A health assessment method and system for pumping station units based on multi-objective dynamic optimization of health degradation index

Abstract

The present invention proposes a pump station unit health assessment method and system based on a multi-objective dynamic optimization health degradation index. The method includes: using preprocessed multidimensional time series data and training labels to train a health benchmark model based on an interactive graph neural network; using the health benchmark model to predict the predicted value of a health status assessment characteristic variable; calculating the distance between the predicted value of the health status assessment characteristic variable and the real-time monitored value of the health status assessment characteristic variable based on the Mahalanobis distance to construct a health degradation index for the health status assessment characteristic variable; using the health degradation index to construct and solve a multi-objective function for determining the weight of the health degradation index; obtaining the weight of the health degradation index; and taking a weighted sum of the weight and the health degradation index to obtain a comprehensive health degradation index. The proposed solution can significantly improve the accuracy and comprehensiveness of the assessment by optimizing the weight distribution, more accurately reflecting the actual health status changes of the equipment during operation.

Description

Technical Field

[0001] The present invention belongs to the fields of health assessment and artificial intelligence, and in particular relates to a health assessment method and system for pump station units based on multi-objective dynamic optimization of health degradation index. Background Art

[0002] Pumping stations play a crucial role in water conservancy projects, shouldering multiple tasks such as flood control and drainage, placing extremely high demands on their operational safety. As a core component of a pumping station, the proper operation of the pumping units is crucial to ensuring the smooth operation of the water conservancy system. However, due to complex operating conditions and harsh working environments, the performance of various pumping station components can gradually deteriorate or even fail over time, impacting the safe, stable, and efficient operation of the pumping station units. In recent years, condition-based maintenance strategies for pumping station units have gained increasing attention. Compared with traditional planned maintenance and post-maintenance, condition-based maintenance can effectively avoid excessive or untimely maintenance, thereby reducing maintenance costs and improving the overall efficiency of the pumping station. Therefore, accurately assessing the real-time operating status of pumping station units is crucial for effectively maintaining their condition and reducing failure rates. This not only helps to modernize the water conservancy system but also promotes the steady development of the national economy.

[0003] With the rapid development of the Internet of Things (IoT) technology, pumping stations are equipped with numerous sensors designed to monitor the operating status of equipment in real time. These sensors generate a large number of multivariate, time-stamped observation sequences, known as multidimensional time series. Although condition monitoring systems can acquire multidimensional time series, extracting key information representing the health status of the unit and conducting effective, real-time assessments from them remains a significant challenge. First, the complexity and diversity of multidimensional time series require health assessment models to be able to handle high dimensionality and nonlinearity. The dynamic interrelationships between factors such as voltage, current, and temperature within pumping stations, as well as the long-term dependencies of sensor time series, complicate accurate health assessment. In particular, single data point analysis cannot fully capture temporal dependencies; instead, the logical order and recursive relationships of sampling points within the series must be comprehensively considered. Furthermore, in practical engineering applications, pumping stations often face complex operating conditions and strong noise backgrounds. The variability of operating conditions and the uncertainty of monitoring signals further exacerbate the challenges of data processing and health assessment. Therefore, effective models must not only filter noise and process complex data but also capture the inherent uncertainty and dynamic correlations of each monitoring signal to accurately assess the health status of pumping stations.

[0004] Currently, the main methods for mechanical equipment health assessment include physical model-driven, knowledge-driven, and data-driven approaches. Pump station units have complex mechanical structures, and their degradation mechanisms in complex environments have not yet been fully understood. The health status of key components exhibits complex degradation patterns, and the components of the unit are coupled to each other, making it difficult to accurately establish an assessment model. Knowledge-driven models establish a mapping relationship between degradation characteristics and health status through reasoning and analysis based on expert experience. However, these methods rely heavily on the domain knowledge of experts, making the constructed health assessment models susceptible to human interference and lacking objectivity. Data-driven health assessment methods, to avoid the limitations of mathematical models and expert prior knowledge, treat the pump station unit as a black box and map the extracted characteristic parameters to the health status. These data-driven methods show considerable promise in estimating the health status of pump station units.

[0005] Data-driven health assessment tasks generally involve two steps: 1) Building a health benchmark model (HBM) based on the unit's health data to fit the unit's health signals. 2) Using this HBM, the constructed health value is predicted under degraded conditions. Based on the difference between the actual and predicted signals, a health deterioration index (HDI) is calculated to characterize the equipment's performance degradation. This indicates that the HBM's ability to fit complex operating conditions and the scientific nature of the HDI construction are prerequisites for effective health assessment.

[0006] To improve the fitting capabilities of HBMs (Highly Deficit Models) under complex operating conditions and the scientificity of HDI construction, deep learning methods have been introduced for health status assessment tasks. Several improved HBMs and HDIs have been studied and proposed. However, time series data from pump station units monitored by multiple sensors exhibit complex nonlinear coupling over time, and the sample data exhibits nonlinear and heterogeneous distributions. Three major issues remain: 1) When constructing models from multidimensional monitoring data, only the temporal characteristics of the parameters are considered, while the internal spatial dependencies formed by the complex correlations between the parameters are ignored. This makes the models inadequate for learning features from the interacting multidimensional monitoring data. Furthermore, in actual operating conditions, noise interference caused by abnormal operating environments, human interference, and data acquisition equipment failures degrades data quality, resulting in slightly lower accuracy of the constructed HBMs and unstable assessment results. 2) Pump station unit operating conditions are variable, and these methods directly learn features from signal amplitudes without considering the randomness and uncertainty of the monitoring signals. When constructing HDIs based on difference measurements, the overall distribution of the data is not considered, resulting in large errors in comparisons between measurements. 3) Pump station unit degradation status uses comprehensive evaluation indicators to comprehensively assess unit health. However, different indicators change in real time due to their inherent characteristics and the degree to which they dynamically change and impact health. Simple mean summation and fusion cannot accurately describe the degradation status of pump station units. Therefore, existing deep learning health status assessment methods are still insufficient in extracting the temporal dependencies and complex coupling characteristics of multidimensional time series to construct a health benchmark model and objectively and scientifically construct a health degradation index. Summary of the Invention

[0007] In order to solve the above technical problems, the present invention proposes a technical solution for a pump station unit health assessment method based on a multi-objective dynamic optimization health degradation index to solve the above technical problems.

[0008] The first aspect of the present invention discloses a health assessment method for a pump station unit based on a multi-objective dynamic optimization health degradation index, the method comprising:

[0009] Step S1, performing data preprocessing on the multidimensional time series data collected by the pump station unit sensor to obtain preprocessed multidimensional time series data;

[0010] Step S2: using random forest to perform feature importance assessment on the pre-processed multidimensional time series data, and screening out unit health status assessment feature variables as training labels;

[0011] Step S3, using the preprocessed multidimensional time series data and training labels to train a health benchmark model based on an interactive graph neural network;

[0012] Step S4: pre-process the multi-dimensional time series data collected in real time and input it into the trained health benchmark model to obtain the predicted values ​​of the characteristic variables of the health status assessment of the real-time running unit;

[0013] Step S5: calculating the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable based on the Mahalanobis distance, and constructing a health degradation index of the unit health status assessment characteristic variable;

[0014] Step S6: Apply the health degradation index to construct a multi-objective function for determining the weight of the health degradation index; use a sequential least squares algorithm to solve the multi-objective function to obtain the weight of the health degradation index; and add the weight and the health degradation index to obtain a comprehensive health degradation index.

[0015] According to the method of the first aspect of the present invention, in step S1, the multidimensional time series data includes: unit vibration, unit swing, stuffing box temperature, upper and lower guide bearing temperatures, upper and lower guide cooling water temperatures, unit noise, blade mechanism temperature, oil pan temperature, flow, stator winding temperature, thrust bearing temperature, frequency conversion current, power, voltage, winding temperature and ambient temperature, pump outlet pressure, cooling water flow, main pipe pressure, water supply main pipe temperature and vacuum pipe pressure.

[0016] According to the method of the first aspect of the present invention, in step S3, the health benchmark model based on the interactive graph neural network includes:

[0017] Firstly, the pre-processed multi-dimensional time series data is input into an adaptive multi-scale identifier to construct multi-scale time series with different noise levels;

[0018] Then, the multi-scale time series is input into the cross-scale GNN to extract the time node features in the multi-scale time series;

[0019] Then, the time node features are input into the cross-variable GNN to extract the variable node features in the multidimensional time series;

[0020] Finally, the variable node features are input into a multilayer perceptron to obtain the predicted values ​​of the unit health status assessment feature variables.

[0021] According to the method of the first aspect of the present invention, in step S5, calculating the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable based on the Mahalanobis distance to construct the health degradation index of the unit health status assessment characteristic variable includes:

[0022] Using a Gaussian cloud model to quantify the real-time monitoring value of the unit health status assessment characteristic variable and the predicted value of the unit health status assessment characteristic variable to obtain a characteristic vector of the real-time monitoring value and a characteristic vector of the predicted value;

[0023] The feature vector of the real-time monitoring value includes: the expectation of the real-time monitoring value, the entropy of the real-time monitoring value and the super entropy of the real-time monitoring value;

[0024] The feature vector of the predicted value includes: the expectation of the predicted value, the entropy of the predicted value and the super entropy of the predicted value;

[0025] The Mahalanobis distance is used to calculate the distance between the eigenvector of the real-time monitoring value and the eigenvector of the predicted value;

[0026]

[0027] in, HDI i Indicates the i The health degradation index of the characteristic variables for the health status assessment of each unit; D ij Indicates the i The distance between the characteristic vector of the monitored value and the characteristic vector of the predicted value of the characteristic variable of the unit health status assessment.

[0028] According to the method of the first aspect of the present invention, in step S6, applying the health degradation index to construct a multi-objective function for determining the weight of the health degradation index includes:

[0029] Applying the health degradation index to construct information entropy;

[0030] Applying the health degradation index, internal consistency parameters are constructed;

[0031] The information entropy and the internal consistency parameter are applied to construct a multi-objective function for determining the weight of the health degradation index.

[0032] According to the method of the first aspect of the present invention, in step S6, applying the information entropy and the internal consistency parameter to construct a multi-objective function for determining the weight of the health degradation index includes:

[0033]

[0034] in, E represents information entropy; C represents the internal consistency parameter; w i Indicates the i The weight of each health degradation index; α is a parameter between 0 and 1.

[0035] According to the method of the first aspect of the present invention, in step S6, applying the health degradation index to construct information entropy includes:

[0036]

[0037]

[0038] in, HDI i Indicates the i The health degradation index of the characteristic variables for the health status assessment of each unit; p i Indicates the i The probability of the health degradation index value of each unit health status assessment characteristic variable; n Indicates the total number of characteristic variables for unit health status assessment;

[0039] The application of the health degradation index to construct internal consistency parameters includes:

[0040]

[0041] Where W represents the weight matrix of the health degradation index; X represents the characteristic variable matrix of the unit health status assessment; corr() represents the correlation calculation function; and mean() represents the mean function.

[0042] A second aspect of the present invention discloses a pump station unit health assessment system based on multi-objective dynamic optimization of health degradation index, the system comprising:

[0043] The first processing module is configured to perform data preprocessing on the multidimensional time series data collected by the pump station unit sensor to obtain preprocessed multidimensional time series data;

[0044] The second processing module is configured to use random forest to perform feature importance evaluation on the preprocessed multidimensional time series data, and screen out unit health status assessment feature variables as training labels;

[0045] A third processing module is configured to apply the preprocessed multidimensional time series data and training labels to train a health benchmark model based on an interactive graph neural network;

[0046] The fourth processing module is configured to pre-process the multi-dimensional time series data collected in real time and input it into the trained health benchmark model to obtain the predicted value of the characteristic variable of the health status assessment of the real-time running unit;

[0047] A fifth processing module is configured to calculate the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable based on the Mahalanobis distance, and construct a health degradation index of the unit health status assessment characteristic variable;

[0048] The sixth processing module is configured to apply the health degradation index to construct a multi-objective function for determining the weight of the health degradation index; solve the multi-objective function using a sequential least squares algorithm to obtain the weight of the health degradation index; and perform a weighted summation of the weight and the health degradation index to obtain a comprehensive health degradation index.

[0049] A third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements the steps of any one of the methods for health assessment of a pump station unit based on multi-objective dynamic optimization of a health degradation index according to the first aspect of the present disclosure.

[0050] A fourth aspect of the present invention discloses a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program implements the steps of any one of the methods for health assessment of a pump station unit based on multi-objective dynamic optimization of a health degradation index according to the first aspect of the present disclosure.

[0051] In summary, the advantages of the solution proposed in the present invention in industrial health status assessment are reflected in its better capture of the time dependence, spatial dynamic correlation and randomness of the monitoring signals, and the significant improvement of the accuracy and comprehensiveness of the assessment by optimizing the weight distribution, which can more accurately reflect the changes in the health status of the actual operation of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0053] Figure 1 Flowchart of a method for health assessment of a pumping station unit based on multi-objective dynamic optimization of health degradation index according to an embodiment of the present invention;

[0054] Figure 2 This is a structural diagram of a pump station unit health assessment system based on multi-objective dynamic optimization of health degradation index according to an embodiment of the present invention;

[0055] Figure 3 FIG. 4 is a structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0057] A first aspect of the present invention discloses a pump station unit health assessment method based on multi-objective dynamic optimization of health degradation index. Figure 1 FIG. 1 is a flow chart of a method for health assessment of a pumping station unit based on multi-objective dynamic optimization of health degradation index according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0058] Step S1, performing data preprocessing on the multidimensional time series data collected by the pump station unit sensor to obtain preprocessed multidimensional time series data;

[0059] Step S2: using random forest to perform feature importance assessment on the pre-processed multidimensional time series data, and screening out unit health status assessment feature variables as training labels;

[0060] Step S3, using the preprocessed multidimensional time series data and training labels to train a health benchmark model based on an interactive graph neural network;

[0061] Step S4: pre-process the multi-dimensional time series data collected in real time and input it into the trained health benchmark model to obtain the predicted values ​​of the characteristic variables of the health status assessment of the real-time running unit;

[0062] Step S5: calculating the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable based on the Mahalanobis distance, and constructing a health degradation index of the unit health status assessment characteristic variable;

[0063] Step S6: Apply the health degradation index to construct a multi-objective function for determining the weight of the health degradation index; use a sequential least squares algorithm to solve the multi-objective function to obtain the weight of the health degradation index; and add the weight and the health degradation index to obtain a comprehensive health degradation index.

[0064] In step S1, data preprocessing is performed on the multidimensional time series data collected by the pump station unit sensors to obtain preprocessed multidimensional time series data.

[0065] In some embodiments, in step S1, the multidimensional time series data includes: unit vibration, unit swing, stuffing box temperature, upper and lower guide bearing temperatures, upper and lower guide cooling water temperatures, unit noise, blade mechanism temperature, oil pan temperature, flow, stator winding temperature, thrust bearing temperature, variable frequency current, power, voltage, winding temperature and ambient temperature, pump outlet pressure, cooling water flow, main pipe pressure, water supply main pipe temperature and vacuum pipe pressure.

[0066] Specifically, data interpolation: Due to sensor failure, human error, data collection issues, and other factors, some detection data may be missing. Missing data may lead to inaccurate results or prevent effective analysis. Therefore, the nearest neighbor interpolation method is used to reconstruct the missing data.

[0067] Data normalization: Since different sensors have different data characteristics, it is necessary to normalize the data of each sensor using Z-Score based on the mean and standard deviation of the data to eliminate dimensional differences in the data.

[0068] In step S2, random forest is used to perform feature importance evaluation on the preprocessed multidimensional time series data, and characteristic variables for unit health status evaluation are screened out as training labels.

[0069] Specifically, feature importance assessment involves selecting the corresponding out-of-bag data for each decision tree and calculating the out-of-bag error, denoted as errOOB1. Noise is then randomly added to the features of all out-of-bag samples (by shuffling or replacing feature values), and the out-of-bag error is calculated again, resulting in errOOB2. For each feature, its importance is assessed by comparing the difference in error rate before and after the introduction of noise.

[0070] Feature selection: Sort features in descending order based on their calculated importance to determine the proportion of features to be removed or the number of features to be retained. Compare the OOB error rates of different features and select the feature set with the lowest error rate and a reasonable number of features as the key feature variables.

[0071] In step S3, the preprocessed multidimensional time series data and training labels are used to train a health benchmark model based on an interactive graph neural network.

[0072] In some embodiments, in step S3, the health benchmark model based on the interactive graph neural network includes:

[0073] First, the pre-processed multidimensional time series data is input into an adaptive multi-scale identifier to construct multi-scale time series with different noise levels;

[0074] Then, the multi-scale time series is input into the cross-scale GNN to extract the time node features in the multi-scale time series;

[0075] Then, the time node features are input into the cross-variable GNN to extract the variable node features in the multidimensional time series;

[0076] Finally, the variable node features are input into a multilayer perceptron to obtain the predicted values ​​of the unit health status assessment feature variables.

[0077] Specifically, an adaptive multi-scale identifier is designed to extract key frequency components from multidimensional time series through FFT, and the time series is reconstructed based on these frequency components, so as to obtain multi-scale time series with different noise levels.

[0078] The fast Fourier transform (FFT) is used to decompose the signal into sine and cosine components of different frequencies. The average amplitude is calculated to evaluate the average intensity of each frequency component in the original signal, thereby determining the characteristics of periodic changes in the data.

[0079]

[0080] in, represents the amplitude calculation, represents the calculated amplitude value of each frequency, Indicates the lookback window length, right variables are averaged.

[0081] Select the first S frequencies with high amplitude , calculate the length of each , thus obtaining a time series of S scales.

[0082]

[0083] Then perform average pooling on each time series to capture the sequence characteristics from coarse scale to fine scale. Connect the time series of all scales to obtain The output of the adaptive multi-scale recognizer After expansion, it is transformed into Then it is input into the cross-scale interaction module.

[0084] Cross-Scale Interaction Module: Scale features extracted by the adaptive multi-scale identifier are used to initialize and configure a cross-scale temporal correlation graph. Each time point or node in this graph represents a time point at a different scale after processing by the adaptive multi-scale identifier. This graph is then constructed by analyzing the periodicity of each node and selecting its neighbors based on edge weights (representing the correlation between nodes). Furthermore, temporal ordering is taken into account to ensure that each node is connected not only to scale-related neighbors but also to nodes that are immediately adjacent in time.

[0085] Graph initialization: The cross-scale graph of the time dimension is represented as , node definition: Each node in represents a time point at a different scale after processing by the adaptive multi-scale recognizer. Edge definition: An edge represents the correlation between time points at different scales.

[0086] Initialization and configuration of edge weights:

[0087] Weight initialization: To reduce the impact of noise on the relevant weights, two learnable vectors are used and Initialize the relationship matrix between time nodes through the outer product form.

[0088] Weight processing: Apply the ReLU nonlinear function to the relationship matrix to enhance the expressiveness of the model, and then normalize it through the Softmax function to ensure that the sum of the edge weights of each node is 1.

