Nuclear equipment pipeline leakage state monitoring method based on voiceprint feature analysis

Abstract

The invention belongs to the technical field of nuclear power plant equipment monitoring, aims at solving the problems that in the prior art, nuclear power plant high-temperature pipeline leakage monitoring sensitivity is insufficient, the false alarm rate is high, and early-stage tiny leakage detection is difficult to achieve, and discloses a nuclear equipment pipeline leakage state monitoring method based on voiceprint feature analysis. The method comprises the steps of preprocessing real-time data, designing a band-pass filter to enhance leakage features at feature frequencies and reduce noise, constructing multi-dimensional indexes, dynamically distributing multi-index weights, constructing a closed-loop online learning early warning model, and updating online learning early warning model parameters. Leakage in the nuclear power plant pipeline system is monitored and diagnosed in real time through the voiceprint recognition technology, and nuclear power plant pipeline health monitoring and fault early warning can be achieved.

Description

Technical Field

[0001] The present application belongs to the technical field of nuclear power plant equipment monitoring, and in particular relates to a method for monitoring the leakage status of nuclear equipment pipelines based on voiceprint feature analysis. Background Art

[0002] Traditional high-temperature pipeline leak detection relies primarily on manual inspections or offline testing methods, such as ultrasonic testing. These methods cannot achieve continuous, real-time monitoring. Monitoring methods based on process parameters such as pressure and temperature, or video images, lack sensitivity and typically only trigger alarms when leaks are severe.

[0003] In recent years, voiceprint recognition technology has been applied in the field of equipment status monitoring due to its non-contact and high sensitivity. However, the following problems still exist in high-temperature pipeline leak detection:

[0004] First, the pipeline operating environment is complex and the background noise interference is large, making it difficult for traditional voiceprint analysis methods to accurately extract leakage characteristics.

[0005] Second, existing methods mostly use fixed frequency band analysis and cannot adapt to different working conditions.

[0006] Third, the lack of an effective mechanism for dynamic adjustment of feature weights leads to a high false alarm rate. Summary of the Invention

[0007] The purpose of this application is to provide a nuclear equipment pipeline leakage status monitoring method based on voiceprint feature analysis to solve the problems of insufficient sensitivity, high false alarm rate and difficulty in early detection of small leaks in nuclear power plant high-temperature pipeline leakage monitoring in the existing technology.

[0008] In order to achieve the above objectives, this application provides the following technical solutions:

[0009] In a first aspect, the present application provides a method for monitoring the leakage status of nuclear equipment pipelines based on voiceprint feature analysis, comprising:

[0010] Step 1: Analyze historical data. If there is historical data, adaptively select the leakage characteristic frequency band by comparing the historical leakage data with the non-leakage data.

[0011] Step 2: Preprocess the real-time data, including setting up a fixed-size buffer for cyclic storage and designing a bandpass filter to enhance the leakage characteristics and reduce noise at the characteristic frequency;

[0012] Step 3: Construct multidimensional indicators, including spectral entropy, time domain peak-to-peak value, and frequency band energy ratio;

[0013] Step 4: Dynamically assign weights to multiple indicators and build a closed-loop online learning early warning model. This involves using the mean of historical data to initialize cluster centers, initialize weights for high-discrimination indicators, perform real-time classification alarms, label samples, and update the dataset.

[0014] Step 5: Update the online learning warning model parameters, including characteristic frequency bands, weights, and thresholds.

[0015] As an implementable approach, step 1 includes:

[0016] Step 1.1: Calculate the power spectral density of each sample to quantify the energy distribution of different frequency components;

[0017] Step 1.2: Compare the power spectral density of the leakage and normal data and calculate the significant difference in each frequency band;

[0018] Step 1.3: Select the frequency band with the largest difference as the feature frequency band.

