Pump station working condition monitoring method and system based on digital twinning and storage medium

Abstract

The invention relates to the technical field of digital twinning, and discloses a pump station working condition monitoring method and system based on digital twinning and a storage medium. The method comprises the following steps: receiving a pump station operation characteristic matrix of pump station equipment through a PLC (Programmable Logic Controller) control cabinet; constructing a hybrid digital twin model based on the pump station operation characteristic matrix and generating operation state prediction data; inputting the operation state prediction data and the actual monitoring data into a Transform network to carry out working condition feature association mode extraction and health index quantitative calculation to obtain an equipment health assessment index; and triggering an early warning response mechanism by using the equipment health assessment index, and outputting a unit start-stop control strategy of the pump station equipment to the scheduling system through the remote management platform. According to the invention, the capability of identifying abnormal working conditions and the accuracy of early warning are improved, the purposes of'few people on duty, remote monitoring and low-consumption operation 'of the pump station are achieved, the operation and maintenance cost is greatly reduced, and the service life of equipment is prolonged.

Description

Technical Field

[0001] The present invention relates to the technical field of digital twins, and in particular, to a method, system, and storage medium for monitoring the operating conditions of a pumping station based on digital twins. Background Art

[0002] With the continuous expansion of the scale of water conservancy project construction and the continuous improvement of the degree of automation, as a key infrastructure in water conservancy projects, the monitoring and control of the operating status of pumping stations face huge challenges. Traditional pumping station monitoring systems mainly rely on manual inspections, fixed-period maintenance, and simple remote monitoring, which cannot meet the modern requirements of "few operators on site, remote monitoring, and low-power operation". This traditional monitoring mode not only consumes a large amount of human resources, but also has a lag in perceiving the equipment status, making it difficult to detect potential faults in a timely manner, resulting in a high equipment damage rate, low operating efficiency, and even serious safety accidents and economic losses.

[0003] The main technical problem faced by current pumping station operating condition monitoring lies in the processing of multi-source heterogeneous data. The pumping station system involves multiple subsystems such as machinery, electricity, and hydraulics, generating a variety of sensor data types, different sampling frequencies, and large signal characteristic differences. Existing monitoring systems are difficult to effectively fuse and analyze this heterogeneous data, especially in complex water conservancy operation environments. Traditional data fusion methods lack the necessary adaptability and cannot handle the inconsistencies and mutual influences between sensor data, resulting in inaccurate and unreliable monitoring results, making it difficult to support refined equipment status assessment and predictive maintenance decisions. Summary of the Invention

[0004] The present invention provides a method, system, and storage medium for monitoring the operating conditions of a pumping station based on digital twins. The present invention improves the ability to identify abnormal operating conditions and the accuracy of early warnings, realizes the goal of "few operators on site, remote monitoring, and low-power operation" for pumping stations, greatly reduces the operation and maintenance costs, and extends the service life of equipment.

[0005] In a first aspect, the present invention provides a method for monitoring the operating conditions of a pumping station based on digital twins. The method for monitoring the operating conditions of a pumping station based on digital twins includes:

[0006] Receiving the pressure, flow rate, vibration, temperature, and electrical energy parameters of the pumping station equipment through a PLC control cabinet to obtain a pumping station operation characteristic matrix;

[0007] Constructing a hybrid digital twin model of a pumping station physical model and a data-driven model based on the pumping station operation characteristic matrix, and generating operation status prediction data of the pumping station equipment through the hybrid digital twin model;

[0008] Input the operation status prediction data and actual monitoring data into the Transformer network to extract the correlation pattern of working conditions characteristics and calculate the quantification of the health index, so as to obtain the equipment health assessment index;

[0009] Use the equipment health assessment index to trigger the early warning response mechanism, and output the unit start-stop control strategy of the pumping station equipment to the dispatching system through the remote management platform.

[0010] In a second aspect, the present invention provides a pumping station working condition monitoring system based on digital twin. The pumping station working condition monitoring system based on digital twin includes:

[0011] A receiving module, configured to receive the pressure, flow rate, vibration, temperature and electrical energy parameters of the pumping station equipment through the PLC control cabinet to obtain the pumping station operation characteristic matrix;

[0012] A construction module, configured to construct a hybrid digital twin model of the pumping station physical model and the data-driven model based on the pumping station operation characteristic matrix, and generate the operation status prediction data of the pumping station equipment through the hybrid digital twin model;

[0013] A calculation module, configured to input the operation status prediction data and actual monitoring data into the Transformer network to extract the correlation pattern of working conditions characteristics and calculate the quantification of the health index, so as to obtain the equipment health assessment index;

[0014] An output module, configured to use the equipment health assessment index to trigger the early warning response mechanism, and output the unit start-stop control strategy of the pumping station equipment to the dispatching system through the remote management platform.

[0015] In a third aspect, a computer-readable storage medium is provided. Instructions are stored in the computer-readable storage medium. When it runs on a computer, it causes the computer to execute the above-mentioned pumping station working condition monitoring method based on digital twin.

[0016] In the technical solution provided by the present invention, through the construction of a five-layer fusion architecture including a data layer, a feature layer, a matching layer, a scenario layer, and a decision layer, the efficient fusion and processing of multi-source heterogeneous sensing data are realized, effectively solving the problem of data inconsistency in traditional monitoring systems and improving the data foundation quality of pump station condition monitoring. By combining physical models with data-driven models to construct a hybrid digital twin model, it can not only accurately reflect the physical characteristics of pump station equipment but also capture the complex dynamic characteristics of equipment operation through deep learning methods, achieving high-precision prediction of pump station conditions. The spatio-temporal feature analysis module based on the Transformer network architecture can effectively capture the complex correlation relationships and temporal evolution laws among pump station condition parameters, and compared with traditional methods, greatly improves the recognition ability of abnormal conditions and the accuracy of early warning. The hierarchical condition mode matching and health index quantization calculation method of the present invention realizes the accurate assessment of the health status of pump station equipment, can identify potential equipment failures at an early stage, and provides a scientific basis for predictive maintenance. The intelligent early warning and maintenance decision support system of the present invention can automatically generate maintenance decision suggestions and unit start-stop control strategies according to the equipment health assessment index, realizing the goal of "few operators on site, remote monitoring, and low-power operation" for pump stations, greatly reducing operation and maintenance costs, and extending the service life of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0018] Figure 1 Schematic diagram of an embodiment of the pump station condition monitoring method based on digital twin in the embodiment of the present invention;

[0019] Figure 2 Schematic diagram of an embodiment of the pump station condition monitoring system based on digital twin in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] An embodiment of the present invention provides a method, system, and storage medium for monitoring the operating conditions of a pumping station based on digital twins. The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims, and above-mentioned drawings of the present invention are used to distinguish similar objects and do not necessarily need to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order different from that illustrated or described herein. In addition, the term "including" or "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0021] For ease of understanding, the specific process of the embodiment of the present invention will be described below. Please refer to Figure 1 , an embodiment of the method for monitoring the operating conditions of a pumping station based on digital twins in the embodiment of the present invention includes:

[0022] Step S101: Receive the pressure, flow rate, vibration, temperature, and electrical energy parameters of the pumping station equipment through the PLC control cabinet to obtain the pumping station operation characteristic matrix;

[0023] It can be understood that the execution subject of the present invention can be a digital twin-based pumping station operating condition monitoring system, or a terminal or a server, and specific limitations are not made here. The embodiment of the present invention takes the server as the execution subject as an example for illustration.

[0024] Specifically, deploy a distributed sensor network at each key part of the pumping station equipment, where the pumping station equipment includes multiple components such as the pumping station main unit, electrical system, inlet and outlet pipelines, and auxiliary equipment. Different types of sensors are arranged at these parts to ensure coverage of various parameters of the pumping station operation. Install a high-precision piezoresistive pressure transmitter on the pumping station main unit to measure the pressure changes in the pumping station in real time; configure an electromagnetic flowmeter at the inlet and outlet pipelines to ensure accurate measurement of flow data; set a three-axis MEMS accelerometer on key mechanical components to monitor the vibration of the pump body; configure a PT100 platinum resistance temperature sensor in the electrical system to collect the temperature rise of the motor and related components; and install sensors such as voltage, current, power factor, and power quality analyzers to obtain electrical energy parameters. Set the sampling frequency for various sensors to optimize the efficiency and quality of data acquisition. Different types of sensors are set with different sampling frequencies according to their physical characteristics and operating condition change characteristics. For example, the measurement accuracy of the pressure sensor should be no less than 0.1% F.S, and the sampling frequency is set to 100 Hz to capture rapidly changing pressure fluctuations; the measurement accuracy of the flow sensor is no less than 0.5%, and the sampling frequency is set to 10 Hz to ensure the stability of flow data; the measurement range of the vibration sensor is ±16 g, and the sampling frequency is set to 1 kHz to accurately obtain the equipment vibration signal; the measurement accuracy of the temperature sensor is no less than ±0.1 °C, and the sampling frequency is set to 1 Hz to monitor the temperature change trend of the equipment; the sampling frequency of the electrical parameter sensor is set to 50 Hz to ensure the accuracy of power quality analysis; and the measurement accuracy of the ultrasonic water level gauge is no less than ±5 mm, and the sampling frequency is set to 1 Hz to provide stable measurement of the water level. Based on the above sampling frequency configuration, establish a sensor acquisition parameter configuration table. Implement an adaptive sampling strategy based on the sensor acquisition parameters configuration. This strategy dynamically adjusts the sampling frequency of the sensors according to different situations of the pumping station operation state. When the equipment is operating normally and stably, reduce data acquisition redundancy and improve storage and calculation efficiency. For example, the sampling frequency of the pressure sensor is reduced to 20 Hz, and the sampling frequency of the vibration sensor is reduced to 200 Hz; while during the start-stop process of the equipment or when the operation state changes suddenly, increase the sampling frequency to capture key operating condition changes and ensure the integrity and accuracy of the monitoring data. When the system detects that the parameter fluctuation amplitude exceeds the set threshold, it will automatically increase the sampling frequency of the relevant sensors to enhance the data acquisition quality at critical moments. All adaptive sampling data is transmitted to the PLC control cabinet through the fieldbus to form the original sensing data. Perform standardization processing on the original sensing data, including multiple steps such as unit conversion, range normalization, timestamp calibration, and outlier screening. During the unit conversion process, unify the data of different types of sensors to the standard measurement unit to eliminate the differences brought by the measurement unit.In the range normalization process, the Min-Max normalization method is adopted to map all sensor data to the interval [0, 1] to reduce the influence of different dimensions on data analysis. In the timestamp calibration step, based on the high-precision clock synchronization mechanism, data with different sampling frequencies are aligned to a unified time series to ensure the temporal consistency of the data. And outlier screening is carried out. The median filter and the 3σ principle are used to identify and remove abnormal data such as mutation points and interference signals. For some data missing cases caused by communication failures, a combination of linear interpolation and forward filling is used to complete data filling to improve the integrity and reliability of the data. After completing the data standardization process, all standardized data are packaged into data packets in a unified format. Each data packet contains a sensor identifier, a timestamp, a measurement value, a quality identifier, and a device operation status flag to ensure the consistency and traceability of the data. These data packets form a multi-source heterogeneous sensing data set. Heterogeneous data fusion and working condition feature extraction are performed on the multi-source heterogeneous sensing data set to obtain the operation feature matrix of the pumping station.