[0089] Construct a cross-scale temporal correlation map:

[0090] In the process of constructing the cross-scale temporal correlation graph, a scale-sensitive neighbor selection strategy is adopted: for each time node , analyze the periodic characteristics of its scale, and then select the node with the strongest correlation (i.e., the largest edge weight) from all available neighbors to ensure that the connections in the graph best reflect the actual correlation between nodes.

[0091]

[0092] in, Extract the highest relevant weight Operations on each node, It is a time node In the The relative weight of the scale.

[0093] Trend sensitive selection: For each time node , in addition to the neighbors selected based on scale sensitivity, the temporal order of the nodes also needs to be considered to ensure that connections are maintained with the nodes immediately before and after them:

[0094]

[0095] in, Indicates the scale of the time node. The trend neighbor set consists of its adjacent time nodes Composition, here It ensures that only nodes that are close in time are selected, i.e. the previous and next time points.

[0096] After the scale-sensitive and trend-sensitive selection process, each node will have two types of neighbor sets: scale-sensitive neighbors and trend-sensitive neighbors By merging these two sets, a comprehensive neighbor set is formed . Then, the relevance weight matrix needs to be updated , to reflect the merged neighbor set, and then construct a cross-scale correlation graph that represents the relationship between nodes at different time scales, ensuring the accuracy and effectiveness of the graph structure.

[0097] Cross-scale interaction: information propagation is performed on the constructed cross-scale graph, and cross-scale interaction is performed on the time dimension based on GNN. The information propagation process will be stacked layer:

[0098]

[0099]

[0100]

[0101] in, is the activation function; is a learnable matrix; is a time node feature; It is the aggregation of neighbor time node features. Aggregates neighbor node features from the previous layer to Then, By aggregated time node features and its upper layer features Finally, Normalization of layers is the output of the cross-scale GNN.

[0102] The Cross-Variable Interaction Module receives as input time node features derived from cross-scale interaction processing. These features integrate information from multiple time scales, providing a foundation for in-depth analysis of interactions between variables. In this module, each variable is treated as a node in a graph, with relationships between nodes defined by correlations. Information propagation is performed using a graph neural network (GNN) to optimize interactions between variables. By setting homogeneous and heterogeneous neighbors and normalizing correlation weights, the module accurately captures and processes complex dynamic relationships between variables.

[0103] The cross-variable plot is defined as . is a variable node set, where is a variable set Each element in represents the correlation weight between two variables. Initialized by two latent vectors and vectors. Each variable is considered a node in the graph .side The weights represent the correlation between variables, through two latent vectors and The product is initialized and processed by ReLU and Softmax functions:

[0104] Heterogeneity decoupling. Select the node with each variable The most relevant nodes as homogeneous (positive) neighbors, and select the one with the weakest correlation nodes as heterogeneous (negative) neighbors. Sorting and selecting the top and bottom K nodes to achieve this: and .

[0105] The relevant weights are normalized. For each variable node , the weights of homogeneous and heterogeneous relationships are normalized separately, and then a cross-variable graph with separated homogeneous and heterogeneous correlations is constructed using the updated weights.

[0106] Cross-variable interaction: Use GNN to propagate information on cross-variable graphs. The feature update of each variable node takes into account the influence from its positive and negative neighbors:

[0107]

[0108]

[0109]

[0110] is the variable node feature; is the aggregation of neighbor variable node features. Aggregated with Related homogeneous and heterogeneous neighbor node features of the previous layer. Variable features through aggregation and the features of the previous layer Finally, Normalization of layers is the output of the cross-variable GNN.

[0111] Multi-step Forecasting: This forecasting process uses the output features from the cross-variable interaction module and relies on two multi-layer perceptrons (MLPs). The first MLP maps the time dimension of the features from the length of the original time series to 1, achieving time compression of the features. The second MLP then maps the compressed time dimension from the length of the historical input sequence to the length of the target output sequence, generating a future forecast sequence.

[0112] In step S5, the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable is calculated based on the Mahalanobis distance, and a health degradation index of the unit health status assessment characteristic variable is constructed.

[0113] In some embodiments, in step S5, calculating the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable based on the Mahalanobis distance to construct the health degradation index of the unit health status assessment characteristic variable includes:

[0114] Using a Gaussian cloud model to quantify the real-time monitoring value of the unit health status assessment characteristic variable and the predicted value of the unit health status assessment characteristic variable to obtain a characteristic vector of the real-time monitoring value and a characteristic vector of the predicted value;

[0115] The feature vector of the real-time monitoring value includes: the expectation of the real-time monitoring value, the entropy of the real-time monitoring value and the super entropy of the real-time monitoring value;

[0116] The feature vector of the predicted value includes: the expectation of the predicted value, the entropy of the predicted value and the super entropy of the predicted value;

[0117] The Mahalanobis distance is used to calculate the distance between the eigenvector of the real-time monitoring value and the eigenvector of the predicted value;

[0118]

[0119] in, HDI i Indicates the i The health degradation index of the characteristic variables for the health status assessment of each unit; D ij Indicates the i The distance between the characteristic vector of the monitored value and the characteristic vector of the predicted value of the characteristic variable of the unit health status assessment.

[0120] Specifically, the Gaussian cloud model is used to quantify the real-time monitoring value of the unit health status assessment characteristic variable and the predicted value of the unit health status assessment characteristic variable, and the characteristic vector of the real-time monitoring value and the characteristic vector of the predicted value are obtained, including:

[0121] First, the multidimensional time series data is transformed into qualitative concepts through the Backward Cloud Generator (BCG) ,in denote expectation, entropy and super entropy respectively; then, the forward cloud generator (FCG) is based on Generate new data points to form the eigenvector GCM of the real-time monitoring value and the eigenvector GCM of the predicted value, which are represented by GCM1 and GCM2 respectively.

[0122] The Mahalanobis distance (MD) is used to calculate the deviation between the real-time monitoring GCM1 and the health prediction GCM2. Each GCM has three numerical characteristics. They are mapped into three-dimensional space, assuming GCM1 and GCM2 Mapped into three-dimensional space, the distance between two points is :

[0123]

[0124] Among them, the vector and , and the matrix is the covariance matrix of vectors I and J The inverse matrix of .

[0125] HDI is defined as:

[0126]

[0127] Health degradation index representing key characteristic variables. Distance The larger the distance, the higher the degree of unit health degradation; conversely, the closer the distance, the lower the degree of unit health degradation.

