points The weighted average of An integer of 0 or 1, 0 represents a false collision signal, and 1 represents a true collision signal. The expected category value of the final output is a value between 0 and 1. If it is less than 0.5, it is determined to be a false impact signal, and if it is greater than 0.5, it is determined to be a true impact signal, as shown in formula (4): (4)

The present invention discloses a universal feature extraction and mechanical learning identification method for metal impact vibration signals, the purpose of which is to quickly and correctly determine whether it is a real metal impact event, and because the method is simple, it will not cause excessive burden on program operations. When used, after capturing the time domain waveform data of the vibration signal, the spectrum distribution of the impact moment is first locked by time-frequency analysis, and then the self-developed equal-bandwidth spectrum energy ratio method is used to simultaneously perform filtering and spectrum feature extraction, and finally the improved mechanical learning method is used to distinguish true and false metal impact signals. Therefore, the present invention is used to identify metal impact vibration signals, and can perform status diagnosis of mechanical equipment without stopping the machine, thereby improving the operating stability of the mechanical equipment to avoid unexpected power outages caused by failures or accidents.

From the above, it can be seen that the method proposed by the present invention is a patented technology related to non-invasive vibration signal diagnosis of foreign body impact on power plant equipment, and can be applied to foreign body impact identification of various metal material machinery.

In order to assist domestic nuclear power plants in distinguishing true and false impact signals of metal loose element detection systems, the present invention has developed a feature extraction and mechanical learning identification method for metal impact vibration signals. In addition to being able to correctly distinguish true and false impact signals in a short period of time, it can also be applied to abnormal detection of other mechanical equipment being invaded by foreign objects. In this way, the present invention can quickly and correctly distinguish true and false metal loose element impact signals with simple calculations, ensuring the safe and stable operation of nuclear power plants. Therefore, it has the following characteristics: 1. It only uses ²ⁿ numerical data, and the program operation is simple, fast and light, and will not cause a burden on the hardware equipment; 2. It uses the self-developed equal-bandwidth spectrum energy ratio method to perform filtering and spectrum feature extraction at the same time; and 3. It is combined with the self-improved machine learning method to accurately distinguish true and false metal impact signals with a high accuracy rate.

In summary, the present invention is a method for extracting features from metal impact vibration signals and mechanical identification, which can effectively improve various shortcomings of the existing methods. In addition to accurately identifying whether it is a foreign body impact signal to ensure the safe and stable operation of nuclear power plants, it can also be applied to the identification of foreign body impacts of other metal material machinery, thereby making the present invention more advanced, more practical, and more in line with the needs of users. It has indeed met the requirements for invention patent applications, and a patent application is filed in accordance with the law.

However, the above is only a preferred embodiment of the present invention and should not be used to limit the scope of implementation of the present invention; therefore, any simple equivalent changes and modifications made according to the scope of the patent application of the present invention and the content of the invention specification should still fall within the scope of the present invention patent.

s1～s3: Steps

s31～s34: Steps

Figure 1 is a schematic diagram of the overall process of the present invention. Figure 2 is a spectrum diagram of the instant of impact captured by time-frequency analysis conversion. Figure 3 is a schematic diagram of the equal-bandwidth spectrum energy ratio method of the present invention. Figure 4 is a schematic diagram of the process of the improved multi-dimensional space distance mechanical learning of the present invention.

s1~s3: Steps

Claims (6)

A method for feature extraction and mechanical learning identification of metal impact vibration signals comprises at least the following steps: Step 1: using a vibration sensing element to measure and receive a vibration signal, and by time-frequency analysis and conversion, the time domain waveform of the vibration signal is processed into time-frequency domain spectrum information having both time domain and frequency domain, from which the spectrum at the moment of impact is extracted; Step 2: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 3: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 4: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 5: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 6: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 7: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 8: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 9: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 10: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 11: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 12: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 13: using equal bandwidth to convert the spectrum at the moment of impact into a time-frequency domain spectrum; Step 14: Spectral energy ratio method, filtering and spectral feature extraction are performed simultaneously to obtain spectral feature coefficients that can distinguish high and low frequency bands; and step 3: the spectral feature coefficients are used to distinguish true and false metal impact signals using improved machine learning method and multi-dimensional space distance search algorithm; if it is determined to be a true impact signal, the warning light will be turned on to notify personnel to handle it, if not, continue monitoring. According to the feature extraction and mechanical learning identification method of metal impact vibration signal described in item 1 of the patent application scope, the time-frequency analysis conversion is short-time Fourier transform (STFT), Hilbert-Huang transform, or wavelet transform. According to the feature extraction and mechanical learning identification method of metal impact vibration signal described in Item 1 of the patent application scope, the horizontal axis of the time-frequency domain spectrum is frequency, and the vertical axis is time. The time segment with the largest energy in the time-frequency domain spectrum is the spectrum at the moment of the impact. According to the feature extraction and mechanical learning identification method of metal impact vibration signal described in Item 1 of the patent application scope, the broadband spectrum energy ratio method is selected from a spectrum feature extraction method. According to the feature extraction and mechanical learning identification method of metal impact vibration signal described in item 1 of the patent application scope, the bandwidth spectrum energy ratio method is to divide the spectrum at the moment of impact into N intervals, the bandwidth of each interval is △f, and the area A _i between each interval curve and the horizontal axis is calculated respectively. Then, the ratio σ _i of each interval area to the total area can be used as the spectrum characteristic coefficient, and its formula is:

Here, the value of N is 2 ⁿ , and n is a positive integer. According to the feature extraction and mechanical learning identification method of metal impact vibration signal described in item 1 of the patent application scope, the improved mechanical learning method uses the calculation process of the multi-dimensional space distance search algorithm, which at least includes the following steps: Step 3.1: Perform Euclidean distance calculation in N-dimensional space on the spectral feature coefficient x _j of the vibration signal to be judged and the spectral feature coefficient yi _,j of all sample points.

Calculation, the formula is:

Step 3.2: From all the sample points, select the first k points with the smallest Euclidean distance as the judgment sample points; Step 3.3: Calculate the inverse of the distance of the selected k judgment sample points as the weight value w _i of the category judgment, which is expressed as:

; and step 3.4: calculate the weighted average of the category values θ _i of the k judgment sample points, where θ _i is an integer of 0 or 1, 0 represents a false impact signal, and 1 represents a true impact signal; the expected category value θ _ex outputted at the end is a value between 0 and 1, and a value less than 0.5 is determined to be a false impact signal, and a value greater than 0.5 is determined to be a true impact signal, and the formula is expressed as follows: