Digital twin pump station intelligent perception construction method

Abstract

The application discloses a digital twin pump station intelligent sensing construction method, and relates to the technical fields of pump station digital twin and intelligent sensing. The method acquires multi-source heterogeneous sensing signals of five types of sensors arranged on a motor, a water pump and a pipeline of a pump station and performs space-time alignment, constructs a mechanism model based on an electric-motor-liquid energy quality conservation relation and calculates a cross-physical quantity residual vector, constructs a working condition parameterized conservation constraint manifold according to working condition parameters, decomposes the residual vector into a conservation consistent component in a tangent space and a conservation violation component in a normal direction, makes the conservation consistent component backfill a hybrid twin model for self-calibration, uses the conservation violation component as a device degradation fingerprint to drive fault sensing and updates a bidirectional synchronization through a virtual-real bidirectional driving engine, and uses whether the residual violates a conservation coupling relation as a criterion to simultaneously improve the mapping fidelity and early fault distinguishability of a physical entity and a digital twin without expanding the sensing arrangement range.

Description

Technical Field

[0001] This invention relates to the field of digital twin and intelligent sensing technology for pumping stations, specifically to a method for constructing intelligent sensing for digital twin pumping stations. Background Technology

[0002] Pumping stations are core hubs in urban water supply, drainage, flood control, and agricultural irrigation projects, responsible for water transport and regulation. Their operational safety and economic efficiency directly impact water supply security and flood control safety. Traditional pumping station management relies primarily on manual inspections and post-event maintenance, which suffers from insufficient sensing breadth, delayed fault response, and limited model mapping accuracy. Digital twin technology, by constructing high-fidelity virtual models of physical entities, enables real-time mapping and interaction between the physical and information worlds, providing a new technological path for intelligent pumping station management. Intelligent sensing components, acting as both data sources and mapping bridges, are crucial in determining the fidelity of the digital twin.

[0003] Chinese patent CN114992106B discloses a pump station unit operation status analysis system based on digital twins. The system adopts an architecture that combines a physical entity subsystem of the pump station unit, a digital twin engine subsystem, a virtual entity subsystem of the pump station unit, and a digital twin service platform. It constructs a virtual digital image of the pump station unit through geometric models, mechanism models, process models, and data models. With the help of a small number of sensors, it solves for performance parameters that are difficult to measure directly, such as flow rate, net positive suction head (NPSH), and power, through model simulation and numerical calculation. Finally, it realizes bidirectional interaction and model iteration between the physical entity and the virtual entity through the digital twin engine. However, in the process of data stimulation and model iteration, this scheme always regards the deviation between the predicted value of the mechanism model and the measured value of the sensor as an error that needs to be eliminated through model iteration. Its goal is to continuously reduce this deviation so that the virtual entity approximates the physical entity. In a full-element system such as a pump station, which is rigidly coupled by the energy and mass conservation relationship, this approach will smooth out the deviation component carrying the early degradation information of the equipment in the process of eliminating deviations across physical quantities and averaging multi-source data fusion. Furthermore, it cannot distinguish between the deviation caused by the switching of operating conditions and the deviation caused by equipment degradation. Ultimately, although the breadth of perception is improved, the mapping fidelity between the physical entity and the digital twin and the early identifiability of equipment failures are limited.

[0004] Chinese patent CN112417619B discloses a pump unit optimization and adjustment system and method based on digital twins. This method combines physical entities, digital twins, and digital space, using data mirroring, mapping reconstruction, information interaction feedback, and simulation, and leveraging optimization algorithms and machine learning to achieve pump unit operation status monitoring, optimization adjustment, and health prediction. However, this method also employs a processing framework of independent modeling and weighted fusion of multi-source sensing data. Its weighted fusion objective remains to reduce the overall deviation amplitude between prediction and measurement, failing to utilize the conservation coupling relationship that the deviation should satisfy between different physical quantities. Therefore, it is also difficult to distinguish between the operating condition drift deviation allowed by the conservation relationship and the local failure deviation that violates the conservation relationship under varying operating conditions.

[0005] It is evident that existing digital twin solutions for pumping stations generally treat the deviation between mechanism prediction and actual measurement as an error to be eliminated. They process multi-source heterogeneous data by independently modeling and then fusing the data for each physical quantity, aiming to reduce the overall deviation amplitude. This dilutes the deviation component carrying equipment degradation information during cross-modal averaging of multi-source fusion, and misjudges operational drift as equipment degradation. Consequently, without expanding the sensor deployment, it is difficult to simultaneously improve the mapping fidelity between the physical entity and the digital twin while maintaining early fault identification. Distinguishing between operational or calibration drift deviations allowed by the energy-mass conservation principle and local failure deviations that violate conservation principles is the core technical challenge restricting the mapping fidelity and early fault identification of pumping station digital twins. Summary of the Invention

[0006] To address the core technical bottleneck in existing digital twin intelligent sensing technologies for pump stations, which treats the discrepancy between mechanism prediction and actual measurement as an error to be eliminated, making it difficult to distinguish between operating condition drift deviation and equipment degradation deviation in multi-source fusion, thus limiting mapping fidelity and early fault identification, this invention provides a method for constructing intelligent sensing for digital twin pump stations. This method decomposes the cross-physical quantity residual vector between multi-source heterogeneous sensing signals and mechanism predictions along the parameterized conservation constraint manifold of the pump station's electromechanical-hydraulic energy-mass conservation relationship into a consistent conservation component and a violation conservation component. The consistent conservation component is then reinjected into the twin for self-calibration, and the violation conservation component drives fault perception and virtual-real synchronization. Without expanding the sensor deployment, this method achieves high-fidelity mapping of all elements of the pump station and early fault identification at the principle level of energy and mass conservation coupling constraints.

[0007] The technical solution of this invention is as follows:

[0008] The intelligent sensing construction method for digital twin pumping stations includes the following steps:

[0009] Step S1: Obtain vibration sensors, temperature sensors, pressure sensors, flow sensors, and current sensors respectively deployed on the motor, water pump, valve, pipeline, and water tank of the pumping station; collect vibration signals, temperature signals, pressure signals, flow signals, and current signals from the motor, water pump, valve, pipeline, and water tank; construct multi-source heterogeneous sensing signals from the vibration signals, temperature signals, pressure signals, flow signals, and current signals; and perform spatiotemporal alignment on the multi-source heterogeneous sensing signals to obtain spatiotemporally aligned sensing data.

[0010] Step S2: Construct a mechanism model based on the energy-mass conservation relationship of the pump station's electromechanical-hydraulic system; generate mechanism prediction values ​​for each physical quantity from the spatiotemporal aligned sensing data through the mechanism model; and calculate the cross-physical quantity residual vector between the mechanism prediction values ​​and the spatiotemporal aligned sensing data.

[0011] Step S3: Construct a parameterized conservation constraint manifold based on the current operating condition parameters, project the cross-physical quantity residual vector onto the parameterized conservation constraint manifold, and decompose it to obtain the conservation consistency component located in the tangent space of the parameterized conservation constraint manifold and the conservation violation component located in the normal of the parameterized conservation constraint manifold.

[0012] Step S4: The conserved consistency component is fed back into the hybrid twin model formed by the fusion of the mechanism model and the data-driven model, and the hybrid twin model is subjected to twin self-calibration.

[0013] Step S5: The conservation violation component is used as a device degradation fingerprint to drive fault perception, and the bidirectional synchronization between the motor, water pump, valve, pipeline and water tank and the hybrid twin model is updated through the virtual and real bidirectional driving engine.