[0128] In step S6, the health degradation index is applied to construct a multi-objective function for determining the weight of the health degradation index; the multi-objective function is solved using a sequential least squares algorithm to obtain the weight of the health degradation index; and the weight is weighted and summed with the health degradation index to obtain a comprehensive health degradation index.

[0129] In some embodiments, in step S6, applying the health degradation index to construct a multi-objective function for determining the weight of the health degradation index includes:

[0130] Applying the health degradation index to construct information entropy;

[0131] Applying the health degradation index, internal consistency parameters are constructed;

[0132] The information entropy and the internal consistency parameter are applied to construct a multi-objective function for determining the weight of the health degradation index.

[0133] The application of the information entropy and the internal consistency parameter to construct a multi-objective function for determining the weight of the health degradation index includes:

[0134]

[0135] in, E represents information entropy; C represents the internal consistency parameter; w i Indicates the i The weight of each health degradation index; α It is a parameter between 0 and 1 that is used to balance the weights of information entropy and intrinsic consistency in the objective function.

[0136] The application of the health degradation index to construct information entropy includes:

[0137]

[0138]

[0139] in, HDI i Indicates the i The health degradation index of the characteristic variables for the health status assessment of each unit; p i Indicates the i The probability of the health degradation index value of each unit health status assessment characteristic variable; n Represents the total number of characteristic variables for unit health status assessment; Entropy is a measure of uncertainty or information content, ensuring the diversity of weight distribution of key characteristic variables so that no single indicator overly dominates the comprehensive health degradation index (RCHDI).

[0140] The application of the health degradation index to construct internal consistency parameters includes:

[0141]

[0142] Where W represents the weight matrix of the health degradation index; X represents the matrix of characteristic variables for unit health assessment; corr() represents the correlation calculation function; and mean() represents the mean function. Internal consistency measures the degree of correlation between key characteristic variables and is measured by calculating the average value of the correlation matrix of weighted indicator values.

[0143] Specifically, a sliding window mechanism is used to extract a data segment of a fixed window size at each time point as the input of the current optimization, which is used to capture the latest dynamic information while avoiding the increase in computational complexity caused by excessive data.

[0144] According to the optimal weight vector solved, The comprehensive health degradation index (RCHDI) is obtained by weighted aggregation of the HDIs to comprehensively and real-timely characterize the changes in health status.

[0145]

[0146] in, Represents the current moment, Represents the current moment i The health degradation index of the characteristic variables for the health status assessment of each unit.

[0147] In summary, the solution proposed in the present invention not only eliminates unexpected noise in the time dimension, but also selects the optimal scale at different noise levels to ensure the accuracy and reliability of the evaluation results, effectively solving the noise problem that is prevalent in industrial environments. Secondly, the cross-scale interactive GNN adopted by the present invention is designed to capture data features under clearer trends and low noise conditions, which is difficult to achieve in traditional single-scale analysis methods. In addition, the cross-variable interactive GNN can simulate the dynamic correlation between different variables and reveal complex spatiotemporal dependencies by learning the homogeneity and heterogeneity relationships between variables. This in-depth multidimensional time series analysis based on cross-scale and cross-variable interactive graph neural networks not only improves the accuracy of health status assessment, but also enhances the robustness of the HBM of pump station units when facing industrial data. The health degradation index HDI of the pump station unit constructed by the MD-GCM method not only takes into account the uncertainty of the monitoring signal, but also uses the Mahalanobis distance to calculate the deviation, avoiding the influence of model hyperparameters and improving the sensitivity to changes in health status. Finally, the optimal weight vector calculated by the multi-objective optimization algorithm is combined with the dynamically updated sliding window weight to obtain the comprehensive health degradation index RCHDI, which fully reveals the intrinsic characteristics and real-time dynamic changes of each HDI, thereby accurately characterizing the changes in the health status of the unit.

[0148] The second aspect of the present invention discloses a pump station unit health assessment system based on multi-objective dynamic optimization of health degradation index. Figure 2 FIG is a structural diagram of a pump station unit health assessment system based on multi-objective dynamic optimization of health degradation index according to an embodiment of the present invention; Figure 2 As shown, the system 100 includes:

[0149] The first processing module 101 is configured to perform data preprocessing on the multidimensional time series data collected by the pump station unit sensor to obtain preprocessed multidimensional time series data;

[0150] The second processing module 102 is configured to use random forest to perform feature importance evaluation on the pre-processed multi-dimensional time series data, and screen out unit health status evaluation feature variables as training labels;

[0151] The third processing module 103 is configured to use the preprocessed multidimensional time series data and training labels to train a health benchmark model based on an interactive graph neural network;

[0152] The fourth processing module 104 is configured to pre-process the multi-dimensional time series data collected in real time and input it into the trained health benchmark model to obtain the predicted value of the characteristic variable of the health status assessment of the real-time unit;

[0153] The fifth processing module 105 is configured to calculate the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable based on the Mahalanobis distance, and construct a health degradation index of the unit health status assessment characteristic variable;

[0154] The sixth processing module 106 is configured to apply the health degradation index to construct a multi-objective function for determining the weight of the health degradation index; solve the multi-objective function using a sequential least squares algorithm to obtain the weight of the health degradation index; and perform a weighted summation of the weight and the health degradation index to obtain a comprehensive health degradation index.

[0155] According to the system of the second aspect of the present invention, the first processing module 101 is specifically configured as follows: the multidimensional time series data includes: unit vibration, unit swing, stuffing box temperature, upper and lower guide bearing temperatures, upper and lower guide cooling water temperatures, unit noise, blade mechanism temperature, oil pan temperature, flow, stator winding temperature, thrust bearing temperature, variable frequency current, power, voltage, winding temperature and ambient temperature, pump outlet pressure, cooling water flow, main pipe pressure, water supply main pipe temperature and vacuum pipe pressure.

[0156] Specifically, data interpolation: Due to sensor failure, human error, data collection issues, and other factors, some detection data may be missing. Missing data may lead to inaccurate results or prevent effective analysis. Therefore, the nearest neighbor interpolation method is used to reconstruct the missing data.

[0157] Data normalization: Since different sensors have different data characteristics, it is necessary to normalize the data of each sensor using Z-Score based on the mean and standard deviation of the data to eliminate dimensional differences in the data.

[0158] According to the system of the second aspect of the present invention, the second processing module 102 is specifically configured to evaluate feature importance: for each decision tree, select corresponding out-of-bag data to calculate the out-of-bag data error, recorded as errOOB1; randomly add noise interference to the features of all out-of-bag data samples (disrupting or replacing feature values), and then calculate the out-of-bag data error again to obtain errOOB2. For each feature, its importance is evaluated by comparing the difference in error rate before and after the introduction of noise.