[0025] Construct a five - layer fusion architecture including a data layer, a feature layer, a matching layer, a scenario layer, and a decision layer as the data fusion processing framework. In this architecture, the data layer receives and pre - processes various sensor data from a distributed sensor network to ensure the quality and temporal consistency of the data; the feature layer performs data transformation and feature extraction for subsequent analysis; the matching layer is used for alignment and association between different types of data, enabling effective integration of redundant data and establishing reasonable parameter associations for physically related data; the scenario layer dynamically adjusts the data fusion method according to the operating state of the pumping station, so that data analysis can adapt to different working conditions; the decision layer conducts health assessment and control optimization based on the fused data. During the data fusion process, analyze the historical reliability indicators and current data quality identifiers of each sensor to calculate the dynamic weight allocation result. The historical reliability indicator is based on the long - term monitoring data of the sensor, and evaluates its reliability level under different working conditions by calculating parameters such as its data variance, outlier ratio, and data missing rate, while the current data quality identifier determines the credibility of the current data based on characteristics such as the volatility and measurement consistency of real - time data. To synthesize these factors, a weighted formula is used to calculate the fusion weight of each sensor, where the weight of each sensor is determined by its historical error rate, current data quality coefficient, and sensor importance coefficient, so that in the data fusion process, sensor data with higher reliability occupies a larger weight, while sensor data with lower reliability or poor data quality is given a smaller weight, and even excluded in specific cases. According to the dynamic weight allocation result, weighted average processing is performed on redundant sensor data that measures the same physical quantity in a multi - source heterogeneous sensor dataset to reduce measurement errors and improve data stability. For example, the pressure values measured by multiple pressure sensors are affected by the equipment operating state, the accuracy of the sensors themselves, and environmental interference. The weighted average method is used to fuse them so that the final pressure measurement value can more accurately reflect the true operating condition of the equipment. For sensor data that measures different physical quantities but has a physical association, a parameter association model is established based on the characteristic curve of the pumping station equipment, and the least - squares fitting method is used for data optimization. For example, there is an inherent physical relationship between flow rate, pressure, and motor power. Therefore, a multi - variable regression equation is established through the hydraulic model and motor model of the pumping station, and the least - squares method is used to calculate the optimal parameter estimation value to obtain the preliminary fused data. After the data fusion is completed, time - domain features and frequency - domain features are extracted from the preliminary fused data to construct a full - feature set.In the process of time-domain feature extraction, various statistical features such as the mean, standard deviation, peak value, peak-to-peak value, waveform factor, impulse factor, and margin factor of the data are calculated to reflect the basic variation law of the data; in the process of frequency-domain feature extraction, the fast Fourier transform is used to convert the time-domain data into frequency-domain data, and features such as power spectral density, main frequency component, and harmonic energy are extracted to reveal the periodic variation characteristics of the data; for vibration signals, non-linear features such as wavelet packet energy entropy and empirical mode decomposition energy moment are extracted through wavelet transform to capture the potential features of equipment faults. Based on the current operating conditions of the pumping station equipment, feature selection is performed on the full feature set to obtain the most discriminative operating condition-related features. In this process, according to the feature requirements of different operating states, an adaptive feature selection strategy is adopted. For example, during the equipment startup process, features such as the rotational speed change rate, vibration amplitude change, and flow establishment time are focused on, while during the stable operation stage, features such as efficiency indicators, harmonic components, and temperature rise conditions are mainly concerned. To improve the scientific nature of feature selection, methods such as information gain, mutual information, and LASSO regression are used for feature screening to remove redundant features and ensure that the selected features can effectively reflect the operating conditions of the equipment. Principal component analysis is performed on the operating condition-related features to eliminate redundant information and reduce the data dimension, and finally the pumping station operation feature matrix is obtained. In the process of principal component analysis, the covariance matrix between features is calculated and its eigenvalue decomposition is performed to obtain the principal component direction; according to the principle of cumulative contribution rate, the principal components that can explain more than 95% of the data variance are selected, and the original features are projected into the principal component space to obtain the reduced-dimension feature matrix.

[0026] Step S102, construct a hybrid digital twin model of the pumping station physical model and the data-driven model based on the pumping station operation feature matrix, and generate the operation state prediction data of the pumping station equipment through the hybrid digital twin model;

[0027] Specifically, a physical model framework of the pumping station is established based on the principles of fluid mechanics. This physical model consists of three core parts, namely, the hydraulic model, the mechanical model, and the electrical model. The hydraulic model is based on the Bernoulli equation and the continuity equation, describes the flow-head characteristics of the pumping station, and determines the model parameters by fitting the actual operation data, so as to accurately describe the hydraulic performance of the pump body. The mechanical model is used to depict the dynamic response characteristics of mechanical components, and the differential equations of the spring-mass-damper system are used to simulate the vibration behavior of components such as the motor rotor, coupling, and impeller, and the finite element analysis method is combined to optimize the dynamic characteristics of each component. The electrical model is used to simulate the operating states of key electrical equipment such as motors, frequency converters, and soft-start devices, establishes a mathematical description of voltage, current, and power changes based on the equivalent circuit theory, and combines the control logic to achieve accurate modeling of the dynamic response of the motor under different operating conditions, thereby constructing the physical model framework of the pumping station. While the physical model is being established, a data-driven deep learning network architecture is constructed according to the operation characteristic matrix of the pumping station. Among them, the bidirectional long short-term memory network (Bi-LSTM) is used to capture the time-series change characteristics of the pumping station conditions. This network architecture includes an encoder-decoder structure. The encoder part consists of three layers of Bi-LSTM units, with 128 neurons in each layer to handle the long-term dependence of time-series data. The decoder part consists of two layers of LSTM units, with 96 neurons in each layer to predict the future changes in condition parameters, and a Dropout layer is added after each layer of LSTM to prevent overfitting, where the Dropout rate is set to 0.2. During the network training process, the weighted sum of the root mean square error (RMSE) and the mean absolute percentage error (MAPE) is used as the loss function, and the Adam optimizer is used for gradient update. The initial learning rate is set to 0.001, and a learning rate decay strategy is adopted, decaying to 0.9 times the original every 50 training epochs to improve the convergence stability of the network. The physical model framework of the pumping station is combined with the deep learning network architecture, and the digital twin technology is used for joint modeling. Among them, the operation characteristic matrix of the pumping station is used as the input data, and at the same time, the data-driven Bi-LSTM network is used to optimize the parameters in the physical model, so that it can better fit the actual working conditions, thereby constructing the first digital twin model. In this process, the deep learning model is pre-trained with the initial condition simulation data provided by the physical model to make it have basic fluid, mechanical, and electrical characteristic constraints, and then the historical operation data is used to further optimize the network parameters, so that the physical model and the data-driven model can complement each other and jointly improve the prediction accuracy. As the model training progresses, the first digital twin model can gradually improve the prediction ability of the operating state of the pumping station equipment and generate preliminary prediction data. In order to improve the accuracy and generalization ability of the model, the prediction data of the first digital twin model is continuously compared with the actual operation data, and the model structure is optimized based on the error feedback mechanism.When the prediction errors of consecutive multiple time windows exceed a set threshold (e.g., MAPE > 5%), the transfer learning method is automatically activated to fine-tune the model. Specifically, the underlying feature extraction part of the Bi-LSTM network is frozen, and only the parameters of the upper-layer prediction part are updated, enabling the model to adjust to the latest operation data while maintaining the general feature extraction ability, resulting in the second digital twin model. This method ensures the long-term stability of the model and avoids the performance degradation problem caused by data drift. Uncertainty quantification calculation is performed on the second digital twin model to improve the reliability of the model prediction results. The Monte Carlo Dropout method is adopted to calculate the confidence interval of the prediction results through multiple forward propagations, thereby quantifying the uncertainty of the model. This method effectively identifies the credibility of the prediction data and accordingly establishes a hybrid digital twin model with reliability assessment, which not only includes the prediction value itself but also the confidence interval of the prediction value, enabling subsequent decisions to be adjusted based on the uncertainty level of the data. For example, when the confidence interval of the prediction results of certain key parameters is too large, it indicates that there is a large uncertainty in the model's prediction of these parameters. At this time, additional data collection or model update is triggered to improve the prediction accuracy. The pump station operation feature matrix is input into the hybrid digital twin model to calculate the predicted values and their confidence intervals of each key parameter in the future time period, obtaining the operation state prediction data of the pump station equipment.

[0028] Step S103: Input the operation state prediction data and the actual monitoring data into the Transformer network to extract the working condition feature correlation pattern and calculate the health index quantification, obtaining the equipment health assessment index;

[0029] Specifically, align the timestamp of the running state prediction data with the actual monitoring data, and use methods such as difference calculation, ratio calculation, and sliding window differentiation to extract differential features to capture the evolution trend of the prediction error over time, obtaining the target input data. This target input data contains the direct difference information between the prediction data and the actual data, as well as the trend change rate, error accumulation effect, and fluctuation characteristics of key parameters. Perform attention weight calculation on the target input data. Use the multi-head self-attention mechanism to process the features of the input data, where the query matrix, key matrix, and value matrix are obtained through linear transformation, and calculate the attention distribution. To ensure that the model can simultaneously focus on the feature patterns of multiple subspaces, 8 different self-attention heads are calculated in parallel. Each attention head independently calculates the weights and then they are concatenated, and the multi-head attention representation is obtained through linear transformation. The multi-head attention mechanism can effectively mine the complex dependence relationships between different physical variables in the data and capture the key operating condition features at different scales, improving the accuracy of the operating condition feature modeling. Enhance the temporal relationship of the multi-head attention representation to better understand the evolution trend of the equipment operating state. Introduce a relative position encoding scheme, which can more effectively capture the dynamic change characteristics of the data over time compared to the traditional absolute position encoding. The relative position encoding learns the relationship weights between different time steps, enabling the model to not only focus on the current state but also establish feature associations over different time spans, thereby more accurately modeling the equipment operating trajectory. On this basis, construct a parameter correlation graph, where the nodes of the graph represent different monitoring parameters, and the weights of the edges represent the correlation strength between the parameters. The construction of the parameter correlation graph depends on the comprehensive calculation of the Pearson correlation coefficient and the mutual information content. The Pearson correlation coefficient is used to measure the linear dependence relationship, while the mutual information content can capture the non-linear correlation. When constructing the parameter correlation graph, dynamically adjust the correlation strength to ensure that the model can adapt to the parameter relationships under different operating states. For example, when the pumping station is operating at a high load state, the correlation between flow rate, pressure, and motor power increases, while it weakens under a low load state. Based on the feature representation enhanced by relative position encoding, combine the parameter correlation graph to calculate the parameter correlation features related to the operating condition. Input the parameter correlation features related to the operating condition into a Transformer deep network composed of 6 layers of encoders. Each layer of the encoder in this network contains a multi-head self-attention sublayer and a feed-forward neural network sublayer, and residual connections and layer normalization are used to ensure gradient stability and model convergence. The multi-head self-attention sublayer is used to calculate the dynamic weight distribution between the key operating condition features, enabling the model to accurately identify the key factors affecting the equipment health state, while the feed-forward neural network sublayer is used for non-linear feature transformation to improve the model's expression ability for complex operating condition patterns. Through the layer-by-layer processing of the 6 layers of encoders, a high-dimensional operating condition representation is obtained, which contains the key features of the pumping station at different times and different operating states and can effectively integrate the deviation information between the prediction data and the actual monitoring data.After obtaining the high-dimensional operating condition representation, in order to reduce the data dimension and improve the calculation efficiency, the principal component analysis method is used for dimensionality reduction. Calculate the covariance matrix of the operating condition features and perform eigenvalue decomposition on it to obtain the most representative principal component direction; according to the principle of cumulative contribution rate, select the principal components that can explain more than 95% of the data variance, and map the original high-dimensional features to the low-dimensional feature space. Map the current operating condition point to the low-dimensional feature space and connect the historical operating condition points to form an evolution trajectory, which can reflect the long-term change trend of the pump station operation state. In order to distinguish normal operating conditions from abnormal operating conditions, the low-dimensional feature space is divided into a normal operation area, a attention area, and an abnormal area. The normal operation area corresponds to the operating conditions of the equipment running stably for a long time in history, the attention area indicates that the operation state of the equipment has a slight deviation but has not reached a serious failure state, while the abnormal area indicates that the equipment operation has entered the failure range and immediate intervention measures need to be taken. In order to quantitatively evaluate the abnormality degree of the current operating condition, the Mahalanobis distance is used to calculate the distance between the current operating condition point and the historical typical operating conditions, and an abnormality metric value is defined according to the size of this distance to form a characteristic map of the pump station operation state. Based on the characteristic map of the pump station operation state, hierarchical operating condition pattern matching and health index quantification calculation are carried out to obtain the equipment health assessment index. In the process of operating condition pattern matching, a support vector machine classifier is used to preliminarily classify the current operating condition to determine which basic operating condition types it belongs to, such as stable operation, load fluctuation, abnormal vibration, or sudden failure; the dynamic time warping method is used to calculate the similarity between the current operating condition trajectory and the historical typical operating condition trajectory, and based on the principle of minimum distance matching, the historical pattern that best matches the current operating condition is identified. On this basis, the equipment health assessment index is calculated. This index consists of five sub-indices, including the efficiency index, vibration index, temperature rise index, electrical index, and stability index. Each sub-index calculates the deviation score by comparing the current monitoring parameters with the standard parameters under the corresponding operating conditions, and the total health index is calculated by the method of weighted fusion, where the weight coefficients are optimized and determined by the genetic algorithm. According to the calculation results, the health assessment index is divided into five grades: excellent, good, general, attention, and warning, and combined with the prediction data to evaluate the evolution trend of the equipment health state in the future for a period of time to provide forward-looking maintenance decision support.