[0019] As an implementable approach, in step 1.1, the power spectral density of each sample is as follows:

[0020]

[0021] Where P(f) is the power spectral density at frequency f; x(n) is the acquired signal; n is the discrete time index; N is the signal length; f is the frequency;

[0022] The power spectral density of a normal sample is:

[0023]

[0024] Where, P normal is the average power spectral density of normal samples; K is the number of normal samples for calculating the average power spectral density; P normal,i (f) is the power spectral density of the i-th normal sample at frequency f;

[0025] The power spectral density of the leakage sample is:

[0026]

[0027] Where, P leak is the average power spectral density of the leakage sample; K1 is the number of normal samples for calculating the average power spectral density; P leak,i (f) is the power spectral density of the i-th normal sample at frequency f;

[0028] As an implementable approach, the difference D(f) is calculated as follows:

[0029]

[0030] Where std(·) is the standard deviation;

[0031] The frequency band with the largest difference is selected as the characteristic frequency band [f low ,f high ],as follows:

[0032]

[0033] In the formula, [f low ,f high ] is the frequency band with the largest difference, where f low is the minimum frequency of the frequency band, f high is the maximum frequency of the band.

[0034] As an implementable approach, step 2 includes:

[0035] Step 2.1: For online continuous collection of high-temperature pipeline data, define a fixed-size buffer to cyclically store real-time data;

[0036] Step 2.2: Design a bandpass filter with a passband frequency that is a characteristic frequency, suppressing non-characteristic frequency band noise and enhancing leakage characteristics.

[0037] As an implementable approach, in step 2.2, the transfer function H(z) of the bandpass filter is:

[0038]

[0039] Where H(z) is the transfer function of the bandpass filter in the Z domain; h[n] is the unit impulse response of the filter; z is the complex variable in the Z transform; n is the discrete time index; L is the length of the unit impulse response of the filter; and H(f) is the frequency response of the bandpass filter.

[0040] As an implementable approach, in step 3, the spectrum entropy is:

[0041]

[0042] Where p i is the normalized energy of the i-th component of the signal in the frequency domain; M is the total number of frequency domain components;

[0043] The peak-to-peak value in the time domain is:

[0044] A pp =max(x(n))-min(x(n))

[0045] Where max(x(n)) is the maximum value of the signal x(n); min(x(n)) is the minimum value of the signal x(n);

[0046] The band energy ratio is:

[0047]

[0048] Where, E band is the leakage frequency band energy; E total is the total energy of the signal; P(f) is the power spectral density of the signal at frequency f.

[0049] As an implementable approach, step 4 includes:

[0050] Step 4.1: Use the feature means of historical normal and leaked data as the initial cluster centers;

[0051] Step 4.2: Calculate the discrimination of each indicator based on FisherScore and normalize it, giving higher discrimination indicators greater weight;

[0052] Step 4.3: Calculate the weighted score and set the initial alarm threshold based on the statistical distribution of historical samples to detect leaks;

[0053] Step 4.4: Data accumulation and labeling. If a leak is detected, the on-site personnel confirm and update the module. If a leak is confirmed, the data is marked as a leak sample. If it is a false alarm, it is marked as a normal sample. The updated data set is used for the next model parameter update.

[0054] As an implementable approach, in step 4.2, the normalized weights are:

[0055]

[0056] in,

[0057] Where, ω i is the normalized weight of the i-th feature; FisherScore i is the score of the i-th feature; j is the summation variable, which means summing the feature scores of j from 1 to 3; μ leak,i is the mean of the i-th feature in the leaked data; μ normal,i is the mean of the i-th feature in normal data; σ normal,i is the standard deviation of the i-th feature in normal data; σ leak,i is the standard deviation of the i-th feature in the leaked data.

[0058] As an implementable approach, step 4.3 includes:

[0059] Step 4.3.1: Calculate the weighted score as follows:

[0060]

[0061] Where, ω i is the normalized weight of the i-th feature; F i is the value of the i-th feature, where F1 = H (spectral entropy), F2 = A pp (peak-to-peak value in time domain), F3 = R (band energy ratio);

[0062] Step 4.3.2: Set the initial alarm threshold based on the statistical distribution of historical samples as follows:

[0063]

[0064] Where, T is the initial alarm threshold; N sample is the number of historical samples; X is the historical sample set; D(X) is the weighted score set of historical samples prctile(D(X),95) is the 95% quantile of D(X); max(D(X)) is the maximum value in D(X).

[0065] Step 4.3.3: Leak detection and alarm, as follows:

[0066]

[0067] Where Alarm is the detection result (True means a leak is detected and an alarm is triggered, False means no leak is detected); Score is the feature weighted score of the sample; and T is the alarm threshold.