[0159] Feature selection: Sort features in descending order based on their calculated importance to determine the proportion of features to be removed or the number of features to be retained. Compare the OOB error rates of different features and select the feature set with the lowest error rate and a reasonable number of features as the key feature variables.

[0160] According to the system of the second aspect of the present invention, the third processing module 103 is specifically configured as follows: the health benchmark model based on the interactive graph neural network includes:

[0161] Firstly, the pre-processed multi-dimensional time series data is input into an adaptive multi-scale identifier to construct multi-scale time series with different noise levels;

[0162] Then, the multi-scale time series is input into the cross-scale GNN to extract the time node features in the multi-scale time series;

[0163] Then, the time node features are input into the cross-variable GNN to extract the variable node features in the multidimensional time series;

[0164] Finally, the variable node features are input into a multilayer perceptron to obtain the predicted values ​​of the unit health status assessment feature variables.

[0165] Specifically, an adaptive multi-scale identifier is designed to extract key frequency components from multidimensional time series through FFT, and the time series is reconstructed based on these frequency components, so as to obtain multi-scale time series with different noise levels.

[0166] The fast Fourier transform (FFT) is used to decompose the signal into sine and cosine components of different frequencies. The average amplitude is calculated to evaluate the average intensity of each frequency component in the original signal, thereby determining the characteristics of periodic changes in the data.

[0167]

[0168] in, represents the amplitude calculation, represents the calculated amplitude value of each frequency, Indicates the lookback window length, right variables are averaged.

[0169] Select the first S frequencies with high amplitude , calculate the length of each , thus obtaining a time series of S scales.

[0170]

[0171] Then perform average pooling on each time series to capture the sequence characteristics from coarse scale to fine scale. Connect the time series of all scales to obtain The output of the adaptive multi-scale recognizer After expansion, it is transformed into Then it is input into the cross-scale interaction module.

[0172] Cross-Scale Interaction Module: Scale features extracted by the adaptive multi-scale identifier are used to initialize and configure a cross-scale temporal correlation graph. Each time point or node in this graph represents a time point at a different scale after processing by the adaptive multi-scale identifier. This graph is then constructed by analyzing the periodicity of each node and selecting its neighbors based on edge weights (representing the correlation between nodes). Furthermore, temporal ordering is taken into account to ensure that each node is connected not only to scale-related neighbors but also to nodes that are immediately adjacent in time.

[0173] Graph initialization: The cross-scale graph of the time dimension is represented as , node definition: Each node in represents a time point at a different scale after processing by the adaptive multi-scale recognizer. Edge definition: An edge represents the correlation between time points at different scales.

[0174] Initialization and configuration of edge weights:

[0175] Weight initialization: To reduce the impact of noise on the relevant weights, two learnable vectors are used and Initialize the relationship matrix between time nodes through the outer product form.

[0176] Weight processing: Apply the ReLU nonlinear function to the relationship matrix to enhance the expressiveness of the model, and then normalize it through the Softmax function to ensure that the sum of the edge weights of each node is 1.

[0177] Construct a cross-scale temporal correlation map:

[0178] In the process of constructing the cross-scale temporal correlation graph, a scale-sensitive neighbor selection strategy is adopted: for each time node , analyze the periodic characteristics of its scale, and then select the node with the strongest correlation (i.e., the largest edge weight) from all available neighbors to ensure that the connections in the graph best reflect the actual correlation between nodes.

[0179]

[0180] in, Extract the highest relevant weight Operations on each node, It is a time node In the The relative weight of the scale.

[0181] Trend sensitive selection: For each time node In addition to selecting neighbors based on scale sensitivity, the temporal order of nodes also needs to be considered to ensure connections with the nodes immediately preceding and following them:

[0182]

[0183] in, Indicates the scale of the time node. The trend neighbor set consists of its adjacent time nodes Composition, here It ensures that only nodes that are close in time are selected, i.e. the previous and next time points.

[0184] After the scale-sensitive and trend-sensitive selection process, each node will have two types of neighbor sets: scale-sensitive neighbors and trend-sensitive neighbors By merging these two sets, a comprehensive neighbor set is formed . Then, the relevance weight matrix needs to be updated , to reflect the merged neighbor set, and then construct a cross-scale correlation graph that represents the relationship between nodes at different time scales, ensuring the accuracy and effectiveness of the graph structure.

[0185] Cross-scale interaction: information propagation is performed on the constructed cross-scale graph, and cross-scale interaction is performed on the time dimension based on GNN. The information propagation process will be stacked layer:

[0186]

[0187]

[0188]

[0189] in, is the activation function; is a learnable matrix; is a time node feature; It is the aggregation of neighbor time node features. Aggregates neighbor node features from the previous layer to Then, By aggregated time node features and its upper layer features Finally, Normalization of layers is the output of the cross-scale GNN.

[0190] The Cross-Variable Interaction Module receives as input time node features derived from cross-scale interaction processing. These features integrate information from multiple time scales, providing a foundation for in-depth analysis of interactions between variables. In this module, each variable is treated as a node in a graph, with relationships between nodes defined by correlations. Information propagation is performed using a graph neural network (GNN) to optimize interactions between variables. By setting homogeneous and heterogeneous neighbors and normalizing correlation weights, the module accurately captures and processes complex dynamic relationships between variables.

[0191] The cross-variable plot is defined as . is a set of variable nodes, where is a variable set Each element in represents the correlation weight between two variables. Initialized by two latent vectors and vectors. Each variable is considered a node in the graph .side The weights represent the correlation between variables, through two latent vectors and The product is initialized and processed by ReLU and Softmax functions:

[0192] Heterogeneity decoupling. Select the node with each variable The most relevant nodes as homogeneous (positive) neighbors, and select the one with the weakest correlation nodes as heterogeneous (negative) neighbors. Sorting and selecting the top and bottom K nodes to achieve this: and .

[0193] The relevant weights are normalized. For each variable node , the weights of homogeneous and heterogeneous relationships are normalized separately, and then a cross-variable graph with separated homogeneous and heterogeneous correlations is constructed using the updated weights.