[0030] Build a three - layer operating condition pattern library including a basic operating condition layer, a composite operating condition layer, and an abnormal operating condition layer, which serves as a benchmark template for operating condition identification. In this pattern library, the basic operating condition layer defines four basic operating states of the pump station, including the startup process, stable operation, variable operating condition operation, and shutdown process. Each operating condition corresponds to different operating parameter characteristics. For example, during the startup process, the focus is on the flow establishment time and current impact; during stable operation, the focus is on efficiency stability and vibration level; during variable operating condition operation, it involves load fluctuations and speed changes; and during the shutdown process, the pressure decay and vibration disappearance are monitored. The composite operating condition layer further expands the basic operating conditions by combining different operating conditions to form more complex operating condition patterns, such as operating states under different load levels and transition operating conditions caused by different start - stop strategies (such as fast start - stop or slow start - stop). The abnormal operating condition layer is used to identify potential equipment failure modes, covering typical failure operating conditions such as bearing abnormalities, impeller wear, motor overheating, and water hammer phenomena. Among them, bearing abnormalities are mainly manifested as an increase in high - frequency vibration components, impeller wear leads to a decrease in flow efficiency, motor overheating is accompanied by an abnormal increase in the temperature index, and the water hammer phenomenon will cause a sharp pressure fluctuation in a short time. Based on this operating condition pattern library, a benchmark template for operating condition identification is formed to support subsequent operating condition matching and health assessment calculations. Perform the first - level matching on the operating state characteristic spectrum of the pump station to determine the current basic operating condition type. Input the characteristic spectrum into a support vector machine classifier, which has been trained on standard samples in the operating condition pattern library and can calculate the probability distribution of the current operating state belonging to the four basic operating conditions according to the input characteristics. The support vector machine transforms the input data into a high - dimensional feature space through a kernel function mapping and realizes the distinction of different operating condition categories by constructing an optimal classification hyperplane. During the classification process, by calculating the distance from the input feature points to the hyperplanes of different operating condition categories and combining probability estimation methods, the probability distribution of the four basic operating conditions is obtained, thereby determining the most likely current operating state of the pump station and obtaining the identification result of the basic operating condition type. Perform the second - level matching on the characteristic spectrum to identify specific operating modes or abnormal operating conditions. In this process, the dynamic time warping (DTW) algorithm is used to calculate the similarity between the current operating condition sequence and each standard operating condition sequence in the operating condition pattern library. The DTW algorithm measures the optimal matching path between two time series by dynamically adjusting the alignment of the time axis, thus effectively handling the non - linear time changes of operating condition states. For example, under different loads, the flow, pressure, and vibration signals of the pump station will show different change patterns, and DTW can automatically align these signals in the time dimension to ensure the accuracy of the matching. By calculating the DTW distance between the current operating condition sequence and each standard operating condition pattern and selecting the operating condition pattern with the smallest distance as the final matching result, accurate operating condition identification can be achieved, which can not only distinguish normal operating modes but also detect potential abnormal states.Compare the matching results of specific working conditions with the standard parameters, and calculate the deviation degrees of the efficiency index, vibration index, temperature rise index, electrical index, and stability index respectively. The efficiency index reflects the energy utilization efficiency of the pumping station under the current working conditions and is closely related to the flow rate and head parameters; the vibration index is used to evaluate the mechanical vibration stability of the pump body and is mainly calculated based on the root mean square value, peak value, and spectral energy distribution of the vibration signal; the temperature rise index is used to monitor the temperature of the motor and bearings and evaluate the thermal stability of the equipment; the electrical index is based on electrical parameters such as voltage, current, and power factor to measure the power quality of the motor operation; the stability index evaluates the overall operation smoothness of the pumping station by statistically analyzing the fluctuation range of key parameters. The calculation method of the deviation degree of each sub-index is to normalize and compare the current monitoring value with the standard value, and calculate the deviation degree in combination with historical data. Different scoring weights are assigned to different health states according to the size of the deviation. After calculating the values of each health sub-index, weight optimization is carried out to comprehensively calculate the final equipment health assessment index. The weight optimization adopts an adaptive weight allocation strategy based on the genetic algorithm, in which the weight coefficients of each sub-index are optimized by learning the historical working condition data, so that the final assessment index can accurately reflect the true health status of the equipment. The final health assessment index is calculated by the weighted sum of the five health sub-indices.

[0031] Step S104: Trigger the early warning response mechanism using the equipment health assessment index, and output the unit start-stop control strategy of the pumping station equipment to the dispatching system through the remote management platform.

[0032] Specifically, multi-level warning trigger conditions are constructed, and the device health assessment index is monitored in real time to determine whether to trigger a warning. This multi-level warning mechanism is based on the hierarchical classification of the health index, and divides the device status into five levels: normal, slightly abnormal, generally abnormal, severely abnormal, and critical state. Among them, the normal state corresponds to a health assessment index greater than 85, and no additional measures are required. The slightly abnormal state corresponds to a health index between 70 and 85, indicating that the device has a slight operation deviation but does not affect normal work for the time being. The generally abnormal state corresponds to a health index between 55 and 70, indicating that the device has a slight fault risk and needs to be closely monitored. The severely abnormal state corresponds to a health index between 40 and 55, indicating that the device has a large fault hidden danger and needs to arrange maintenance in advance. The critical state corresponds to a health index lower than 40, meaning that the device is about to have a serious fault or has already had a significant performance decline, and emergency measures must be taken immediately. When the health assessment index changes, it is compared with the preset multi-level warning threshold according to the current index value, and combined with the historical data trend analysis to judge whether the warning trigger condition is met, so as to generate a warning trigger result. After the warning is triggered, the intelligent diagnosis module is immediately started to determine the specific reason for the decrease in the health assessment index and provide corresponding fault analysis. This module uses the feature matching method to retrieve historical fault cases similar to the current device state in the case library and calculates the similarity score between the current fault mode and each case in the case library. The similarity calculation is based on a variety of feature matching methods, including Euclidean distance, dynamic time warping, and cosine similarity, etc., and comprehensively considers the operating environment, load conditions, and historical health data of the device. In this way, the historical fault mode closest to the current working condition is identified, and combined with the diagnostic conclusions in the case library, a fault diagnosis result is obtained, including the fault type, possible fault reasons, influence range, and recommended disposal plan. If the system fails to find a highly matching case in the case library, it will call the rule-based reasoning method, combined with the physical model of the pumping station equipment and expert experience rules, to deduce possible fault reasons to ensure the comprehensiveness and accuracy of the diagnosis. After obtaining the fault diagnosis result, an optimal maintenance strategy is formulated based on this information. For this purpose, a decision-making framework with a double deep Q-network structure is introduced. This framework uses reinforcement learning to continuously optimize the decision-making strategy through continuous optimization to ensure that the maintenance decision of the device can balance cost-effectiveness and operation reliability. This framework consists of an online network and a target network. The online network is responsible for selecting maintenance actions, while the target network is used to evaluate the value of maintenance actions. In the decision-making process, the fault diagnosis result, device health assessment index, maintenance history, and operation plan are used as state inputs, and a weighted combination of maintenance cost, downtime loss, device reliability, and remaining life is used as the objective function to optimize the maintenance decision of the device.Through the reinforcement learning algorithm, the system continuously adjusts the weights of each maintenance option, selects the maintenance strategy with the highest long-term benefit, and obtains the optimal maintenance decision, including when to perform maintenance, which maintenance tasks need to be executed, and how to reduce maintenance costs and downtime. The optimal maintenance strategy is refined to generate maintenance decision suggestions, and combined with the unit start-stop optimization strategy to form the final joint control decision. The maintenance decision suggestions need to cover multiple key elements, including maintenance timing, maintenance items, maintenance resource requirements, and expected effect evaluation. Among them, the maintenance timing is determined by the decision framework to perform maintenance before the equipment status further deteriorates, while the maintenance items are arranged according to the fault diagnosis results and historical maintenance records to ensure that all potential fault points can be processed in a timely manner. At the same time, the calculation of maintenance resource requirements involves the optimization of labor, materials, and time costs to ensure the efficient execution of maintenance tasks, and the expected effect evaluation is used to predict the change trend of the equipment health index after maintenance to measure the effectiveness of maintenance. At the same time, based on the fault diagnosis results of the current working conditions, optimization suggestions for unit start-stop are generated. For example, in some cases, if the equipment health index continues to decline but has not reached the maintenance trigger threshold, by adjusting the operation mode of the pumping station unit, such as reducing the load, switching the operating unit, or optimizing the start-stop strategy, to reduce the further wear of the equipment and extend the buffer time before the fault occurs. In case of severe anomalies or critical states, some units need to be stopped immediately and standby units need to be started to prevent the fault from spreading to the entire system and ensure the continuous operation ability of the pumping station. After the joint control decision is completed, relevant instructions are sent to the dispatching system through the remote management platform to execute the final unit start-stop control strategy. All maintenance suggestions and start-stop control strategies are encapsulated in a standardized data format and transmitted through the standard data interface of the remote management platform to ensure that the dispatching system can seamlessly receive and parse the control instructions. At the same time, a visual monitoring report is generated, including information such as the current health index, fault diagnosis details, maintenance suggestions, and start-stop control strategies, so that the operation and maintenance personnel can clearly understand the equipment status and perform manual intervention if necessary. By executing the unit start-stop control strategy, the system can maximize the equipment life, reduce the failure rate, and optimize the maintenance cost without affecting the overall operation of the pumping station.