[0068] As an implementable approach, step 5: Update the online learning warning model parameters, including characteristic frequency bands, weights, and thresholds, as follows:

[0069] Step 5.1: Update the characteristic frequency band: recalculate the power spectrum density difference between normal and leakage samples, and select the new maximum difference frequency band as the characteristic frequency band;

[0070] Step 5.2: Weight update: Recalculate FisherScore based on the new sample, and update the feature weight ω after normalization i ;

[0071] Step 5.3: Threshold update: Recalculate the threshold based on the weighted score distribution of the latest historical samples:

[0072]

[0073] Where, T new is the updated alarm threshold; N sample,new is the number of historical samples after update; X new is the updated historical sample set; D new (X new ) is the weighted score set of the updated historical samples; Nsamples,new The updated sample quantity.

[0074] Compared with the prior art, the nuclear equipment pipeline leakage state monitoring method based on voiceprint feature analysis provided by the application has the following beneficial effects:

[0075] The application uses voiceprint recognition technology to monitor and diagnose the leakage in the pipeline system of the nuclear power plant in real time, and can realize the health monitoring and fault early warning of the pipeline of the nuclear power plant.

[0076] By using the voiceprint recognition technology, the application can capture the acoustic characteristics generated by the pipeline leakage in real time, and combine the adaptive leakage feature frequency band selection, multi-index dynamic weight distribution method and closed-loop online learning warning model to realize the accurate monitoring and early warning of the high-temperature pipeline leakage state of the nuclear power plant. The method aims to improve the safety and reliability of the pipeline system of the nuclear power plant, reduce the potential risks caused by leakage, and provide technical support for the safe operation of the nuclear power plant.

[0077] The application is an intelligent voiceprint analysis technology that can automatically optimize the detection frequency band, dynamically adjust the feature weight and has continuous learning ability, which can improve the real-time performance and accuracy of high-temperature pipeline leakage detection. BRIEF DESCRIPTION OF DRAWINGS

[0078] In order to more clearly illustrate the technical solutions of the application, the following will briefly introduce the drawings needed to be used in the technical description.

[0079] Figure 1 The flowchart of the nuclear equipment pipeline leakage state monitoring method based on voiceprint feature analysis provided by the application;

[0080] Figure 2 The flowchart of the nuclear equipment pipeline leakage state monitoring method based on voiceprint feature analysis provided by the application; DETAILED DESCRIPTION

[0081] The following will be further described in detail through specific embodiments.

[0082] As shown in the Figure 1 The application provides a nuclear equipment pipeline leakage state monitoring method based on voiceprint feature analysis, which comprises:

[0083] Step 1: analyze historical data and select adaptive leakage feature frequency band.

[0084] If there is historical data, the leakage feature frequency band is adaptively selected by comparing the historical leakage data and the non-leakage data, which specifically includes the following:

[0085] Step 1.1: calculate the power spectral density. Calculate the power spectral density of each sample to quantify the energy distribution of different frequency components and provide a basis for frequency band selection.

[0086] Step 1.2: Calculate the difference. Compare the power spectral density of the leakage and normal data and calculate the significant difference in each frequency band.

[0087] Step 1.3: Select the frequency band with the largest difference as the characteristic frequency band. Focus on the leakage-sensitive frequency band to improve the signal-to-noise ratio.

[0088] Step 1 adopts an adaptive leakage characteristic frequency band selection method, which automatically selects leakage-sensitive frequency bands based on the difference analysis of historical or experimental high-temperature pipeline data, reduces human intervention, can continuously update historical data, adapt to different operating environments, and improve the sensitivity of leakage signals.

[0089] Step 2: Analyze real-time monitoring data and perform preprocessing (feature extraction), including:

[0090] Step 2.1: Data buffer setup. For online continuous data collection of high-temperature pipelines, define a fixed-size buffer to cyclically store real-time data. This ensures consistent data volume each time it is processed, meeting real-time requirements.

[0091] Step 2.2: Leakage feature enhancement and noise reduction: Design a bandpass filter with a passband frequency that is the characteristic frequency to suppress non-characteristic frequency band noise and enhance the leakage feature.

[0092] Step 3: Construct multi-dimensional indicators to evaluate pipeline leakage.

[0093] In high-temperature pipeline leakage scenarios, the characteristics of leakage signals are complex, so it is necessary to build more targeted multi-dimensional indicators, including:

[0094] (1) Spectral entropy, which measures the complexity of the signal in the frequency domain. Leakage signals usually lead to a decrease in spectral entropy.