[0194] Cross-variable interaction: Use GNN to propagate information on cross-variable graphs. The feature update of each variable node takes into account the influence from its positive and negative neighbors:

[0195]

[0196]

[0197]

[0198] is the variable node feature; is the aggregation of neighbor variable node features. Aggregated with Related homogeneous and heterogeneous neighbor node features of the previous layer. Variable features through aggregation and the features of the previous layer Finally, Normalization of layers is the output of the cross-variable GNN.

[0199] Multi-step Forecasting: This forecasting process uses the output features from the cross-variable interaction module and relies on two multi-layer perceptrons (MLPs). The first MLP maps the time dimension of the features from the length of the original time series to 1, achieving time compression of the features. The second MLP then maps the compressed time dimension from the length of the historical input sequence to the length of the target output sequence, generating a future forecast sequence.

[0200] According to the system of the second aspect of the present invention, the fifth processing module 105 is specifically configured to calculate the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable based on the Mahalanobis distance, and construct the health degradation index of the unit health status assessment characteristic variable, including:

[0201] Using a Gaussian cloud model to quantify the real-time monitoring value of the unit health status assessment characteristic variable and the predicted value of the unit health status assessment characteristic variable to obtain a characteristic vector of the real-time monitoring value and a characteristic vector of the predicted value;

[0202] The feature vector of the real-time monitoring value includes: the expectation of the real-time monitoring value, the entropy of the real-time monitoring value and the super entropy of the real-time monitoring value;

[0203] The feature vector of the predicted value includes: the expectation of the predicted value, the entropy of the predicted value and the super entropy of the predicted value;

[0204] The Mahalanobis distance is used to calculate the distance between the eigenvector of the real-time monitoring value and the eigenvector of the predicted value;

[0205]

[0206] in, HDI i Indicates the i The health degradation index of the characteristic variables for the health status assessment of each unit; D ij Indicates the i The distance between the eigenvector of the real-time monitoring value and the eigenvector of the predicted value of the characteristic variable of the unit health status assessment.

[0207] Specifically, the Gaussian cloud model is used to quantify the real-time monitoring value of the unit health status assessment characteristic variable and the predicted value of the unit health status assessment characteristic variable, and the characteristic vector of the real-time monitoring value and the characteristic vector of the predicted value are obtained, including:

[0208] First, the multidimensional time series data is transformed into qualitative concepts through the Backward Cloud Generator (BCG) ,in denote expectation, entropy and super entropy respectively; then, the forward cloud generator (FCG) is based on Generate new data points to form the eigenvector GCM of the real-time monitoring value and the eigenvector GCM of the predicted value, which are represented by GCM1 and GCM2 respectively.

[0209] The Mahalanobis distance (MD) is used to calculate the deviation between the real-time monitoring GCM1 and the health prediction GCM2. Each GCM has three numerical characteristics. They are mapped into three-dimensional space, assuming GCM1 and GCM2 Mapped into three-dimensional space, the distance between two points is :

[0210]

[0211] Among them, the vector and , and the matrix is the covariance matrix of vectors I and J The inverse matrix of .

[0212] HDI is defined as:

[0213]

[0214] Health degradation index representing key characteristic variables. Distance The larger the distance, the higher the degree of unit health degradation; conversely, the closer the distance, the lower the degree of unit health degradation.

[0215] According to the system of the second aspect of the present invention, the sixth processing module 106 is specifically configured to apply the health degradation index to construct a multi-objective function for determining the weight of the health degradation index, including:

[0216] Applying the health degradation index to construct information entropy;

[0217] Applying the health degradation index, internal consistency parameters are constructed;

[0218] The information entropy and the internal consistency parameter are applied to construct a multi-objective function for determining the weight of the health degradation index.

[0219] The application of the information entropy and the internal consistency parameter to construct a multi-objective function for determining the weight of the health degradation index includes:

[0220]

[0221] in, E represents information entropy; C represents the internal consistency parameter; w i Indicates the i The weight of each health degradation index; α It is a parameter between 0 and 1 that is used to balance the weights of information entropy and intrinsic consistency in the objective function.

[0222] The application of the health degradation index to construct information entropy includes:

[0223]

[0224]

[0225] in, HDI i Indicates the i The health degradation index of the characteristic variables for the health status assessment of each unit; p i Indicates the i The probability of the health degradation index value of each unit health status assessment characteristic variable; n Represents the total number of characteristic variables for unit health status assessment; Entropy is a measure of uncertainty or information content, ensuring the diversity of weight distribution of key characteristic variables so that no single indicator overly dominates the comprehensive health degradation index (RCHDI).

[0226] The application of the health degradation index to construct internal consistency parameters includes:

[0227]

[0228] Where W represents the weight matrix of the health degradation index; X represents the matrix of characteristic variables for unit health assessment; corr() represents the correlation calculation function; and mean() represents the mean function. Internal consistency measures the degree of correlation between key characteristic variables and is measured by calculating the average value of the correlation matrix of weighted indicator values.

[0229] Specifically, a sliding window mechanism is used to extract a data segment of a fixed window size at each time point as the input of the current optimization, which is used to capture the latest dynamic information while avoiding the increase in computational complexity caused by excessive data.

[0230] According to the optimal weight vector solved, The comprehensive health degradation index (RCHDI) is obtained by weighted aggregation of the HDIs to comprehensively and real-timely characterize the changes in health status.

[0231]

[0232] in, Represents the current moment, Represents the current moment i The health degradation index of the characteristic variables for the health status assessment of each unit.

[0233] A third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements the steps of any one of the methods for health assessment of a pump station unit based on multi-objective dynamic optimization of a health degradation index as disclosed in the first aspect of the present invention.

[0234] Figure 3 FIG. 1 is a structural diagram of an electronic device according to an embodiment of the present invention. Figure 3 As shown, the electronic device includes a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface of the electronic device is used to communicate with an external terminal via wired or wireless communication. The wireless communication method can be achieved through Wi-Fi, a carrier network, near-field communication (NFC), or other technologies. The display of the electronic device can be a liquid crystal display or an electronic ink display. The input device of the electronic device can be a touch layer covering the display, or it can be buttons, a trackball, or a touchpad provided on the electronic device housing, or it can be an external keyboard, touchpad, or mouse.

[0235] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a structural diagram of the part related to the technical solution of the present disclosure, and does not constitute a limitation on the electronic device to which the solution of the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0236] A fourth aspect of the present invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when executed by a processor, implements any of the steps in the method for health assessment of a pumping station unit based on multi-objective dynamic optimization of a health degradation index as disclosed in the first aspect of the present invention.