[0033] In the embodiments of the present invention, by constructing a five - layer fusion architecture including a data layer, a feature layer, a matching layer, a scenario layer, and a decision layer, the efficient fusion and processing of multi - source heterogeneous sensing data are realized, effectively solving the problem of data inconsistency in traditional monitoring systems and improving the data quality of pump station condition monitoring. By combining physical models with data - driven models, a hybrid digital twin model is constructed, which can not only accurately reflect the physical characteristics of pump station equipment but also capture the complex dynamic characteristics of equipment operation through deep learning methods, realizing high - precision prediction of pump station conditions. The spatio - temporal feature analysis module based on the Transformer network architecture can effectively capture the complex correlation relationships and temporal evolution laws among pump station condition parameters, greatly improving the ability to identify abnormal conditions and the accuracy of early warning compared with traditional methods. The hierarchical condition pattern matching and health index quantification calculation method of the present invention realizes the accurate assessment of the health status of pump station equipment, can identify potential equipment failures at an early stage, and provides a scientific basis for predictive maintenance. The intelligent early warning and maintenance decision - making support system of the present invention can automatically generate maintenance decision suggestions and unit start - stop control strategies according to the equipment health assessment index, realizing the goal of "few operators on site, remote monitoring, and low - consumption operation" of the pump station, greatly reducing the operation and maintenance costs, and extending the service life of equipment.

[0034] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0035] Deploy a distributed sensor network at the pump station equipment, where the pump station equipment includes a pump station main unit, an electrical system, inlet and outlet pipelines, and auxiliary equipment;

[0036] Based on the distributed sensor network, set the sampling frequencies of pressure sensors, flow sensors, vibration sensors, temperature sensors, and electrical energy parameter sensors to obtain sensor acquisition parameter configurations;

[0037] Based on the sensor acquisition parameter configurations, perform adaptive sampling to obtain adaptive sampling data, and transmit the adaptive sampling data to the PLC control cabinet through the field bus to obtain the original sensing data;

[0038] Perform unit conversion, range normalization, timestamp calibration, and outlier screening on the original sensing data to obtain standardized processed data;

[0039] Package the standardized processed data into a unified format data packet including sensor identifiers, timestamps, measured values, quality identifiers, and equipment operation status markers to obtain a multi - source heterogeneous sensing data set;

[0040] Perform heterogeneous data fusion and condition feature extraction on the multi - source heterogeneous sensing data set to obtain a pump station operation feature matrix.

[0041] Specifically, a distributed sensor network needs to be deployed at the pump station equipment. The pump station equipment mainly includes four core components: the pump station main unit, the electrical system, the inlet and outlet pipelines, and auxiliary equipment. The pump station main unit is the core part of the entire system, responsible for the lifting and transportation of water bodies; the electrical system is responsible for providing power and control for the pump station main unit; the inlet and outlet pipelines connect the water source and the transportation target; the auxiliary equipment includes support equipment such as the cooling system and the lubrication system. Different types of sensors are arranged at these key positions to form an all-round monitoring network. A high-precision piezoresistive pressure transmitter is installed on the pump station main unit to measure the pressure changes in the pump station in real time. These pressure data directly reflect the working status and efficiency of the water pump; an electromagnetic flowmeter is configured at the inlet and outlet pipelines to monitor the fluid flow rate and ensure accurate measurement of the flow data; a three-axis MEMS accelerometer is set on the mechanical components to monitor the vibration of the pump body. The vibration data is an important indicator for judging mechanical failures; a PT100 platinum resistance temperature sensor is configured in the electrical system to collect the temperature rise of the motor and related components; at the same time, sensors such as voltage, current, power factor, and power quality analyzers are installed to obtain electrical energy parameters.

[0042] After the sensor network is deployed, the sampling frequencies of various sensors need to be set. The measurement accuracy of the pressure sensor is not less than 0.1% F.S, and the sampling frequency is set to 100 Hz to capture fast-changing pressure fluctuations; the measurement accuracy of the flow sensor is not less than 0.5%, and the sampling frequency is set to 10 Hz; the measurement range of the vibration sensor is ±16 g, and the sampling frequency is set to 1 kHz; the measurement accuracy of the temperature sensor is not less than ±0.1 °C, and the sampling frequency is set to 1 Hz; the sampling frequency of the electrical parameter sensor is set to 50 Hz. Different sensors use different sampling frequencies based on the consideration of the change characteristics of each physical quantity to ensure that the change laws of each parameter can be accurately captured.

[0043] Based on the sensor-collected parameter configuration, an adaptive sampling strategy is implemented. Adaptive sampling refers to dynamically adjusting the sampling frequency of the sensors according to different situations of the pump station operation status. When the pump station is in the start-stop process or the operation status changes suddenly, the system automatically increases the sampling frequency to capture key operating condition changes. For example, when it is detected that the pressure fluctuation amplitude exceeds the set threshold, the sampling frequency of the pressure sensor will increase from 20 Hz during normal operation to 100 Hz, and the sampling frequency of the vibration sensor will increase from 200 Hz to 1 kHz. This adaptive mechanism can reduce the data storage and transmission burden while ensuring data quality.

[0044] All the data obtained by adaptive sampling are transmitted to the PLC control cabinet via the fieldbus to form the original sensing data. These original data need to be standardized, including unit conversion, range normalization, timestamp calibration, and outlier screening. Unit conversion unifies the data of different sensors into standard measurement units; range normalization uses the Min-Max normalization method to map all sensor data to the interval [0, 1], reducing the impact of different dimensions on data analysis; timestamp calibration ensures the temporal consistency of data through a high-precision clock synchronization mechanism; outlier screening uses median filtering and the 3σ principle to identify and remove abnormal data, and linear interpolation and forward filling are combined to complete the data missing situation.

[0045] The standardized data are packed into data packets in a unified format. Each data packet contains a sensor identifier, a timestamp, a measurement value, a quality identifier, and a device operation status flag, forming a multi-source heterogeneous sensing data set. Next, heterogeneous data fusion and operating condition feature extraction are performed on these data sets to obtain the pump station operation feature matrix. This process uses a five-layer fusion architecture, including a data layer, a feature layer, a matching layer, a scenario layer, and a decision layer. The data layer receives and preprocesses sensor data; the feature layer performs data conversion and feature extraction; the matching layer realizes the alignment and association between different types of data; the scenario layer dynamically adjusts the fusion method according to the operation status of the pump station; the decision layer performs health assessment based on the fused data. For example, when the vibration sensor of a certain pump station detects that the vibration value of a certain bearing suddenly rises from the normal value of 0.5 mm / s to 2.8 mm / s, the system automatically increases the sampling frequency of this sensor from 200 Hz to 1 kHz, and at the same time triggers the increase in the sampling frequency of the relevant temperature sensor. These data are integrated and analyzed through heterogeneous data fusion. The system identifies the correlation between the increase in vibration amplitude and the increase in energy in the 6 kHz frequency band in the spectrum, and combines the temperature rise trend to form a characteristic pattern indicating a potential bearing failure. These features are extracted and dimension-reduced through principal component analysis to form the pump station operation feature matrix, providing high-quality data input for the subsequent digital twin model, enabling the system to accurately predict the equipment status and timely send warning information.

[0046] In a specific embodiment, the process of performing heterogeneous data fusion and operating condition feature extraction on the multi-source heterogeneous sensing data set to obtain the pump station operation feature matrix may specifically include the following steps:

[0047] Construct a five-layer fusion architecture including a data layer, a feature layer, a matching layer, a scenario layer, and a decision layer to obtain a data fusion processing framework;

[0048] Use the data fusion processing framework to analyze and calculate the historical reliability indicators and current data quality identifiers of each sensor to obtain the dynamic weight allocation result;

[0049] According to the dynamic weight allocation results, the redundant sensor data measuring the same physical quantity in the multi-source heterogeneous sensor dataset is weighted and averaged, and for the sensor data measuring different physical quantities but having physical associations in the multi-source heterogeneous sensor dataset, a parameter association model is established based on the characteristic curve of the pumping station equipment for least squares fitting to obtain the preliminary fusion data;

[0050] Extract the time-domain features and frequency-domain features from the preliminary fusion data to obtain the full feature set, and perform feature selection on the full feature set based on the current operating conditions of the pumping station equipment to obtain the condition-related features;

[0051] Perform principal component analysis on the condition-related features to obtain the pumping station operation feature matrix.

[0052] Specifically, a distributed sensor network is deployed at the pumping station equipment. The pumping station equipment mainly includes four core components: the pumping station main unit, the electrical system, the inlet and outlet pipelines, and the auxiliary equipment. The pumping station main unit is the core part of the whole system, responsible for the lifting and transportation of water bodies, including key components such as the water pump body, the motor, and the bearings; the electrical system is responsible for providing power and control for the pumping station main unit, including the frequency converter, the soft start device, and the control panel, etc.; the inlet and outlet pipelines connect the water source and the transportation target, including the inlet pipe, the outlet pipe, the valve, and the joint, etc.; the auxiliary equipment includes support equipment such as the cooling system, the lubrication system, and the monitoring system, etc. Based on these pumping station equipment, a five-layer fusion architecture including the data layer, the feature layer, the matching layer, the scenario layer, and the decision layer is constructed to form a data fusion processing framework. The data layer is responsible for receiving and preprocessing various sensor data from the distributed sensor network to ensure the quality and timing consistency of the data; the feature layer performs data conversion and feature extraction, converting the original data into a feature representation for analysis; the matching layer is used for alignment and association between different types of data, enabling effective integration of redundant data and establishing reasonable parameter associations for physically related data; the scenario layer dynamically adjusts the data fusion method according to the operating state of the pumping station, enabling data analysis to adapt to different operating conditions; the decision layer performs health assessment and control optimization based on the fused data and outputs the final decision result.

[0053] Using this data fusion processing framework, analyze and calculate the historical reliability index and the current data quality identifier of each sensor to obtain the dynamic weight allocation result. The historical reliability index is based on the long-term monitoring data of the sensor, and by calculating parameters such as its data variance, the proportion of outliers, and the data missing rate, etc., to evaluate its reliability level under different operating conditions; the current data quality identifier is based on characteristics such as the volatility and measurement consistency of the real-time data to judge the credibility of the current data. Considering these factors, the following formula is used to calculate the fusion weight of each sensor:

[0054] W sensor =α·(1 - E hist ) + β·Qcurr +γ·I sens

[0055] Wherein, W sensor represents the dynamic weight value of the sensor; E hist represents the historical error rate of the sensor, ranging from 0 to 1, and the lower the value, the more reliable the historical performance; Q curr represents the current data quality coefficient, ranging from 0 to 1, and the higher the value, the better the current data quality; I sens represents the sensor importance coefficient, reflecting the importance of the sensor under specific working conditions; α, β, and γ are the weight coefficients of the three factors respectively, and satisfy α + β + γ = 1.

[0056] According to the calculated dynamic weight distribution result, weighted average processing is performed on the redundant sensor data measuring the same physical quantity in the multi-source heterogeneous sensor dataset. For example, for multiple pressure sensors measuring the pressure at the same location, its fusion value is calculated as:

[0057]

[0058] Wherein, P fused is the fused pressure value, P k is the measured value of the k-th pressure sensor, W k is the dynamic weight of this sensor, and K is the total number of sensors. At the same time, for sensor data measuring different physical quantities but having physical associations, a parameter association model is established based on the characteristic curve of the pumping station equipment for least squares fitting. For example, there is an inherent physical relationship between flow rate, pressure, and motor power. A multi-variable regression equation is established through the hydraulic model, motor model, etc. of the pumping station, and the optimal parameter estimation value is calculated using the least squares method to obtain the preliminary fusion data.