[0014] The beneficial effects of this invention are as follows:

[0015] First, by decomposing the cross-physical quantity residual vector between multi-source heterogeneous sensing signals and mechanism predictions along the condition-parameterized conservation constraint manifold into conservation consistency components and conservation violation components, a mechanism-level distinction is achieved between condition drift deviations and equipment degradation deviations. The mechanism is as follows: pump station motors, pumps, valves, pipelines, and water tanks are rigidly coupled by the same electro-mechanical-hydraulic energy-mass conservation chain. Any mechanism prediction deviation of a physical quantity will inevitably propagate to adjacent physical quantities according to the conservation relationship and a definite coupling ratio. Therefore, deviations generated by healthy equipment under changes in operating conditions or calibration can only fall within the tangent space of the conservation constraint manifold. Only local failures that disrupt a certain conservation link will generate deviation components pointing towards the manifold normal. Accordingly, the conservation violation component accurately isolates the deviation carrying degradation information, while the conservation consistency component corresponds to condition drift that does not require alarms. Compared to treating the deviation as a whole as an error to be eliminated and smoothing it out during model iteration, this scheme no longer aims to reduce the magnitude of the deviation, but instead uses whether the deviation violates the conservation coupling ratio as the criterion. Thus, it preserves the degraded signal diluted by traditional schemes without expanding the breadth of sensor deployment.

[0016] Secondly, by identifying the parameters of the conservation constraint manifold online based on the inverter speed, valve opening, and water level in the pool, and aligning the phase of the residuals during operating condition switching, stable extraction of conservation violation components under varying operating conditions is achieved. The mechanism is as follows: the shape of the conservation constraint manifold in the pumping station continuously drifts with the operating conditions, and the transient state during operating condition switching causes time misalignment in the electromechanical-hydraulic links. If a fixed manifold is used as the criterion, the legitimate deviations of the transient state during operating condition switching will be misprojected onto the manifold normal and misjudged as conservation violations. This scheme identifies the manifold parameters online using operating condition parameters as independent variables and performs phase alignment of the residuals in each link, ensuring that the extraction of conservation violation components is always relative to the manifold matching the current operating condition. Compared to the approach of using overall deviation amplitude weighted fusion, which cannot adaptively distinguish the causes of deviations according to operating conditions, this scheme significantly reduces false alarms under varying operating conditions.

[0017] Third, by constructing a conservation violation residual space field on a physical topology diagram consisting of the connections between motors, pumps, valves, pipelines, and water tanks, and then inverting the root cause location based on its propagation direction along the topology, the root cause location of faults under sparse sensing conditions is achieved. The mechanism is as follows: local failures act as sources of conservation violations, and their effects propagate node by node along the conservation coupling relationships in the physical topology. Applying the diffusion inverse operation of the topology diagram to the conservation violation residual space field restores the source distribution, thereby converging the dispersed multi-point deviations towards a single root cause location.

[0018] Fourth, the aforementioned self-calibration of the conserved consistency component and the perception driven by the conserved violation component are generated by the same residual decomposition, and further driven by the field strength distribution of the conserved violation residual spatial field, resulting in non-uniform virtual-real bidirectional synchronization. This allows the two objectives of improving mapping fidelity and improving early fault identifiability, which are mutually constrained in traditional schemes, to produce a synergistic effect that exceeds the sum of their individual effects. Specifically, the same conservation decomposition outputs both a conservation consistency component for twin self-calibration and improved fidelity, and a conservation violation component for fault detection and improved identifiability. Both components share the conservation constraint manifold as the computational basis without repeated modeling. Simultaneously, the field strength distribution of the conservation violation residual space field is used to allocate synchronization frequencies to each node in the physical topology graph. The synchronization frequency is increased for nodes with high field strength and decreased for nodes with low field strength. This concentrates synchronization resources on the nodes that carry the most degradation information within the limited edge cloud bandwidth, ensuring both the real-time response of key nodes and the global mapping fidelity. This breaks the constraint between cost and fidelity in traditional uniform high-frequency synchronization, achieving a synergistic gain greater than the sum of its parts. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the intelligent sensing construction method for digital twin pump stations according to the present invention.

[0020] Figure 2 This is a flowchart of the multi-source heterogeneous sensing signal acquisition and spatiotemporal alignment in step S1 of the present invention.

[0021] Figure 3 This is a flowchart of the mechanism model construction and cross-physical quantity residual vector calculation in step S2 of the present invention.

