CN-121993421-A - Vibration anomaly detection method and system for nuclear power cold source pump

CN121993421ACN 121993421 ACN121993421 ACN 121993421ACN-121993421-A

Abstract

The invention provides a method and a system for detecting vibration abnormality of a nuclear power cold source pump, comprising the steps of obtaining an original vibration signal; in the start-stop stage of the cold source pump, an original vibration signal is processed through a Savitzky-Golay smooth function and dynamic polynomial regression to obtain a trend predicted value, a fitting residual sequence is calculated according to the trend predicted value and the original vibration signal, a dynamic standard deviation and a corresponding self-adaptive confidence interval are calculated, anomaly detection is carried out according to the self-adaptive confidence interval, after the cold source pump finishes the start-stop process and enters a relatively stable operation stage, the predicted value of a time sequence characteristic, the predicted value of a multi-source cooperative characteristic and a theoretical predicted value are subjected to dynamic weighted fusion to obtain a final predicted value, a residual error is calculated according to the original vibration signal and the final predicted value, anomaly detection is carried out according to a dynamic threshold interval to obtain an anomaly judgment result, and real-time anomaly identification and intelligent early warning of the cold source pump in the start-stop and stable operation processes are realized.

Inventors

Lang Chaohao
WU JIANMING

Assignees

浙江远算科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. The method for detecting the vibration abnormality of the nuclear power cold source pump is characterized by comprising the following steps of: Acquiring an original vibration signal; in the start-stop stage of the cold source pump, the original vibration signal is processed through a Savitzky-Golay smooth function and dynamic polynomial regression to obtain a trend prediction value; calculating a fitting residual sequence according to the trend predicted value and the original vibration signal; calculating a dynamic standard deviation and a corresponding self-adaptive confidence interval by adopting a dynamic estimation method combining a sliding window and exponential weighting for the fitting residual sequence; Performing anomaly detection according to the self-adaptive confidence interval; When the cold source pump finishes the start-stop process and enters a relatively stable operation stage, the original vibration signal is respectively output by an LSTM time sequence prediction model, a gradient lifting tree model and a physical twin model to obtain a predicted value of a time sequence characteristic, a predicted value of a multi-source cooperative characteristic and a theoretical predicted value; Dynamically weighting and fusing the predicted value of the time sequence feature, the predicted value of the multi-source cooperative feature and the theoretical predicted value to obtain a final predicted value; calculating residual errors according to the original vibration signals and the final predicted values, and constructing a dynamic threshold interval; and carrying out abnormality detection according to the dynamic threshold interval to obtain an abnormality judgment result.
2. The method for detecting vibration anomalies of a nuclear power cold source pump according to claim 1, wherein processing the original vibration signal through a Savitzky-Golay smoothing function and dynamic polynomial regression to obtain a trend prediction value comprises: denoising and trend maintaining the original vibration signal by adopting the Savitzky-Golay smoothing function to obtain a smoothed signal; and fitting the smoothed signal by adopting the dynamic polynomial regression to obtain the trend predicted value, wherein the fitting order of the dynamic polynomial regression is determined by stepwise regression and Bayesian information criterion.
3. The method for detecting vibration anomalies of a nuclear power cold source pump according to claim 2, wherein the fitting order of the dynamic polynomial regression is determined by stepwise regression and bayesian information criteria, comprising: setting a candidate order range, gradually increasing the polynomial order according to the candidate order range, and establishing a corresponding fitting model; Calculating BIC values of the candidate models; In the step-by-step selection process, when the BIC value is not reduced any more due to the new increasing order item, stopping model expansion, and taking the order corresponding to the minimum BIC value as the optimal order.
4. The vibration anomaly detection method of a nuclear power cold source pump according to claim 1, wherein outputting the predicted value of the time series characteristic, the predicted value of the multi-source cooperative characteristic and the theoretical predicted value by the LSTM time series prediction model, the gradient lifting tree model and the physical twin model, respectively, comprises: the original vibration signal passes through the LSTM time sequence prediction model to obtain a predicted value of the time sequence characteristic; Constructing the gradient lifting tree model based on multi-source static or slow-varying features; The original vibration signal is passed through the gradient lifting tree model to obtain a predicted value of the multi-source cooperative characteristic; Constructing the physical twin model based on pump structural parameters and a kinetic equation; And the original vibration signal passes through the physical twin model to obtain the theoretical prediction value.
5. The method for detecting vibration abnormality of a nuclear power cold source pump according to claim 1, wherein the adaptive confidence interval includes a confidence lower limit and a confidence upper limit, and performing abnormality detection according to the adaptive confidence interval includes: triggering an anomaly when the original vibration signal is less than the lower confidence limit or the original vibration signal is greater than the upper confidence limit and continuously exceeds a set time.
6. The method for detecting vibration abnormality of a nuclear power cold source pump according to claim 1, wherein after obtaining the abnormality determination result, the method further comprises: Acquiring an abnormal residual sequence, the feature importance output by the stable operation stage, the attention weight and dynamic variance estimation of the LSTM time sequence prediction model; carrying out weighted fusion on the abnormal residual sequence, the feature importance output by the stable operation stage, the attention weight of the LSTM time sequence prediction model and the dynamic variance estimation, and calculating a real-time risk scoring value; and comparing the real-time risk score value with a preset threshold value to determine a risk level.
7. The method of claim 6, wherein comparing the real-time risk score value with a preset threshold value, and determining a risk level comprises: When the real-time risk score value is smaller than a first preset threshold value, determining that the risk is low; When the real-time risk score value is greater than or equal to the first preset threshold value and smaller than a second preset threshold value, judging that the risk is in a stroke; And when the real-time risk score value is greater than or equal to the second preset threshold value, judging that the risk is high.
8. A vibration anomaly detection system for a nuclear power cold source pump, the system comprising: the acquisition module is used for acquiring an original vibration signal; The processing module is used for processing the original vibration signal through a Savitzky-Golay smooth function and dynamic polynomial regression in the start-stop stage of the cold source pump to obtain a trend prediction value; the fitting residual sequence calculating module is used for calculating a fitting residual sequence according to the trend predicted value and the original vibration signal; the self-adaptive confidence interval construction module is used for calculating a dynamic standard deviation and constructing a corresponding self-adaptive confidence interval by adopting a dynamic estimation method combining a sliding window and exponential weighting for the fitting residual sequence; the first abnormality detection module is used for carrying out abnormality detection according to the self-adaptive confidence interval; The output module is used for outputting a predicted value of a time sequence characteristic, a predicted value of a multi-source cooperative characteristic and a theoretical predicted value through an LSTM time sequence predicted model, a gradient lifting tree model and a physical twin model respectively after the cold source pump completes a start-stop process and enters a relatively stable operation stage; the weighted fusion module is used for dynamically weighting and fusing the predicted value of the time sequence feature, the predicted value of the multi-source cooperative feature and the theoretical predicted value to obtain a final predicted value; The dynamic threshold interval construction module is used for calculating residual errors according to the original vibration signals and the final predicted values and constructing a dynamic threshold interval; and the second abnormality detection module is used for carrying out abnormality detection according to the dynamic threshold interval to obtain an abnormality judgment result.
9. An electronic device comprising a memory, a processor, the memory having stored thereon a computer program executable on the processor, characterized in that the processor implements the method of any of the preceding claims 1 to 7 when the computer program is executed.
10. A computer readable medium having non-volatile program code executable by a processor, the program code causing the processor to perform the method of any one of the preceding claims 1 to 7.