[0059] The time-domain features and frequency-domain features of the preliminary fusion data are extracted to obtain the full feature set. The time-domain features include various statistical features such as the mean, standard deviation, peak value, peak-to-peak value, waveform factor, pulse factor, and margin factor of the data, which reflect the basic change law of the data; the frequency-domain features convert the time-domain data into frequency-domain data through fast Fourier transform and extract features such as power spectral density, main frequency component, and harmonic energy to reveal the periodic change characteristics of the data. For vibration signals, non-linear features such as wavelet packet energy entropy and empirical mode decomposition energy moment are also extracted through wavelet transform to capture potential features of equipment faults.

[0060] Based on the current operating conditions of the pump station equipment, feature selection is performed on the full feature set to obtain the most discriminative operating condition-related features. According to the feature requirements of different operating states, an adaptive feature selection strategy is adopted. During the equipment startup process, features such as the rotational speed change rate, vibration amplitude change, and flow establishment time are mainly concerned; during the stable operation stage, features such as efficiency indicators, harmonic components, and temperature rise conditions are mainly concerned. Methods such as information gain, mutual information, and LASSO regression are used for feature screening to remove redundant features and ensure that the selected features can effectively reflect the operating conditions of the equipment.

[0061] Finally, principal component analysis is performed on the operating condition-related features to eliminate redundant information and reduce the data dimension, obtaining the pump station operation feature matrix. During the principal component analysis process, the covariance matrix between features is calculated, and eigenvalue decomposition is performed on it to obtain the principal component direction; according to the principle of cumulative contribution rate, the principal components that can explain more than 95% of the data variance are selected, and the original features are projected into the principal component space to obtain the feature matrix after dimension reduction.

[0062] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0063] Based on the principles of fluid mechanics, a hydraulic model describing the flow-head characteristics of the pump station, a mechanical model describing the response characteristics of mechanical components, and an electrical model describing the characteristics and control logic of electrical equipment are established to obtain the pump station physical model framework;

[0064] According to the pump station operation feature matrix, a bidirectional long short-term memory network is constructed to obtain the deep learning network architecture;

[0065] The pump station physical model framework is combined with the deep learning network architecture, and model training is carried out through the pump station operation feature matrix and digital twin technology to obtain the first digital twin model;

[0066] The training status prediction data of the first digital twin model is continuously compared with the actual operation data, and the model fine-tuning process is started through the transfer learning method. The underlying feature extraction part is frozen and only the parameters of the upper prediction part are updated to obtain the second digital twin model;

[0067] Uncertainty quantification calculation is performed on the second digital twin model to generate the confidence interval of the training status prediction data, and a hybrid digital twin model with reliability assessment is generated according to the confidence interval;

[0068] The pump station operation feature matrix is input into the hybrid digital twin model, and the predicted values and their confidence intervals of each key parameter in the future time period are calculated to obtain the pump station equipment operation status prediction data.

[0069] Specifically, a physical model framework for describing the characteristics of each part of the pumping station equipment is established based on the principles of fluid mechanics. The hydraulic model is based on the Bernoulli equation and the continuity equation, and is used to describe the flow-head characteristics of the pumping station. The Bernoulli equation represents the principle of energy conservation when the fluid moves along the streamline, and the continuity equation represents the principle of mass conservation of the fluid. Through these basic equations, combined with the characteristic parameters of the actual pumping station, a mathematical model for describing the performance of the water pump is established, which can accurately reflect the relationship between the flow rate and the head under different working conditions. The mechanical model uses a system of differential equations of the spring-mass-damper system to simulate the vibration behavior of mechanical components such as the motor rotor, coupling, and impeller. This model regards each mechanical component as an element with mass, elasticity, and damping characteristics, and forms a complete mechanical dynamic response model by establishing the interaction relationship between them, which can predict the vibration characteristics of the mechanical components of the pumping station under different load and speed conditions. The electrical model is based on the equivalent circuit theory, establishes a mathematical description of the voltage, current, and power changes, and combines the control logic to simulate the dynamic response of key electrical equipment such as motors, frequency converters, and soft start devices under different operating conditions.

[0070] Based on the obtained pump station operation characteristic matrix, a Bidirectional Long Short-Term Memory Network (Bi-LSTM) is constructed to form a deep learning network architecture. Bi-LSTM is a special type of recurrent neural network that can consider both the forward and backward information of sequential data, and is particularly suitable for processing time series data and capturing long-term dependencies therein. The Bi-LSTM network adopts an encoder-decoder structure. The encoder part consists of three layers of Bi-LSTM units, with each layer containing 128 neurons, which are used to process the long-term dependencies of time series data; the decoder part consists of two layers of LSTM units, with each layer containing 96 neurons, which are used to predict future changes in operating condition parameters. To prevent overfitting, a Dropout layer is added after each layer of LSTM, and the Dropout rate is set to 0.2. The network training uses the weighted sum of Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) as the loss function, and the Adam optimizer is used for gradient update. The initial learning rate is set to 0.001, and a learning rate decay strategy is used. Every 50 training epochs, the learning rate is decayed to 0.9 times the original value to improve the convergence stability of the network. The physical model framework of the pump station is combined with the deep learning network architecture, and the model is trained through the pump station operation characteristic matrix and digital twin technology to obtain the first digital twin model. Digital twin technology refers to constructing a digital model of a physical entity in a virtual space to achieve real-time mapping, monitoring, and prediction of the physical entity. In this step, the pump station operation characteristic matrix is used as the input data, and the deep learning model is pre-trained using the initial operating condition simulation data provided by the physical model to enable it to have basic fluid, mechanical, and electrical property constraints. Then, historical operation data is used to further optimize the network parameters, so that the physical model and the data-driven model complement each other and jointly improve the prediction accuracy. This process makes full use of the rigor of the physical model based on mechanism and the adaptive learning ability of the data-driven model to form a hybrid model with both interpretability and high accuracy.

[0071] Continuously compare the training status prediction data of the first digital twin model with the actual operation data. When the prediction errors in multiple consecutive time windows exceed the set threshold (e.g., MAPE > 5%), automatically start the transfer learning method to fine-tune the model. Transfer learning is a machine learning method that uses the knowledge already learned to solve new related problems and transfer knowledge between different domains. In this method, the specific operation of transfer learning is to freeze the underlying feature extraction part of the Bi-LSTM network and only update the parameters of the upper prediction part. The underlying feature extraction part learns the basic feature representation of the data and has good generalization ability, while the upper prediction part is more concerned with specific tasks and is more susceptible to data changes. In this way, the model can maintain its general feature extraction ability while adjusting to the latest operation data, obtaining the second digital twin model, ensuring the long-term stability of the model, and avoiding the performance degradation problem caused by data drift.

[0072] Perform uncertainty quantification calculation on the second digital twin model to generate the confidence interval of the training status prediction data. Uncertainty quantification is an important means to evaluate the reliability of model predictions. In this method, the Monte Carlo Dropout method is used to achieve this. Monte Carlo Dropout is a technique that keeps the Dropout layer turned on during the prediction stage. By performing multiple forward propagations, different prediction results are obtained, thereby estimating the uncertainty of the prediction. The specific operation is to perform multiple (e.g., 100 times) forward propagations during the prediction stage, each time keeping the Dropout randomly inactivated, obtaining multiple sets of different prediction results. Then, calculate the mean of these results as the final prediction value, calculate the standard deviation as a measure of uncertainty, and generate a 95% confidence interval. The size of the confidence interval reflects the confidence level of the model in the prediction result, providing an important reference for subsequent decision-making. Based on these confidence intervals, a hybrid digital twin model with reliability assessment is generated. This model not only includes the prediction value itself but also the confidence interval of the prediction value, enabling subsequent decisions to be adjusted according to the uncertainty level of the data. Input the pump station operation feature matrix into the hybrid digital twin model to calculate the predicted values and their confidence intervals of each key parameter in the future time period, obtaining the operation status prediction data of the pump station equipment. These prediction data will serve as an important basis for subsequent working condition assessment and early warning decision-making.

[0073] Taking a large drainage pump station as an example, this pump station is equipped with 4 pumps, and a complete digital twin model is constructed by the above method. First, a hydraulic model is established based on the Bernoulli equation and the continuity equation to describe the flow-head curve of the pump at different rotational speeds. At the same time, a mechanical model is established based on the kinetic principle to simulate the vibration characteristics of components such as the pump shaft and the motor rotor. The electrical model describes the variation laws of current, voltage, and power during the motor startup, operation, and braking processes. These physical models are combined with a bidirectional long short-term memory network trained based on historical operation data to form the first digital twin model. In practical applications, when it is found that the mean absolute percentage error between the flow predicted by the model and the actual flow exceeds 4.8% within 5 consecutive hours, the system automatically starts the transfer learning process, freezes the first two layers of Bi-LSTM units of the network, and only updates the parameters of the latter layer of Bi-LSTM units and the decoder part. After 50 training batches, the prediction error is reduced to 2.3%, forming a more accurate second digital twin model. Then, 100 Monte Carlo Dropout calculations are performed on this model to obtain the 95% confidence interval of the predicted values. The final model can accurately predict the flow rate, pressure, vibration, and electrical energy parameters within the next 24 hours and give the corresponding confidence intervals. For example, the predicted flow rate of pump No. 1 after 12 hours is 3500 cubic meters per hour, and the confidence interval is [3420, 3580] cubic meters per hour.

[0074] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0075] Align the timestamps of the operation state prediction data and the actual monitoring data and calculate the differential features to obtain the target input data;

[0076] Perform attention weight calculation on the target input data, and simultaneously calculate 8 different self-attention heads in parallel and splice the results for linear transformation to obtain the multi-head attention representation;

[0077] Apply the relative position encoding scheme to the multi-head attention representation to obtain the feature representation with enhanced temporal relationship, and construct a parameter correlation graph based on the feature representation with enhanced temporal relationship to obtain the parameter correlation features related to the working conditions;

[0078] Input the parameter correlation features related to the working conditions into a Transformer deep network composed of 6 layers of encoders. Each layer of encoder contains a multi-head self-attention sublayer and a feed-forward neural network sublayer, and apply residual connection and layer normalization to obtain the high-dimensional working condition representation;

[0079] The high-dimensional operating conditions are represented and dimension reduction is performed through the principal component analysis method to obtain a low-dimensional feature space. The current operating condition point is mapped in the low-dimensional feature space and connected to historical operating condition points to form an evolution trajectory. At the same time, the low-dimensional feature space is divided into a normal operation area, a attention area, and an abnormal area. The abnormal metric value is calculated based on the Mahalanobis distance between the current operating condition point and historical typical operating conditions, and a characteristic map of the pump station operating state is obtained.

[0080] Hierarchical operating condition pattern matching and health index quantification calculation are performed on the characteristic map of the pump station operating state to obtain the equipment health assessment index.

[0081] Specifically, timestamp alignment means calibrating data obtained at different times according to a unified time reference to ensure the chronological correspondence relationship between data. The specific operation is to match the prediction data and the actual monitoring data according to the time tags. For possible time deviations in the prediction data, calibration is performed through the linear interpolation method. After completing the time alignment, the difference features of the two sets of data are calculated, including direct difference calculation, ratio calculation, and sliding window difference. The direct difference refers to calculating the difference between the predicted value and the actual value point by point, reflecting the absolute magnitude of the prediction error; the ratio calculation is the ratio of the predicted value to the actual value, reflecting the relative error; the sliding window difference captures the evolution pattern of the prediction error by calculating the change trend of the differences within consecutive time windows. These difference features together constitute the target input data for subsequent deep learning analysis.