[0022] Figure 4 This is a flowchart of the construction of the working condition parameterized conservation constraint manifold and residual decomposition in step S3 of the present invention.

[0023] Figure 5 This is a flowchart of the twin self-calibration and conservation violation component-driven fault perception and virtual-real synchronization in steps S4 and S5 of the present invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0025] like Figure 1As shown, this embodiment provides a digital twin pump station intelligent perception construction method, applied to a city water supply pump station. The pump station includes three parallel centrifugal pump units and their drive motors, outlet valves, water supply pipelines, and a forebay. Steps S1 to S5 are executed sequentially to form a closed-loop mapping from the physical entity through a data link to the virtual model and back to the physical entity. Each step is explained in detail below.

[0026] Step S1: Acquisition and spatiotemporal alignment of multi-source heterogeneous sensing signals. For example... Figure 2 As shown, step S1 first establishes a monitoring index system to meet the sensing needs of all elements of the pumping station, and based on this, completes the optimized deployment of sensors and the unified access of multi-protocol heterogeneous devices. Specifically, vibration sensors, temperature sensors, pressure sensors, flow sensors, and current sensors are deployed at key locations of the five core physical entities of the pumping station: motors, pumps, valves, pipelines, and water tanks. Vibration sensors are deployed at the bearing housings of the motor and pumps, with a sampling frequency of 10000Hz, to collect vibration signals reflecting the mechanical and hydraulic coupling state; temperature sensors are deployed at the motor windings and bearings, with a sampling frequency of 1Hz, to collect temperature signals; pressure sensors are deployed at the pump inlet and outlet and key nodes of the pipeline, with a sampling frequency of 100Hz, to collect pressure signals; flow sensors are deployed at the main water supply pipeline, with a sampling frequency of 10Hz, to collect flow signals; and current sensors are deployed at the motor power supply circuit, with a sampling frequency of 1000Hz, to collect current signals. The vibration, temperature, pressure, flow, and current signals together constitute a multi-source heterogeneous sensing signal.

[0027] To address the issue of sensors on motors, pumps, valves, pipelines, and water tanks outputting data using different communication protocols, a protocol adaptation unit is set up at the edge IoT layer. This unit converts data output based on different protocols such as Modbus, OPC UA, and MQTT into structured records with device identifiers, physical quantity types, timestamps, and spatial coordinates. This enables unified access for heterogeneous devices using multiple protocols, providing a consistent data foundation for subsequent processing.

[0028] Because the sampling frequencies, transmission delays, and spatial locations of the five types of sensors differ, the multi-source heterogeneous sensing signals exhibit temporal and spatial inconsistencies when reaching the edge IoT layer. Therefore, spatiotemporal alignment of these signals is necessary. In the time dimension, using the time grid corresponding to the highest sampling frequency as a reference, physical quantities with lower sampling frequencies are upsampled using linear interpolation. Physical quantities with fixed transmission delays are timestamped according to their calibrated delays, ensuring alignment of all physical quantities on the same time grid. In the spatial dimension, using the pump station's three-dimensional coordinate system as a unified reference, the spatial coordinates of each sensor are mapped to this coordinate system, and the physical quantities are grouped according to their physical entity affiliation. After the aforementioned timestamping correction, interpolation upsampling, and coordinate mapping, the multi-source heterogeneous sensing signals are organized into spatiotemporally aligned sensing data where each physical quantity corresponds one-to-one on a unified spatiotemporal grid. The timestamping correction and coordinate mapping described above are conventional data preprocessing methods in this field; in this step, the corresponding parameters can be determined based on the actual sampling frequency and spatial layout of the pump station.

[0029] In one possible implementation, before spatiotemporally aligning the multi-source heterogeneous sensing signals, the edge computing node sequentially performs data cleaning, anomaly detection, and missing value imputation on the signals: data cleaning removes out-of-range values ​​and communication garbled characters; anomaly detection marks instantaneous jumps in a single physical quantity; and missing value imputation uses linear interpolation of nearby time points to fill in short-term packet loss. When the anomaly detection result meets preset local response conditions, such as when the effective value of the vibration signal continuously exceeds a preset safety threshold within 50ms, the edge computing node directly issues a local closed-loop response of speed reduction or shutdown to the corresponding unit without waiting for cloud processing, thereby avoiding the response lag caused by traditional pure cloud processing.