[0237] Please note that the technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above embodiments only express several implementation methods of the present application. The description is relatively specific and detailed, but it cannot be understood as a limitation on the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, without departing from the concept of this application, several variations and improvements can be made, which all fall within the scope of protection of this application. Therefore, the scope of protection of the patent in this application shall be based on the attached claims.

Claims

1. A health assessment method for pumping station units based on multi-objective dynamic optimization of health degradation index, characterized in that: The method comprises: Step S1, performing data preprocessing on the multidimensional time series data collected by the pump station unit sensor to obtain preprocessed multidimensional time series data; Step S2: using random forest to perform feature importance assessment on the pre-processed multidimensional time series data, and screening out unit health status assessment feature variables as training labels; Step S3: Using the pre-processed multidimensional time series data and training labels to train a health benchmark model based on an interactive graph neural network; the health benchmark model based on an interactive graph neural network includes: Firstly, the pre-processed multi-dimensional time series data is input into an adaptive multi-scale identifier to construct multi-scale time series with different noise levels; Then, the multi-scale time series is input into a cross-scale graph neural network (GNN) to extract time node features in the multi-scale time series; Then, the time node features are input into a cross-variable graph neural network (GNN) to extract variable node features in the multidimensional time series; Finally, the variable node features are input into a multilayer perceptron to obtain the predicted values ​​of the unit health status assessment feature variables; Step S4: pre-process the multi-dimensional time series data collected in real time and input it into the trained health benchmark model to obtain the predicted values ​​of the characteristic variables of the health status assessment of the real-time running unit; Step S5: Calculating the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable based on the Mahalanobis distance to construct a health degradation index of the unit health status assessment characteristic variable; the step of calculating the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable based on the Mahalanobis distance to construct a health degradation index of the unit health status assessment characteristic variable includes: Using a Gaussian cloud model to quantify the real-time monitoring value of the unit health status assessment characteristic variable and the predicted value of the unit health status assessment characteristic variable to obtain a characteristic vector of the real-time monitoring value and a characteristic vector of the predicted value; The feature vector of the real-time monitoring value includes: the expectation of the real-time monitoring value, the entropy of the real-time monitoring value and the super entropy of the real-time monitoring value; The feature vector of the predicted value includes: the expectation of the predicted value, the entropy of the predicted value and the super entropy of the predicted value; The Mahalanobis distance is used to calculate the distance between the eigenvector of the real-time monitoring value and the eigenvector of the predicted value; in, HDI i Indicates the i The health degradation index of the characteristic variables for the health status assessment of each unit; D ij Indicates the i The distance between the characteristic vector of the monitored value and the characteristic vector of the predicted value of each unit health status assessment characteristic variable; Step S6: Apply the health degradation index to construct a multi-objective function for determining the weight of the health degradation index; use a sequential least squares algorithm to solve the multi-objective function to obtain the weight of the health degradation index; and add the weight and the health degradation index to obtain a comprehensive health degradation index.

2. A pump station unit health assessment method based on multi-objective dynamic optimization of health degradation index according to claim 1, characterized in that: In step S1, the multidimensional time series data includes: unit vibration, unit swing, stuffing box temperature, upper and lower guide bearing temperatures, upper and lower guide cooling water temperatures, unit noise, blade mechanism temperature, oil pan temperature, flow, stator winding temperature, thrust bearing temperature, frequency conversion current, power, voltage, winding temperature and ambient temperature, pump outlet pressure, cooling water flow, main pipe pressure, water supply main pipe temperature and vacuum pipe pressure.

3. The method for health assessment of pumping station units based on multi-objective dynamic optimization of health degradation index according to claim 1 is characterized in that: In step S6, applying the health degradation index to construct a multi-objective function for determining the weight of the health degradation index includes: Applying the health degradation index to construct information entropy; Applying the health degradation index, internal consistency parameters are constructed; The information entropy and the internal consistency parameter are applied to construct a multi-objective function for determining the weight of the health degradation index.

4. A pump station unit health assessment method based on multi-objective dynamic optimization of health degradation index according to claim 3, characterized in that: In step S6, applying the information entropy and the internal consistency parameter to construct a multi-objective function for determining the weight of the health degradation index includes: in, E represents information entropy; C represents the internal consistency parameter; w i Indicates the i The weight of each health degradation index; α is a parameter between 0 and 1.

5. A pump station unit health assessment method based on multi-objective dynamic optimization of health degradation index according to claim 4, characterized in that: In step S6, applying the health degradation index to construct information entropy includes: in, HDI i Indicates the i The health degradation index of the characteristic variables for the health status assessment of each unit; p i Indicates the i The probability of the health degradation index value of each unit health status assessment characteristic variable; n Indicates the total number of characteristic variables for unit health status assessment; The application of the health degradation index to construct internal consistency parameters includes: Where W represents the weight matrix of the health degradation index; X represents the characteristic variable matrix of the unit health status assessment; corr() represents the correlation calculation function; and mean() represents the mean function.

6. A pump station unit health assessment system based on multi-objective dynamic optimization of health degradation index, implementing the steps of a pump station unit health assessment method based on multi-objective dynamic optimization of health degradation index according to any one of claims 1 to 5, characterized in that: The system comprises: The first processing module is configured to perform data preprocessing on the multidimensional time series data collected by the pump station unit sensor to obtain preprocessed multidimensional time series data; The second processing module is configured to use random forest to perform feature importance evaluation on the preprocessed multidimensional time series data, and screen out unit health status assessment feature variables as training labels; A third processing module is configured to apply the preprocessed multidimensional time series data and training labels to train a health benchmark model based on an interactive graph neural network; The fourth processing module is configured to pre-process the multi-dimensional time series data collected in real time and input it into the trained health benchmark model to obtain the predicted value of the characteristic variable of the health status assessment of the real-time running unit; A fifth processing module is configured to calculate the distance between the predicted value of the unit health status assessment characteristic variable and the real-time monitored value of the unit health status assessment characteristic variable based on the Mahalanobis distance, and construct a health degradation index of the unit health status assessment characteristic variable; The sixth processing module is configured to apply the health degradation index to construct a multi-objective function for determining the weight of the health degradation index; solve the multi-objective function using a sequential least squares algorithm to obtain the weight of the health degradation index; and perform a weighted summation of the weight and the health degradation index to obtain a comprehensive health degradation index.

7. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it implements the steps in the pump station unit health assessment method based on multi-objective dynamic optimization of health degradation index described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the pump station unit health assessment method based on multi-objective dynamic optimization of health degradation index described in any one of claims 1 to 5 are implemented.

Images