[0082] Attention weight calculation is performed on the target input data using the self-attention mechanism. The self-attention mechanism is a technique that can measure the mutual relationships between elements within a sequence. The core idea is to assign different degrees of importance to different elements by learning the internal correlations in the data. In this method, the input data is linearly transformed to generate a query matrix (Query), a key matrix (Key), and a value matrix (Value) respectively. Then, the dot product of the query and the key is calculated, and the attention weights are obtained by normalizing through the softmax function. Finally, the weights are multiplied by the values to obtain the weighted result. To enhance the model's expressive ability, 8 different self-attention heads are calculated in parallel. Each attention head focuses on different aspects of the data, and then the output results of the 8 heads are concatenated and linearly transformed to obtain the multi-head attention representation. This process enables the model to capture the complex relationships between data features from multiple perspectives simultaneously.

[0083] Apply a relative position encoding scheme to the multi-head attention representation to enhance the expression of temporal relationships. Different from traditional absolute position encoding, relative position encoding focuses on the relative distances between elements and is more suitable for processing temporal data. The specific implementation is to learn a set of trainable relative position embedding matrices to encode the distance relationships between different time steps, and then incorporate this position information into the attention calculation. In this way, the model can more effectively capture the dynamic change characteristics of the data over time, not only paying attention to the current state but also being able to establish feature correlations over different time spans. Based on the enhanced temporal feature representation, construct a parameter correlation graph, where nodes represent different monitoring parameters and the weights of the edges represent the correlation strength between the parameters. The correlation strength is obtained through the comprehensive calculation of the Pearson correlation coefficient and the mutual information content. The Pearson correlation coefficient measures the linear dependence relationship, while the mutual information content captures the non-linear correlation. Dynamically adjust the correlation strength under different operating states to ensure that the model can adapt to the changes in the parameter relationships under different working conditions, and finally obtain the parameter correlation features related to the working conditions.

[0084] The parameter correlation features related to the working conditions are input into a Transformer deep network composed of 6 layers of encoders for processing. Transformer is a neural network architecture based on the self-attention mechanism, which performs excellently in processing sequence data. Each layer of the encoder contains a multi-head self-attention sub-layer and a feed-forward neural network sub-layer. The multi-head self-attention sub-layer is responsible for calculating the dynamic weight distribution between key working condition features, enabling the model to accurately identify the key factors affecting the health state of the equipment; the feed-forward neural network sub-layer contains two linear transformations and a non-linear activation function for further processing of the features. Residual connections and layer normalization are applied after each sub-layer. The residual connection solves the problem of gradient disappearance in the training of deep networks by adding the sub-layer input to the sub-layer output; layer normalization improves the training stability by normalizing each feature vector. Through the deep processing of 6 layers of encoders, a high-dimensional working condition representation containing the key features of the pumping station at different times and different operating states is obtained. The high-dimensional working condition representation is processed by the principal component analysis method for dimensionality reduction to obtain a low-dimensional feature space. Principal component analysis (PCA) is a commonly used linear dimensionality reduction technique that reduces the data dimension by finding the main directions of data variation. First, the covariance matrix of the working condition features is calculated, and then the covariance matrix is eigen-decomposed to obtain the eigenvectors representing the main directions of data variation, that is, the principal components. The principal components that can explain more than 95% of the data variance are selected, and the high-dimensional features are mapped into the low-dimensional space spanned by these principal components. In the low-dimensional space, the current working condition point is mapped, and the historical working condition points are connected to form an evolution trajectory, reflecting the change trend of the operating state of the pumping station. At the same time, the low-dimensional feature space is divided into three regions: the normal operation region, the attention region, and the abnormal region. The normal operation region corresponds to the working conditions of the equipment running stably for a long time; the attention region indicates that the equipment state deviates slightly but has not reached a serious fault; the abnormal region indicates that the equipment has entered the fault range.

[0085] Calculate the anomaly metric value according to the Mahalanobis distance between the current working condition point and the historical typical working conditions, and generate the characteristic map of the operating state of the pumping station. The Mahalanobis distance is a distance metric that considers the characteristics of data distribution, and its calculation formula is:

[0086]

[0087] where, D Mah (x, C) represents the Mahalanobis distance from the working condition point x to the category C; μ C is the mean vector of all samples in the category C; Σ C is the covariance matrix of the category C; is the inverse matrix of the covariance matrix. The Mahalanobis distance takes into account the correlation between features and the shape of data distribution, and can more accurately measure the anomaly degree between points and categories in a multi-dimensional space than the Euclidean distance.

[0088] Calculate the anomaly metric value \(A_v\) based on the Mahalanobis distance:

[0089]

[0090] where \(A\) v is the anomaly metric value; \(D\) Mah (\(x, C\) normal ) is the Mahalanobis distance from the current operating point \(x\) to the normal operating condition category; \(D\) threshold is the preset threshold distance, usually determined based on historical data statistics. The larger the value of \(A\) v , the higher the degree of anomaly of the current operating condition.

[0091] Perform hierarchical operating condition pattern matching and health index quantification calculation on the characteristic spectrum of the pump station operating state to obtain the equipment health assessment index. Hierarchical operating condition pattern matching first constructs a three-layer operating condition pattern library: the basic operating condition layer, the composite operating condition layer, and the abnormal operating condition layer. The basic operating condition layer includes four basic operating conditions: start-up process, stable operation, variable operating condition operation, and shutdown process; the composite operating condition layer includes transition operating conditions under different loads and start-stop strategies; the abnormal operating condition layer includes fault operating conditions such as bearing anomalies and impeller wear. Perform the first-level matching through a support vector machine classifier to calculate the probability distribution of the four basic operating conditions; then perform the second-level matching, calculate the distance between the current operating condition sequence and each standard operating condition sequence through the dynamic time warping (DTW) algorithm, and select the operating condition pattern with the smallest distance.

[0092] The health index quantification calculation is based on the following formula:

[0093]

[0094] where \(HI\) total is the total health assessment index; \(HI\) r represents five sub-indicators: efficiency index (\(HI_1\)), vibration index (\(HI_1\)), temperature rise index (\(HI_1\)), electrical index (\(HI_1\)), and stability index (\(HI_1\)); \(\omega\) r is the corresponding weight coefficient, satisfying

[0095] The calculation formulas for each sub-indicator are:

[0096]

[0097] where \(HI\) r is the health value of the \(r\)-th sub-indicator; \(V\) actual,r is the actual value of the \(r\)-th parameter currently monitored; \(V\) standard,r is the standard parameter value under the corresponding operating condition; \(\lambda\) r is the deviation penalty coefficient, reflecting the impact of different degrees of deviation on the health state.

[0098] In a specific embodiment, the process of performing hierarchical working condition pattern matching and health index quantification calculation on the operation state characteristic map of the pump station to obtain the equipment health assessment index may specifically include the following steps:

[0099] Construct a three - layer working condition pattern library including a basic working condition layer, a composite working condition layer, and an abnormal working condition layer. The basic working condition layer includes four basic working conditions: startup process, stable operation, variable working condition operation, and shutdown process. The composite working condition layer includes operating conditions under different loads and transition conditions under different start - stop strategies. The abnormal working condition layer includes fault conditions such as bearing abnormalities, impeller wear, motor overheating, and water hammer phenomenon, to obtain a working condition recognition reference template;

[0100] Input the operation state characteristic map of the pump station into a support vector machine classifier, perform the first - level matching through the working condition recognition reference template, calculate the probability distribution of the four basic working conditions, and obtain the recognition result of the basic working condition type;

[0101] Based on the recognition result of the basic working condition type, perform the second - level matching on the operation state characteristic map of the pump station. By calculating the dynamic time warping distance between the current working condition sequence and each standard working condition sequence in the working condition pattern library, select the working condition pattern with the smallest distance to obtain the specific working condition matching result;

[0102] Compare and analyze the specific working condition matching result with the standard parameters, calculate the deviation scores of five sub - indicators: efficiency index, vibration index, temperature rise index, electrical index, and stability index, to obtain the health sub - index values of the pump station equipment, and optimize the weights of the health sub - index values to obtain the equipment health assessment index.

[0103] Specifically, the basic working condition layer defines four basic operation states of the pump station, including startup process, stable operation, variable working condition operation, and shutdown process. The startup process represents the transition stage of the pump station from a stationary state to normal operation, characterized by the gradual establishment of flow rate from zero, current impact, and large pressure fluctuations; stable operation represents the constant operation state of the pump station under design conditions, characterized by stable flow rate, pressure, and vibration, and the efficiency being in the optimal range; variable working condition operation represents the state where the load or speed of the pump station changes, characterized by parameter fluctuations with the adjustment of the working condition; the shutdown process represents the process of the pump station from the operating state to a complete stop, characterized by a decrease in flow rate, a drop in pressure, and the gradual disappearance of vibration.

[0104] The composite operating condition layer is a more complex operating condition mode formed by combining different operating conditions on the basis of the basic operating conditions, including the operating conditions under different loads and the transition conditions under different start-stop strategies. The operating conditions under different loads cover three situations: low load (30%-50% of the rated load), medium load (50%-80% of the rated load), and high load (80%-100% of the rated load). The parameter characteristics under each load are different; the transition conditions under different start-stop strategies include two methods: rapid start-stop and slow start-stop. Rapid start-stop completes start-up or stop in a short time, while slow start-stop adopts a gradual approach to avoid water hammer and current impact.

[0105] The abnormal operating condition layer includes common pump station fault modes, such as abnormal bearings, impeller wear, motor overheating, and water hammer phenomena. Abnormal bearings are mainly manifested by an increase in the high-frequency vibration component, especially obvious peaks in the bearing natural frequency and its harmonics; impeller wear leads to a decrease in the pump station efficiency, a reduction in flow rate, and an abnormal pressure-power ratio; motor overheating is manifested by a rapid increase in temperature and an increase in current fluctuation; the water hammer phenomenon generates a short-term pressure mutation and pipeline vibration when the pump station starts and stops. Through the definition and parameter characteristic representation of these three layers of operating condition modes, a benchmark template for operating condition identification is formed, providing a reference standard for subsequent operating condition matching.

[0106] Input the characteristic spectrum of the pump station operating state into the support vector machine classifier for the first-level matching. The support vector machine is a supervised learning algorithm that realizes the distinction of different categories by constructing the optimal classification hyperplane in the feature space. In the training stage, first collect the standard samples of each basic operating condition, extract the relevant features to form a training set, and then use the radial basis function (RBF) as the kernel function to map the samples to a high-dimensional space. The optimal classification hyperplane and its support vectors are solved through the convex optimization method. In the classification stage, the input characteristic spectrum is mapped to the feature space through the same mapping function, calculate its distance to different category hyperplanes, and convert these distances into probability values through Platt's method to obtain the probability distribution of the input data belonging to the four basic operating conditions. The operating condition type with the highest probability is identified as the current basic operating condition, forming the recognition result of the basic operating condition type.

[0107] Based on the recognition result of the basic operating condition type, the second-level matching is performed on the operation state characteristic spectrum of the pumping station. The second-level matching uses the Dynamic Time Warping (DTW) algorithm to calculate the similarity between the current operating condition sequence and each standard operating condition sequence in the operating condition pattern library. The DTW algorithm can measure the similarity between two time series, even if they have different lengths or there are stretches on the time axis. DTW finds the optimal alignment path between the two sequences through dynamic programming, making the cumulative distance of the corresponding points on the path the smallest. The specific steps include: First, construct a distance matrix, where each element in the matrix represents the Euclidean distance between the corresponding points of the two sequences; then use the dynamic programming algorithm to calculate the cumulative distance matrix, and each element represents the minimum cumulative distance from the starting point to this point; finally, trace back from the end point to the starting point to find the optimal alignment path. The DTW distance is defined as the sum of the distances of all corresponding points on the optimal path, and the smaller the distance, the more similar the two sequences are. By calculating the DTW distances between the current operating condition sequence and each standard operating condition pattern and selecting the operating condition pattern with the smallest distance as the matching result, the specific operating condition matching result is obtained.