[0030] Step S2: Mechanism model construction and calculation of residual vectors across physical quantities. For example... Figure 3 As shown, step S2 constructs a mechanism model based on the energy-mass conservation relationship of the pump station's electromechanical-hydraulic system, and generates mechanism prediction values ​​for each physical quantity from the spatiotemporal aligned sensing data through the mechanism model, and then calculates the cross-physical quantity residual vector between the mechanism prediction values ​​and the spatiotemporal aligned sensing data.

[0031] The electromechanical-hydraulic energy-mass conservation relationship includes the electromechanical energy conversion stage, the hydraulic work stage, the pipeline head loss stage, and the pool mass balance stage. The electromechanical energy conversion stage describes the process by which the power supplied by the motor is converted into the pump shaft power after electromagnetic conversion and mechanical transmission; the hydraulic work stage describes the process by which the pump shaft power is converted into the hydraulic power represented by the product of the water head and flow rate; the pipeline head loss stage describes the head loss of the water body in the water conveyance pipeline due to frictional resistance and local resistance; and the pool mass balance stage describes the mass conservation relationship between the inflow and outflow rates and water level changes in the forebay. Under steady-state conditions, the above stages satisfy the conservation of energy and mass, and their basic relationship can be summarized as follows:

[0032] ,

[0033] in: The power supplied to the motor is a scalar quantity, ranging from 0 to the rated power, and is measured in W. It is determined by the current signal and the supply voltage and represents the electrical power input to the pumping station system. The combined electromechanical and hydraulic efficiency is a scalar quantity with a value range of (0,1). It is dimensionless and is determined by the efficiency characteristic curve of the pump unit, representing the efficiency of energy conversion. Let be the density of water, and be a scalar with a value of 1000, expressed in kg / m³, representing the density of the transported water. is the gravitational acceleration, is a scalar with a value of 9.8, and the unit is m / s², representing the intensity of the gravitational field; is the flow rate, is a scalar quantity, with a value range from 0 to the rated flow rate, and the unit is m³ / s. It is determined by the flow rate signal and represents the volume of water transported per unit time. The head is a scalar quantity with a value range from 0 to the rated head, and the unit is m. It is determined by conversion from the pressure signal and represents the head of the water pump lifting the water body by doing work on it. The equivalent power of the pipeline head loss is a scalar quantity, ranging from 0 to the rated power, with units of W. It is determined by the pipeline characteristics and characterizes the power consumed along the pipeline and by local resistance. The above conservation relationship is a well-known basic relationship between hydraulics and electromechanical systems in pumping stations. In this step, it serves as the basis for constructing the mechanism model. Its specific form can be determined by parameter calibration using the measured characteristic curves of the pump unit.

[0034] Based on the aforementioned conservation relationships and combined with the pump station's geometry, the mechanistic model provides mechanistic prediction values ​​for each physical quantity in the spatiotemporally aligned sensing data. The measured values ​​of five physical quantities—vibration, temperature, pressure, flow rate, and current—after spatiotemporal alignment at a certain moment constitute a measured vector. The mechanistic model, based on conservation relationships, calculates the expected values ​​of each physical quantity from independently measurable physical quantities, thus forming a mechanistic prediction vector. The differences between the measured values ​​and the mechanistic prediction values ​​of each physical quantity are normalized according to the reference dimensions of each physical quantity and arranged to obtain a cross-physical quantity residual vector. ,in: The residual vector spanning physical quantities is a five-dimensional column vector, with each component taking values ​​ranging from... Common values ​​within the interval, dimensionless, calculated by this formula, characterize the normalized deviation between the predicted and measured mechanisms of the five types of physical quantities, and are the objects of subsequent conservation decomposition; superscript Indicates vector transpose; subscript Number the physical quantities. The values ​​1 through 5 correspond to vibration, temperature, pressure, flow rate, and current, respectively. For the first The normalized residual of a physical quantity is a scalar, dimensionless, and is calculated by this formula. It characterizes the magnitude and direction of the deviation of the physical quantity. For the first The actual measurement of a physical quantity is a scalar, and the unit is the measurement unit of that physical quantity. It is determined by the spatiotemporal alignment sensing data and represents the value actually measured by the sensor. For the first The predicted mechanism value of a physical quantity is a scalar, with units of and . Similarly, the value of this physical quantity, calculated from the aforementioned mechanism model, represents the proper value of the physical quantity under the conservation relationship; For the first The reference quantity for each physical quantity is a scalar quantity, and its unit is . The same value is taken as the rated value or range of the physical quantity, which is determined by the design parameters of the pump station. It is used to normalize the deviations of different dimensions to the same dimensionless scale, so that the residuals of the five types of physical quantities can be compared in the same vector.