[0095] (2) Peak-to-peak value in the time domain, capturing the transient impact amplitude changes caused by leakage;

[0096] (3) Frequency band energy ratio: Based on the characteristics of high-temperature pipeline leakage signals, the ratio of the energy of a specific frequency band to the total energy is calculated. Leakage signals usually cause a significant increase in the energy of certain frequency bands.

[0097] Step 4: Assign dynamic weights to multiple indicators and use a closed-loop online learning early warning model. The entire process from weight assignment to classification alarm to operator confirmation and parameter update is a closed-loop online learning model.

[0098] Step 4.1: Use the feature means of historical normal and leaky data as the initial cluster centers.

[0099] Step 4.2: Weight initialization. Calculate the discrimination of each indicator and normalize it, giving higher discrimination indicators greater weights to improve model sensitivity.

[0100] Step 4.3: Dynamic classification and alarm: First, calculate the weighted score, then set the initial alarm threshold based on the statistical distribution of historical samples, and finally implement leak detection and alarm.

[0101] Step 4.4: Data Accumulation and Labeling. If a leak is detected, wait for on-site personnel to confirm and update the module. If a leak is confirmed, the data is labeled as a leak sample; if it is a false alarm, it is labeled as a normal sample. The updated dataset is used for the next model parameter update.

[0102] Step 4 adopts a multi-indicator dynamic weight allocation method to construct a variety of leakage feature evaluation indicators, enhance the ability to capture leakage impact signals, and dynamically optimize feature weights to improve model robustness.

[0103] Step 5: Update the parameters of the online learning warning model. As data accumulates (including normal and operator-confirmed leakage data), update the feature frequency bands, weights, and thresholds.

[0104] This application adopts a closed-loop online learning early warning model, which starts from the closed-loop online learning process of "data accumulation-model update-dynamic early warning" and combines unsupervised classification with operator feedback to continuously update model parameters, thereby improving the robustness and accuracy of the method.

[0105] Example

[0106] like Figure 2 As shown, this embodiment is a specific implementation step of a method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis, including:

[0107] Step 1: Historical data analysis and adaptive selection of leakage characteristic frequency bands.

[0108] Step 1.1: Calculate the power spectral density. Calculate the power spectral density (PSD) of each sample as follows:

[0109]

[0110] Where P(f) is the power spectral density at frequency f; x(n) is the acquired signal; n is the discrete time index; N is the signal length; and f is the frequency.

[0111]

[0112] Where: P normal is the average power spectral density of normal samples; K is the number of normal samples for calculating the average power spectral density; P normal,i (f) is the power spectral density of the i-th normal sample at frequency f.

[0113] Step 1.2: Calculate the difference degree D(f) as follows:

[0114]

[0115] where std(·) is the standard deviation.

[0116] Step 1.3: Select the frequency band with the largest difference degree as the characteristic frequency band [f low ,f high ] as follows:

[0117]

[0118] where [f low ,f high ] is the selected frequency band with the largest difference degree, where f low is the minimum frequency of the frequency band, and f high is the maximum frequency of the frequency band.

[0119] Step 2: Real-time data analysis and feature extraction, including:

[0120] Step 2.1: Data buffer setting. For online continuous collection of high-pressure pipeline data, define a fixed-size buffer to store real-time data in a loop.

[0121] Step 2.2: Leakage feature enhancement and noise reduction. Design a band-pass filter with passband frequency at the characteristic frequency [f low ,f high ] to suppress non-characteristic frequency band noise and enhance the leakage feature. The transfer function H(z) of the band-pass filter can be expressed as:

[0122]

[0123] where H(z) is the transfer function of the band-pass filter in the Z domain; h[n] is the unit impulse response of the filter; z is a complex variable in Z transform; n is the discrete time index; L is the length of the filter unit impulse response; and H(f) is the frequency response of the band-pass filter.

[0124] Step 3: Multi-dimensional index construction and evaluation. Multi-dimensional indexes include:

[0125] (1) Spectral entropy, expressed as follows:

[0126]

[0127] where p i is the normalized energy of the i-th component of the signal in the frequency domain; and M is the total number of frequency domain components.