[0108] Compare and analyze the specific operating condition matching result with the standard parameters, and calculate five health score indicators. The efficiency index reflects the energy utilization efficiency of the pumping station. The actual efficiency is calculated through the actually measured flow rate, head, and power, and compared with the standard efficiency under this operating condition. The smaller the deviation, the higher the efficiency index; the vibration index is used to evaluate the mechanical vibration condition, and the deviation is calculated by comparing the difference between the actual root mean square value of vibration and the standard value, combined with the frequency spectrum characteristics; the temperature rise index reflects the thermal state of the equipment, by measuring the temperatures of key parts such as the motor and bearings and comparing with the standard operating temperature; the electrical index evaluates the operating state of the motor, including the deviation of parameters such as voltage, current balance, and power factor; the stability index is determined by calculating the ratio of the fluctuation amplitude of key parameters to the standard fluctuation range. The calculation process of each sub-index is to calculate the deviation between the current measured parameters and the standard parameters, and then map it to the 0 - 100 score interval according to the preset scoring rules to obtain the health score indicator value.

[0109] Finally, optimize the weights of the five health score indicator values to obtain the equipment health assessment index. The weight optimization uses the genetic algorithm, which solves the optimization problem by simulating natural selection and genetic mechanisms. First, define the chromosome encoding, where each chromosome contains five genes, respectively representing the weights of the five sub-indicators; then design the fitness function to evaluate the quality of the weight combination; generate a new generation of population through selection, crossover, and mutation operations, and continuously iterate and optimize to finally obtain the optimal weight combination. Multiply the optimal weights by each health score indicator and sum them to obtain the final equipment health assessment index, which comprehensively reflects the overall health state of the pumping station equipment.

[0110] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0111] Construct multi - level warning trigger conditions, and monitor the equipment health assessment index in real time to obtain the warning trigger result;

[0112] Based on the warning trigger result, start the intelligent diagnosis module, retrieve similar cases in the case base through feature matching to calculate the similarity score, and obtain the fault diagnosis result;

[0113] Input the fault diagnosis result, equipment health assessment index, maintenance history and operation plan into the decision - making framework of the double - deep Q - network structure. Select actions through the online network and evaluate the action value through the target network. Use the weighted combination of maintenance cost, downtime loss, equipment reliability and remaining life as the objective function to obtain the optimal maintenance strategy;

[0114] Refine the optimal maintenance strategy to generate maintenance decision suggestions including maintenance timing, maintenance items, maintenance resource requirements and expected effect evaluation. At the same time, generate unit start - stop suggestions for the current working condition based on the fault diagnosis result to obtain the joint control decision;

[0115] Transmit the joint control decision to the dispatching system through the standard data interface of the remote management platform, and output the unit start - stop control strategy of the pumping station equipment.

[0116] Specifically, in the digital - twin - based pumping station working condition monitoring method, constructing multi - level warning trigger conditions is the key first step to achieve intelligent warning. The multi - level warning mechanism divides the equipment state into five levels: normal, slightly abnormal, generally abnormal, severely abnormal and critical state. The normal state corresponds to a health assessment index greater than 85, indicating that the equipment is in good operating condition and no additional measures are required; the slightly abnormal state corresponds to a health index between 70 and 85, indicating that there are slight operating deviations in the equipment; the generally abnormal state corresponds to a health index between 55 and 70, indicating a slight fault risk; the severely abnormal state corresponds to a health index between 40 and 55, indicating a greater potential for faults; the critical state corresponds to a health index below 40, indicating that the equipment is about to have a serious fault. Monitor the equipment health assessment index in real time, compare it with the multi - level warning thresholds, and analyze it in combination with the change trend of the health index. When it is detected that the decline rate of the health index exceeds the preset threshold, even if the current health index has not fallen below the next - level warning threshold, a warning will be triggered. For example, if the health index drops from 83 to 72 within 24 hours, although it is still in the slightly abnormal range, the decline rate is large, and at this time, a warning at the generally abnormal level will be triggered in advance. Through this dynamic warning mechanism, the warning trigger result is obtained.

[0117] Based on the early warning trigger result, the intelligent diagnosis module is activated, and similar cases are retrieved from the case base through feature matching. The feature matching adopts a retrieval method based on similarity. First, key feature vectors are extracted from the current working condition data, including time-domain features, frequency-domain features, and various health index values. Then, these features are matched with the historical fault cases stored in the case base. The matching process uses a combination of multiple similarity calculation methods, including Euclidean distance, cosine similarity, and dynamic time warping algorithm. Euclidean distance measures the absolute distance of feature vectors in space; cosine similarity measures the directional similarity of feature vectors and is insensitive to data scale; the dynamic time warping algorithm is specifically used to compare the similarity of time series data and can handle sequences of different lengths and different sampling rates. For each historical case, these three similarity metrics are calculated and weighted and summed according to preset weights to obtain a comprehensive similarity score. Sort the cases from high to low according to the similarity score, select several cases with the highest scores as references, and at the same time consider context information such as the equipment operating environment, load conditions, and historical health data, and comprehensively analyze to obtain the most likely fault cause, impact range, and severity, forming a fault diagnosis result.

[0118] The fault diagnosis result, equipment health assessment index, maintenance history, and operation plan are input into the decision-making framework of the double deep Q-network structure to obtain the optimal maintenance strategy. The double deep Q-network is an algorithm in the field of reinforcement learning, which realizes stable value estimation and action selection through two network structures: an online network and a target network. The online network is responsible for selecting maintenance actions according to the current state, and the target network is used to evaluate the value of the selected actions. The two network structures are the same but the parameter update frequencies are different. The parameters of the target network lag behind those of the online network, and the training stability is improved by slow update. During the decision-making process, information such as the fault diagnosis result, equipment health assessment index, maintenance history, and operation plan is encoded into a state vector and input into the double deep Q-network. The network selects the maintenance action with the highest expected return according to experience, such as "repair immediately", "plan to repair at the next shutdown", "continue to run and monitor", etc. The objective function of the maintenance decision is a weighted combination of maintenance cost, downtime loss, equipment reliability, and remaining life. By adjusting the weight coefficients of each factor, the short-term economic benefits and long-term reliability are balanced. The network optimizes the action selection strategy by continuously learning the results and benefits of historical maintenance decisions, and finally outputs the optimal maintenance strategy for the current state.

[0119] Refine the optimal maintenance strategy to generate specific maintenance decision suggestions. The refinement process includes four key elements: determining the maintenance timing, maintenance items, maintenance resource requirements, and evaluation of expected effects. The maintenance timing is determined comprehensively based on the current state of the equipment, the predicted fault development trend, and the operation plan. If it is predicted that the fault will develop into a serious fault in the short term, arrange for maintenance in the near future; if the predicted fault develops slowly, it can be postponed until the planned shutdown for maintenance. The maintenance items are determined based on the fault diagnosis results and historical maintenance records. For the identified potential fault points, a targeted list of maintenance items is formulated. The calculation of maintenance resource requirements is based on the list of maintenance items, estimating the required manpower, materials, tools, and time, and reasonably allocating resources while considering the maintenance difficulty and urgency. The evaluation of expected effects simulates the state of the equipment after maintenance through a digital twin model, predicting the change trend of the health index and performance improvement after maintenance. At the same time, based on the fault diagnosis results, start-stop suggestions for the unit under the current working conditions are generated, such as suggesting reducing the load operation in case of abnormal bearings and suggesting immediate shutdown for cooling in case of overheating of the motor, etc., to form a joint control decision including maintenance and operation adjustment.

[0120] Transmit the joint control decision to the dispatching system through the standard data interface of the remote management platform, and output the start-stop control strategy of the pumping station equipment. The standard data interface uses the OPC UA (OPC Unified Architecture) protocol, which is a widely used communication standard in the field of industrial automation and supports the transmission of complex data structures and semantic information. The remote management platform encapsulates the joint control decision according to a predefined data model, including decision type, priority, execution time, specific operation commands, and additional explanatory information, and sends it to the dispatching system through an encrypted channel. The dispatching system receives and analyzes the control instructions, automatically executes or reminds the operation and maintenance personnel to execute the corresponding control operations according to the instruction content and the current system state, and at the same time feeds back the execution results to the remote management platform to complete the closed-loop control.

[0121] Taking a municipal water supply pumping station as an example, this pumping station adopts a working condition monitoring method based on digital twin to monitor the operation status of equipment in real time. During a monitoring, the health assessment index of Pump No. 1 dropped from 82 to 68, and the downward trend continued, triggering a warning at the general abnormal level. The system automatically activated the intelligent diagnosis module, extracted feature vectors from multiple dimensions such as vibration, temperature, and efficiency, and retrieved similar cases in the case library. By calculating the Euclidean distance, cosine similarity, and dynamic time warping distance, it was found that the current features were most similar to the historical case of "inner ring wear of the bearing", and the similarity score reached 0.87, far higher than the scores of other fault types. The fault diagnosis result showed that there was early wear in the non-driving end bearing inner ring of Pump No. 1, and it was predicted that it would develop into a serious fault after continuous operation for 300 hours. The system input the diagnosis result, health index, maintenance records in the past year (the bearing was replaced last time and has been running for 2,000 hours), and the operation plan for the next week into the double deep Q network. The online network selected the maintenance action of "planned shutdown to replace the bearing within 72 hours" based on this information, and the target network evaluated that the value of this action was higher than that of "immediate shutdown for repair" and "continue running until a fault occurs". After the strategy was refined, it was determined to arrange a 4-hour shutdown for repair after 48 hours, prepare the required bearing model SKF-6308, special disassembly and assembly tools, and two maintenance personnel. It was expected that the health index would recover to over 90 after the repair. At the same time, to control the development of the fault, it was recommended to reduce the load of Pump No. 1 to 70% before the repair and start the No. 2 standby pump to share part of the load. This joint control decision was transmitted to the dispatching system through the OPC UA protocol. After receiving the instruction, the dispatching system automatically adjusted the operation parameters of Pump No. 1 and started the No. 2 pump, and at the same time sent a maintenance plan notice to the maintenance department, successfully controlling the development of the fault and avoiding the risk of water supply interruption caused by sudden shutdown.

[0122] The above described the method for monitoring the working conditions of a pumping station based on digital twin in the embodiments of the present invention. Next, the system for monitoring the working conditions of a pumping station based on digital twin in the embodiments of the present invention will be described. Please refer to Figure 2 , an embodiment of the system for monitoring the working conditions of a pumping station based on digital twin in the embodiments of the present invention includes:

[0123] A receiving module 201, configured to receive the pressure, flow, vibration, temperature, and electrical energy parameters of the pumping station equipment through a PLC control cabinet to obtain a pumping station operation feature matrix;

[0124] A construction module 202, configured to construct a hybrid digital twin model of a pumping station physical model and a data-driven model based on the pumping station operation feature matrix, and generate operation status prediction data of the pumping station equipment through the hybrid digital twin model;

[0125] The calculation module 203 is configured to input the operation status prediction data and the actual monitoring data into a Transformer network for extracting the correlation pattern of working condition features and calculating the health index quantification, so as to obtain the equipment health assessment index;

[0126] The output module 204 is configured to trigger an early warning response mechanism by using the equipment health assessment index, and output the unit start-stop control strategy of the pumping station equipment to the dispatching system through a remote management platform.

