CN-121614803-B - Comprehensive evaluation method for state of pump station unit

CN121614803BCN 121614803 BCN121614803 BCN 121614803BCN-121614803-B

Abstract

The invention discloses a comprehensive evaluation method for a state of a pump station unit, and belongs to the technical field of pump station unit detection. The method comprises the steps of obtaining multi-source monitoring data and generating a standardized monitoring data set, calculating time-varying mutual information among indexes, determining a dynamic threshold value by combining current working condition parameters and water level difference correction items, constructing a dynamic association network, performing physical mechanism characteristic decoupling on the data, wherein the step of stripping vibration signal working condition drift components based on a reference curve to obtain vibration residual errors, the step of calculating equivalent standard working condition temperature based on a heat balance principle, the step of respectively calculating a first weight based on data statistics and a second weight based on network topology, generating comprehensive coupling weight in a self-adaptive mode according to consistency coefficients of the first weight and the second weight, and the step of judging health state grades by utilizing a cloud model. The invention solves the problems of difficult extraction of fault characteristics and poor model robustness under variable working conditions through deep fusion of a physical mechanism and data driving, and realizes accurate assessment of the state of a unit.

Inventors

ZHANG YU
CHEN TIANYU
ZHU YU
QU QIU
ZHANG NING

Assignees

水利部交通运输部国家能源局南京水利科学研究院

Dates

Publication Date: 20260508
Application Date: 20260202

Claims (7)