[0128] (2) Time-domain peak-to-peak value, expressed as follows:

[0129] A pp =max(x(n))-min(x(n))

[0130] Where max(x(n)) is the maximum value of the signal x(n); min(x(n)) is the minimum value of the signal x(n).

[0131] (3) Band energy ratio, which is expressed as follows:

[0132]

[0133] Where, E band is the leakage frequency band energy; E total is the total energy of the signal; P(f) is the power spectral density of the signal at frequency f.

[0134] Step 4: Assign dynamic weights to multiple indicators and adopt a closed-loop online learning early warning model.

[0135] Step 4.1: Weight initialization. Calculate the discrimination of each indicator and normalize it, giving higher discrimination indicators greater weights to improve model sensitivity, as follows:

[0136]

[0137] Where, ω i is the normalized weight of the i-th feature; FisherScore i is the score of the i-th feature; j is the summation variable, which means summing the feature scores of j from 1 to 3; μ leak,i is the mean of the i-th feature in the leaked data; μ normal,i is the mean of the i-th feature in normal data; σ normal,i is the standard deviation of the i-th feature in normal data; σ leak,i is the standard deviation of the i-th feature in the leaked data.

[0138] Step 4.2: Dynamic classification and alarm: First, calculate the weighted score, then set the initial alarm threshold based on the statistical distribution of historical samples, and finally implement leak detection and alarm.

[0139] Step 4.2.1: Calculate the weighted score as follows:

[0140]

[0141] Where, ω i is the normalized weight of the i-th feature; F i is the value of the i-th feature, where F1 = H (spectral entropy), F2 = A pp (peak-to-peak value in time domain), F3=R (band energy ratio).

[0142] Step 4.2.2: Set the initial alarm threshold based on the statistical distribution of historical samples as follows:

[0143]

[0144] Where, T is the initial alarm threshold; N sample is the number of historical samples; X is the historical sample set; D(X) is the weighted score set of historical samples prctile(D(X),95) is the 95% quantile of D(X); max(D(X)) is the maximum value in D(X).

[0145] Step 4.2.3: Leak detection and alarm, as follows:

[0146]

[0147] Where, T is the initial alarm threshold; N sample is the number of historical samples; X is the historical sample set; D(X) is the weighted score set of historical samples prctile(D(X),95) is the 95% quantile of D(X); max(D(X)) is the maximum value in D(X).

[0148] Step 4.3: Data Accumulation and Labeling. If a leak is detected, wait for on-site personnel to confirm and update the module. If a leak is confirmed, the data is labeled as a leak sample; if it is a false alarm, it is labeled as a normal sample. The updated dataset is used for the next model parameter update.

[0149] Step 5: Update the online learning warning model parameters. As data accumulates (including normal and operator-confirmed leakage data), update the following parameters:

[0150] Step 5.1: Update the characteristic frequency band: recalculate the power spectrum density difference between normal and leakage samples, and select the new maximum difference frequency band as the characteristic frequency band;

[0151] Step 5.2: Weight update: Recalculate FisherScore based on the new sample, and update the feature weight ω after normalization i ;

[0152] Step 5.3: Threshold update: Recalculate the threshold based on the weighted score distribution of the latest historical samples:

[0153]

[0154] Where, T new is the updated alarm threshold; N sample,new is the number of historical samples after update; X newis the updated historical sample set; D new (X new ) is the weighted score set of the updated historical samples; N samples,new is the updated sample size.

[0155] The above description is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed in this application should be covered by the scope of protection of the present application.

Claims

1. A method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis, characterized in that: include: Step 1: Analyze historical data. If there is historical data, adaptively select the leakage characteristic frequency band by comparing the historical leakage data with the non-leakage data. Step 2: Preprocess the real-time data and design a bandpass filter to enhance the leakage characteristics and reduce noise at the characteristic frequency; Step 3: Construct multidimensional indicators, including spectral entropy, time domain peak-to-peak value, and frequency band energy ratio; Step 4: Dynamically assign weights to multiple indicators, initialize the weights of high-discrimination indicators, perform real-time classification alarms, mark samples, and update the dataset; Step 5: Update the online learning warning model parameters, Includes characteristic frequency bands, weights and thresholds.