Description

Vibration anomaly detection method and system for nuclear power cold source pump Technical Field The invention relates to the technical field of nuclear power station equipment state detection, in particular to a method and a system for detecting vibration abnormality of a nuclear power cold source pump. Background The nuclear power cold source pump is used as a core component of a nuclear power station circulating cooling system, and the vibration state of the nuclear power cold source pump directly influences the running stability of the pump set and the safety of a nuclear power unit. The existing vibration anomaly detection method mainly comprises the following steps: 1) And the fixed threshold method is to judge by setting a fixed threshold value of vibration acceleration or displacement. However, the method has poor adaptability in the start-stop stage of the cold source pump, and false alarm or missing alarm is easy to occur due to large vibration fluctuation amplitude. 2) The statistical threshold method is to set a discrimination boundary according to the statistical distribution of the historical operation data. Because the data in the start-stop stage is insufficient and the change is quick, the statistical threshold value is difficult to accurately reflect the characteristic of the non-stable process. 3) A single machine learning method such as a Support Vector Machine (SVM), a common neural network and the like generally has a good effect only in a stable operation stage, but cannot simultaneously consider transient severe fluctuation in a start-stop stage and high-precision prediction in the stable stage. Disclosure of Invention In view of the above, the invention aims to provide a vibration anomaly detection method and a vibration anomaly detection system for a nuclear power cold source pump, which are used for realizing real-time anomaly identification and intelligent early warning of the cold source pump in the processes of start-stop and stable operation by combining start-stop stage segmentation trend fitting, stable stage multi-model fusion prediction and multi-source information interpretable analysis. In a first aspect, an embodiment of the present invention provides a method for detecting vibration anomalies of a nuclear power cold source pump, where the method includes: Acquiring an original vibration signal; in the start-stop stage of the cold source pump, the original vibration signal is processed through a Savitzky-Golay smooth function and dynamic polynomial regression to obtain a trend prediction value; calculating a fitting residual sequence according to the trend predicted value and the original vibration signal; calculating a dynamic standard deviation and a corresponding self-adaptive confidence interval by adopting a dynamic estimation method combining a sliding window and exponential weighting for the fitting residual sequence; Performing anomaly detection according to the self-adaptive confidence interval; When the cold source pump finishes the start-stop process and enters a relatively stable operation stage, the original vibration signal is respectively output by an LSTM time sequence prediction model, a gradient lifting tree model and a physical twin model to obtain a predicted value of a time sequence characteristic, a predicted value of a multi-source cooperative characteristic and a theoretical predicted value; Dynamically weighting and fusing the predicted value of the time sequence feature, the predicted value of the multi-source cooperative feature and the theoretical predicted value to obtain a final predicted value; calculating residual errors according to the original vibration signals and the final predicted values, and constructing a dynamic threshold interval; and carrying out abnormality detection according to the dynamic threshold interval to obtain an abnormality judgment result. Further, the original vibration signal is processed through a Savitzky-Golay smoothing function and a dynamic polynomial regression to obtain a trend prediction value, which comprises the following steps: denoising and trend maintaining the original vibration signal by adopting the Savitzky-Golay smoothing function to obtain a smoothed signal; and fitting the smoothed signal by adopting the dynamic polynomial regression to obtain the trend predicted value, wherein the fitting order of the dynamic polynomial regression is determined by stepwise regression and Bayesian information criterion. Further, the fitting order of the dynamic polynomial regression is determined by stepwise regression and bayesian information criteria, including: setting a candidate order range, gradually increasing the polynomial order according to the candidate order range, and establishing a corresponding fitting model; Calculating BIC values of the candidate models; In the step-by-step selection process, when the BIC value is not reduced any more due to the new increasing order item, stopping model expan