1. The comprehensive evaluation method for the state of the pump station unit is characterized by comprising the following steps of: acquiring multi-source monitoring data of a pump station unit, and performing time alignment and standardization processing to obtain a standardized monitoring data set; calculating time-varying mutual information among all monitoring indexes in the standardized monitoring data set, determining a dynamic threshold value based on current working condition parameters extracted from the standardized monitoring data set, and constructing a dynamic association network; The physical mechanism characteristic decoupling of the standardized monitoring data set comprises the steps of stripping working condition drift components of vibration signals based on a preset working condition-vibration reference curve library to obtain vibration residual characteristics, and converting temperature signals into equivalent standard working condition temperatures based on a heat balance principle; respectively calculating a first weight based on the statistical characteristics of the standardized monitoring data set and a second weight based on the topological characteristics of the dynamic association network, and calculating the consistency coefficient of the first weight and the second weight, thereby generating a comprehensive coupling weight; based on the comprehensive coupling weight, the vibration residual error characteristic and the equivalent standard working condition temperature, judging the health state grade of the unit by utilizing a pre-constructed health grade cloud model, and obtaining an evaluation result; The current working condition parameters comprise the current calculated water level difference and the running power of the unit, and the dynamic threshold value is determined based on the current working condition parameters extracted from the standardized monitoring data set, and the method comprises the following steps: according to the current working condition parameters and a preset working condition clustering center, determining the working condition category to which the current working condition belongs; Calculating a dynamic threshold by adopting a self-adaptive threshold model containing a water level difference correction term; The self-adaptive threshold model is defined as that on the basis of mutual information benchmark statistical characteristics of the class of the working conditions, water level difference correction items are overlapped; the water level difference correction term and the calculated deviation degree of the water level difference relative to the pump station design water level difference form positive correlation, and the correction strength is controlled by a preset water level sensitive factor; Constructing a dynamic association network, comprising: Calculating time-varying mutual information among all monitoring indexes in the standardized monitoring data set to obtain a time-varying mutual information matrix; Comparing elements in the time-varying mutual information matrix with a dynamic threshold value element by element, establishing a connecting edge between corresponding monitoring index nodes when the mutual information value is greater than or equal to the dynamic threshold value, and generating a weighted adjacency matrix by taking the mutual information value as an edge weight; Based on the weighted adjacency matrix, calculating the degree centrality, the medium number centrality and the clustering coefficient of each monitoring index node respectively to be used as the topological characteristic of the dynamic association network; Stripping the working condition drifting component of the vibration signal based on a preset reference curve to obtain a vibration residual error characteristic, wherein the method comprises the following steps: Determining the working condition type of the unit according to the current working condition parameters, and calling a vibration mean value curve corresponding to the working condition type from a preset working condition-vibration reference curve library; the real-time vibration signals in the standardized monitoring data set are subjected to point-by-point subtraction with the vibration mean curve, and a vibration baseline determined by working conditions is removed to obtain a residual signal containing non-stationary abnormal components; And carrying out energy normalization processing on the residual signal to obtain the vibration residual characteristic.
2. The method of claim 1, wherein calculating time-varying mutual information between the monitoring metrics in the normalized monitoring dataset comprises: Sliding window division is carried out on the standardized monitoring data set, and a continuous time window sequence is generated; Aiming at the monitoring index data in each time window, calculating an adaptive bandwidth according to the standard deviation and kurtosis coefficient of the sample, and estimating probability density by adopting a Gaussian kernel function based on the adaptive bandwidth; Based on the estimated edge probability density and the joint probability density, calculating mutual information values among all monitoring index pairs to form a time-varying mutual information matrix sequence corresponding to the time window sequence.
3. The method of claim 1, further comprising, after obtaining the residual vibration feature, performing sparse fault feature extraction on the residual vibration feature: loading a pre-built component sensitive atom library, wherein the component sensitive atom library comprises a plurality of characteristic atoms associated with specific fault types; carrying out sparse decomposition on the vibration residual error characteristics by adopting an orthogonal matching pursuit algorithm, and solving sparse coefficient vectors under a component sensitive atom library; And classifying and accumulating the vibration energy contribution value of each component according to the atomic label corresponding to the non-zero element in the sparse coefficient vector to generate a component sensitive vibration characteristic vector.
4. The method of claim 1, wherein converting the temperature signal to an equivalent standard operating temperature based on a heat balance principle comprises: Establishing a pump station unit heat balance equation comprising an electric loss heat source item, a mechanical friction heat source item, a hydraulic loss heat source item and a cooling heat dissipation item; Substituting the current working condition operation parameters in the standardized monitoring data set into a thermal balance equation to calculate the theoretical steady-state temperature under the current working condition; And compensating the temperature monitoring value acquired in real time by utilizing the difference value between the theoretical steady-state temperature and the reference steady-state temperature, and calculating to obtain the equivalent standard working condition temperature for eliminating the influence of the working condition difference.
5. The method of claim 1, wherein separately calculating a first weight based on statistical features of the standardized monitoring dataset and a second weight based on topological features of the dynamically associated network comprises: Constructing a random forest classification model by using a standardized monitoring data set and a corresponding historical health state label, and determining a first weight of each monitoring index by calculating the feature replacement accuracy reduction of the data outside the bag; Extracting the degree centrality index, the medium number centrality index and the clustering coefficient index of each monitoring index node in the dynamic association network, and carrying out weighted fusion on the three indexes according to a preset physical meaning importance degree weight coefficient to determine the second weight of each monitoring index.
6. The method of claim 5, wherein calculating a consistency coefficient of the first weight and the second weight, thereby generating the integrated coupling weight, comprises: Respectively sequencing the first weight and the second weight to obtain a rank order vector, and calculating a spearman rank correlation coefficient between the first weight and the second weight as a consistency coefficient; Adaptively calculating a network topology fusion coefficient according to the consistency coefficient, wherein the lower the consistency coefficient is, the larger the value of the network topology fusion coefficient is, and increasing the contribution proportion based on the weight of the physical mechanism when the weight distribution is diverged; and carrying out weighted summation on the first weight and the second weight based on the network topology fusion coefficient to obtain the comprehensive coupling weight.
7. The method of claim 1, wherein determining the health status level of the unit using a pre-built health level cloud model comprises: the vibration residual error characteristics, the equivalent standard working condition temperature and the swing degree characteristics, the electric characteristics and the hydraulic characteristics in the standardized monitoring data set are input into a preset health grade cloud model, and a forward cloud generator is utilized to calculate single index membership certainty of each index relative to different health grades; Carrying out weighted summation on the single index membership degree based on the comprehensive coupling weight to generate a comprehensive membership degree vector reflecting the distribution probability of the whole state of the unit on each health grade; and judging the current health state level according to the maximum value in the comprehensive membership degree vector, calculating the distribution entropy and the concentration degree of the comprehensive membership degree vector, and evaluating the credibility of the judging result.