2. The method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis according to claim 1 is characterized in that: Step 1 includes: Step 1.1: Calculate the power spectral density of each sample to quantify the energy distribution of different frequency components; Step 1.2: Compare the power spectral density of the leakage and normal data and calculate the significant difference in each frequency band; Step 1.3: Select the frequency band with the largest difference as the feature frequency band.

3. The method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis according to claim 2 is characterized in that: Calculate the difference D(f) as follows: Where std(·) is the standard deviation; The frequency band with the largest difference is selected as the characteristic frequency band [f low ,f high ],as follows: In the formula, [f low ,f high ] is the frequency band with the largest difference, where f low is the minimum frequency of the frequency band, f high is the maximum frequency of the band.

4. The method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis according to claim 1 is characterized in that: Step 2 includes: Step 2.1: For online continuous collection of high-temperature pipeline data, define a fixed-size buffer to cyclically store real-time data; Step 2.2: Design a bandpass filter with a passband frequency that is a characteristic frequency, suppressing non-characteristic frequency band noise and enhancing leakage characteristics.

5. The method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis according to claim 4 is characterized in that: In step 2.2, the transfer function H(z) of the bandpass filter is: Where H(z) is the transfer function of the bandpass filter in the Z domain; h[n] is the unit impulse response of the filter; z is the complex variable in the Z transform; n is the discrete time index; L is the length of the unit impulse response of the filter; and H(f) is the frequency response of the bandpass filter.

6. The method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis according to claim 1 is characterized in that: In step 3, the spectral entropy is: Where p i is the normalized energy of the i-th component of the signal in the frequency domain; M is the total number of frequency domain components; The peak-to-peak value in the time domain is: A pp =max(x(n))-min(x(n)) Where max(x(n)) is the maximum value of the signal x(n); min(x(n)) is the minimum value of the signal x(n); The band energy ratio is: Where, E band is the leakage frequency band energy; E total is the total energy of the signal; P(f) is the power spectral density of the signal at frequency f.

7. The method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis according to claim 1 is characterized in that: Step 4 includes: Step 4.1: Use the feature means of historical normal and leaked data as the initial cluster centers; Step 4.2: Calculate the discrimination of each indicator based on FisherScore and normalize it, giving higher discrimination indicators greater weight; Step 4.3: Calculate the weighted score and set the initial alarm threshold based on the statistical distribution of historical samples to detect leaks; Step 4.4: Data accumulation and labeling. If a leak is detected, confirm and update the module. If a leak is confirmed, mark the data as a leak sample. If it is a false alarm, mark it as a normal sample. The updated data set is used for the next model parameter update.

8. The method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis according to claim 7 is characterized in that: In step 4.2, the normalized weight ω i for: in, Where, ω i is the normalized weight of the i-th feature; FisherScore i is the score of the i-th feature; j is the summation variable, which means summing the feature scores of j from 1 to 3; μ leak,i is the mean of the i-th feature in the leaked data; μ normal,i is the mean of the i-th feature in normal data; σ normal,i is the standard deviation of the i-th feature in normal data; σ leak,i is the standard deviation of the i-th feature in the leaked data.

9. The method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis according to claim 7 is characterized in that: Step 4.3 includes: Step 4.3.1: Calculate the weighted score as follows: Where, ω i is the normalized weight of the i-th feature; F i is the value of the i-th feature; Step 4.3.2: Set the initial alarm threshold based on the statistical distribution of historical samples as follows: Where, T is the initial alarm threshold; N sample is the number of historical samples; X is the historical sample set; D(X) is the feature weighted score set of historical samples; prctile(D(X),95) is the 95% quantile of D(X); Step 4.3.3: Leak detection and alarm, as follows: Where Alarm is the detection result; Score is the feature weighted score of the sample; T is the alarm threshold.

10. The method for monitoring nuclear equipment pipeline leakage status based on voiceprint feature analysis according to claim 8, characterized in that: Step 5 includes: Step 5.1: Recalculate the power spectrum density difference between normal and leaky samples, and select the new maximum difference frequency band as the feature frequency band; Step 5.2: Recalculate FisherScore based on the new sample, normalize and update the feature weight ω i ; Step 5.3: Recalculate the threshold based on the weighted score distribution of the latest historical samples: Where, T new is the updated alarm threshold; N sample,new is the number of historical samples after update; X new is the updated historical sample set; D new (X new ) is the weighted score set of the updated historical samples; N samples,new is the updated sample size.

Images