[0127] Through the collaborative cooperation of the above-mentioned various components, by constructing a five-layer fusion architecture including a data layer, a feature layer, a matching layer, a scenario layer, and a decision layer, the efficient fusion and processing of multi-source heterogeneous sensing data are realized, effectively solving the data inconsistency problem in the traditional monitoring system, and improving the data basic quality of the pumping station working condition monitoring. By combining the physical model with the data-driven model, a hybrid digital twin model is constructed, which can not only accurately reflect the physical characteristics of the pumping station equipment, but also capture the complex dynamic characteristics of the equipment operation through deep learning methods, realizing the high-precision prediction of the pumping station working condition. The spatio-temporal feature analysis module based on the Transformer network architecture can effectively capture the complex correlation relationships and temporal evolution laws among the pumping station working condition parameters, and compared with the traditional methods, greatly improves the recognition ability of abnormal working conditions and the accuracy of early warning. The hierarchical working condition pattern matching and health index quantification calculation method of the present invention realizes the accurate assessment of the health state of the pumping station equipment, can identify potential equipment failures in the early stage, and provides a scientific basis for predictive maintenance. The intelligent early warning and maintenance decision support system of the present invention can automatically generate maintenance decision suggestions and unit start-stop control strategies according to the equipment health assessment index, realize the goal of "few operators on duty, remote monitoring, and low-power operation" of the pumping station, greatly reduce the operation and maintenance costs, and extend the service life of the equipment.

[0128] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions run on a computer, the computer is made to execute the steps of the method for monitoring the working conditions of a pumping station based on digital twin.

[0129] Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the above-described system, system, and unit can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0130] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a digital twin-based pump station condition monitoring device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0131] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A pump station operating condition monitoring method based on digital twin, characterized in that: include: The pressure, flow, vibration, temperature and electric energy parameters of the pump station equipment are received through the PLC control cabinet to obtain the pump station operation characteristic matrix; Building a hybrid digital twin model of a pump station physical model and a data-driven model based on the pump station operation characteristic matrix, and generating operation status prediction data of the pump station equipment through the hybrid digital twin model; The operation status prediction data and the actual monitoring data are input into the Transformer network to extract the operating condition feature association pattern and quantify the health index to obtain the equipment health assessment index; The equipment health assessment index is used to trigger an early warning response mechanism, and the unit start-stop control strategy of the pump station equipment is output to the dispatching system through a remote management platform.

2. The pump station operating condition monitoring method based on digital twin according to claim 1 is characterized in that: The pressure, flow, vibration, temperature and electric energy parameters of the pump station equipment are received through the PLC control cabinet to obtain the pump station operation characteristic matrix, including: Deploy a distributed sensor network at the pump station equipment, wherein the pump station equipment includes a pump station host, an electrical system, water inlet and outlet pipes, and auxiliary equipment; Based on the distributed sensor network, the sampling frequency of the pressure sensor, the flow sensor, the vibration sensor, the temperature sensor and the electric energy parameter sensor is set to obtain the sensor acquisition parameter configuration; Adaptively sampling is performed based on the sensor acquisition parameter configuration to obtain adaptive sampling data, and the adaptive sampling data is transmitted to the PLC control cabinet through the field bus to obtain original sensor data; Performing unit conversion, range normalization, time stamp calibration and outlier screening on the raw sensor data to obtain standardized processed data; Packing the standardized processed data into a unified format data packet including a sensor identifier, a timestamp, a measurement value, a quality identifier, and a device operation status tag to obtain a multi-source heterogeneous sensor data set; Heterogeneous data fusion and operating condition feature extraction are performed on the multi-source heterogeneous sensor data set to obtain a pump station operation feature matrix.

3. The pump station operating condition monitoring method based on digital twin according to claim 2 is characterized in that: The step of performing heterogeneous data fusion and operating condition feature extraction on the multi-source heterogeneous sensor data set to obtain a pump station operation feature matrix includes: A five-layer fusion architecture including data layer, feature layer, matching layer, scene layer and decision layer is constructed to obtain a data fusion processing framework; The data fusion processing framework is used to analyze and calculate the historical reliability index and current data quality indicator of each sensor to obtain a dynamic weight allocation result; According to the dynamic weight allocation result, weighted average processing is performed on redundant sensor data measuring the same physical quantity in the multi-source heterogeneous sensor data set, and least squares fitting is performed on sensor data measuring different physical quantities but physically associated in the multi-source heterogeneous sensor data set based on the pump station equipment characteristic curve to establish a parameter association model to obtain preliminary fusion data; Extracting time domain features and frequency domain features from the preliminary fusion data to obtain a full feature set, and performing feature selection on the full feature set based on the current operating conditions of the pump station equipment to obtain operating condition-related features; The principal component analysis is performed on the operating condition related characteristics to obtain the pump station operation characteristic matrix.

4. The pump station operating condition monitoring method based on digital twin according to claim 1 is characterized in that: The hybrid digital twin model of the pump station physical model and the data-driven model is constructed based on the pump station operation characteristic matrix, and the operation status prediction data of the pump station equipment is generated by the hybrid digital twin model, including: Based on the principles of fluid mechanics, a hydraulic model describing the flow-lift characteristics of the pump station, a mechanical model describing the response characteristics of mechanical components, and an electrical model describing the characteristics and control logic of electrical equipment are established to obtain the physical model framework of the pump station. Building a bidirectional long short-term memory network based on the pump station operation characteristic matrix to obtain a deep learning network architecture; The pump station physical model framework is combined with the deep learning network architecture, and model training is performed through the pump station operation feature matrix and digital twin technology to obtain a first digital twin model; Continuously compare the training state prediction data of the first digital twin model with the actual operation data, and start the model fine-tuning process through the transfer learning method, freeze the underlying feature extraction part and only update the parameters of the upper prediction part to obtain the second digital twin model; Perform uncertainty quantification calculation on the second digital twin model to generate a confidence interval of the training state prediction data, and generate a hybrid digital twin model with reliability evaluation according to the confidence interval; The pump station operation characteristic matrix is ​​input into the hybrid digital twin model, the predicted values ​​and confidence intervals of each key parameter in the future period are calculated, and the operation status prediction data of the pump station equipment is obtained.

5. The pump station operating condition monitoring method based on digital twin according to claim 1 is characterized in that: The operation status prediction data and the actual monitoring data are input into the Transformer network to extract the working condition feature association pattern and quantify the health index to obtain the equipment health assessment index, including: Aligning the time stamps of the operation status prediction data with the actual monitoring data and calculating the difference features to obtain the target input data; Perform attention weight calculation on the target input data, and simultaneously calculate 8 different self-attention heads in parallel and concatenate the results for linear transformation to obtain a multi-head attention representation; Applying a relative position encoding scheme to the multi-head attention representation to obtain a feature representation of enhanced temporal relationship, and constructing a parameter association graph based on the feature representation of enhanced temporal relationship to obtain parameter association features related to working conditions; Inputting the parameter association features related to the working condition into a Transformer deep network composed of 6 layers of encoders, each layer of encoders includes a multi-head self-attention sublayer and a feedforward neural network sublayer, applying residual connection and layer normalization to obtain a high-dimensional working condition representation; The high-dimensional operating condition representation is subjected to dimensionality reduction processing by a principal component analysis method to obtain a low-dimensional feature space, the current operating condition point is mapped in the low-dimensional feature space and connected to the historical operating condition points to form an evolution trajectory, and the low-dimensional feature space is divided into a normal operating area, a caution area, and an abnormal area, and the abnormality measurement value is calculated according to the Mahalanobis distance between the current operating condition point and the historical typical operating condition to obtain a pump station operating status characteristic map; Hierarchical working condition pattern matching and health index quantitative calculation are performed on the pump station operation status characteristic map to obtain the equipment health assessment index.

6. The method for monitoring pump station operating conditions based on digital twin according to claim 5 is characterized in that: The step of performing hierarchical working condition pattern matching and health index quantitative calculation on the pump station operation status characteristic map to obtain the equipment health assessment index includes: A three-layer operating condition model library is constructed, including a basic operating condition layer, a composite operating condition layer, and an abnormal operating condition layer. The basic operating condition layer includes four basic operating conditions: startup process, stable operation, variable operating condition operation, and shutdown process. The composite operating condition layer includes operating conditions under different loads and transition conditions under different start-stop strategies. The abnormal operating condition layer includes fault conditions such as bearing abnormality, impeller wear, motor overheating, and water hammer phenomenon. A benchmark template for operating condition identification is obtained. Input the pump station operation status characteristic map into the support vector machine classifier, perform the first-level matching through the working condition identification reference template, calculate the probability distribution of four basic working conditions, and obtain the basic working condition type identification result; Based on the basic working condition type identification result, the pump station operation status characteristic map is matched at the second level, and the working condition mode with the smallest distance is selected by calculating the dynamic time warping distance between the current working condition sequence and each standard working condition sequence in the working condition mode library to obtain a specific working condition matching result; The specific working condition matching results are compared and analyzed with the standard parameters, and the deviation scores of the five sub-indicators of efficiency index, vibration index, temperature rise index, electrical index and stability index are calculated to obtain the health sub-indicator values ​​of the pump station equipment. The health sub-indicator values ​​are weighted and optimized to obtain the equipment health assessment index.

7. The pump station operating condition monitoring method based on digital twin according to claim 1 is characterized in that: The triggering of the early warning response mechanism by using the equipment health assessment index and outputting the start-stop control strategy of the pump station equipment to the dispatching system through the remote management platform include: Constructing multi-level early warning trigger conditions, and monitoring the equipment health assessment index in real time to obtain early warning trigger results; Based on the warning trigger result, the intelligent diagnosis module is started, similar cases are retrieved from the case library through feature matching to calculate similarity scores, and the fault diagnosis result is obtained; The fault diagnosis result, the equipment health assessment index, the maintenance history and the operation plan are input into the decision framework of the double-depth Q network structure, and the action is selected through the online network and the action value is evaluated through the target network. The weighted combination of maintenance cost, downtime loss, equipment reliability and remaining life is used as the objective function to obtain the optimal maintenance strategy; Refining the optimal maintenance strategy to generate maintenance decision suggestions including maintenance timing, maintenance items, maintenance resource requirements and expected effect evaluation, and generating unit start-up and shutdown suggestions for the current working conditions based on the fault diagnosis results to obtain a joint control decision; The joint control decision is transmitted to the dispatching system through the standard data interface of the remote management platform, and the unit start-stop control strategy of the pump station equipment is output.

8. A pump station operating condition monitoring system based on digital twins, characterized in that: Used to implement the pump station operating condition monitoring method based on digital twins as described in any one of claims 1 to 7, the pump station operating condition monitoring system based on digital twins includes: The receiving module is used to receive the pressure, flow, vibration, temperature and electric energy parameters of the pump station equipment through the PLC control cabinet to obtain the pump station operation characteristic matrix; A construction module is used to construct a hybrid digital twin model of a pump station physical model and a data-driven model based on the pump station operation characteristic matrix, and generate operation status prediction data of the pump station equipment through the hybrid digital twin model; A calculation module is used to input the operation status prediction data and the actual monitoring data into the Transformer network to extract the working condition feature association pattern and quantitatively calculate the health index to obtain the equipment health assessment index; The output module is used to trigger the early warning response mechanism using the equipment health assessment index, and output the unit start-stop control strategy of the pump station equipment to the dispatching system through the remote management platform.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor executes the pump station condition monitoring method based on digital twin as described in any one of claims 1 to 7.

Images