Description

Comprehensive evaluation method for state of pump station unit Technical Field The invention belongs to the technical field of pump station unit detection, and particularly relates to a comprehensive state evaluation method for a pump station unit. Background Along with the construction and development of large-scale industrial equipment, the pump station unit is developed towards high capacity and high rotating speed, and the inside of the pump station unit not only relates to the complex coupling of mechanical, electric and hydraulic systems, but also faces the operation challenges of large water level amplitude, frequent working condition switching and the like. The method has important engineering significance for early identification of equipment degradation trend, prevention of catastrophic failure and optimization of scheduling operation, and is a link for guaranteeing water resource allocation safety. Currently, state monitoring of pump station units mainly depends on vibration, temperature, swing and electrical parameters acquired by a data acquisition and monitoring control SCADA system. The existing evaluation methods are generally divided into two types, namely a discrimination method based on a fixed threshold value and setting a single alarm limit value according to an industry standard, and an intelligent diagnosis method based on data driving and the like, wherein the mapping relation between monitoring data and fault states is mined by utilizing algorithms such as a neural network, a support vector machine and the like or the linear correlation among multiple parameters is analyzed by pearson correlation coefficients. The method has a certain application effect under a steady-state working condition and is widely deployed in an automatic monitoring system. However, the existing scheme still has obvious limitations under the complex environment of variable working conditions, and is mainly characterized by incomplete decoupling of physical mechanisms and statistical data and insufficient adaptability in heterogeneous feature fusion. Specifically, vibration and association characteristics of a pump station unit are affected by upstream and downstream water level differences and blade angles, static threshold values or pure data correlation analysis is mostly adopted in the existing method, physical modulation action of hydraulic working conditions on monitoring indexes is ignored, false association or false alarm is easy to occur under non-design water level working conditions, normal baseline components which change along with loads are contained in temperature and vibration signals, an existing data model often directly processes original signals without stripping working condition drift components, early weak fault characteristics are submerged by working condition background noise which fluctuates greatly, pure statistical weights (such as random forest) and mechanism weights (such as expert experience) frequently collide in the face of outlier data in weight distribution, a dynamic fusion mechanism based on consistency detection is lacked in the prior art, and robustness correction is difficult to effectively utilize a physical topological structure when data rules fail. Disclosure of Invention The invention aims to provide a comprehensive evaluation method for the state of a pump station unit, which aims to solve one of the problems in the prior art. The technical scheme is that the comprehensive evaluation method for the state of the pump station unit comprises the following steps: acquiring multi-source monitoring data of a pump station unit, and performing time alignment and standardization processing to obtain a standardized monitoring data set; calculating time-varying mutual information among all monitoring indexes in the standardized monitoring data set, determining a dynamic threshold value based on current working condition parameters extracted from the standardized monitoring data set, and constructing a dynamic association network; The physical mechanism characteristic decoupling of the standardized monitoring data set comprises the steps of stripping working condition drift components of vibration signals based on a preset reference curve to obtain vibration residual error characteristics, and converting temperature signals into equivalent standard working condition temperatures based on a heat balance principle; respectively calculating a first weight based on the statistical characteristics of the standardized monitoring data set and a second weight based on the topological characteristics of the dynamic association network, and calculating the consistency coefficient of the first weight and the second weight, thereby generating a comprehensive coupling weight; Based on the comprehensive coupling weight, the vibration residual error characteristic and the equivalent standard working condition temperature, the health state grade of the unit is judged by utilizing a pre-cons