[0035] The residual vectors across physical quantities are interconnected according to the coupling coefficients of the electromechanical energy conversion link, the hydraulic work link, the pipeline head loss link, and the water tank mass balance link. That is, the residuals of the five types of physical quantities are not independent of each other, but are interdependent by the constraint of conservation relations. This correlation is the physical basis for the conservation decomposition in the subsequent step S3.

[0036] Step S3: Construction of the conserved constraint manifold and residual decomposition for the operating conditions. For example... Figure 4As shown, step S3 constructs a parameterized conservation constraint manifold based on the current operating parameters, and projects the cross-physical quantity residual vector onto the parameterized conservation constraint manifold to decompose it into conservation consistency components and conservation violation components. This step is the core difference between this invention and the prior art: the prior art aims to reduce the overall magnitude of the residual, while this step uses whether the residual violates the conservation coupling relationship as the criterion to decompose the residual in direction.

[0037] The operating parameters include inverter speed, valve opening, and water tank level. Since the conservation relationships exhibit different coupling forms under different operating conditions, the parameterized conservation constraint manifold changes continuously with the operating parameters. At each moment, the electromechanical-hydraulic energy-mass conservation relationship is linearized at the current operating parameters to obtain a constraint Jacobian matrix. Each row of this matrix represents the coupling ratio constraint imposed by each conservation link on the residuals of the five physical quantities. Residuals satisfying all conservation constraints fall within the tangent space spanned by the null space of this constraint Jacobian matrix, and are thus considered conservation-consistent components; residuals violating conservation constraints fall within the normal direction spanned by the row space of this constraint Jacobian matrix, and are thus considered conservation-violation components. Based on this, the cross-physical quantity residual vector is orthogonally decomposed along the parameterized conservation constraint manifold:

[0038] ,in: The component that violates conservation is a five-dimensional column vector, and each component is dimensionless. It is calculated by this formula and is located in the normal direction of the parameterized conservation constraint manifold of the working condition. It represents the deviation caused by the local failure of the conservation link and is used as the equipment degradation fingerprint. The conserved uniform components are five-dimensional column vectors, each component is dimensionless, and are calculated by this formula. They are located in the tangent space of the parameterized conservation constraint manifold of the working condition, representing the working condition or calibration drift deviation allowed by the conservation relationship, and are used for twin self-calibration. To constrain the Jacobian matrix, for A dimensionless matrix obtained by linearizing the electromechanical-hydraulic energy-mass conservation relationship at the current operating parameters, with a number of rows. Equal to the number of conservation elements involved in the constraint, in this embodiment The value of 4 corresponds to the four stages of electromechanical energy conversion, hydraulic work, pipeline head loss, and water tank mass balance, respectively, representing the coupling ratio constraint of each conservation stage on the residuals of the five types of physical quantities; superscript Indicates matrix transpose; superscript This represents finding the inverse of a matrix. The definition is the same as the aforementioned formula for the residual vector across physical quantities. To quantitatively characterize the degree of conservation violation, the degree of conservation violation is further calculated:

[0039] ,in: Let be the degree of conservation violation, and let be a scalar with a range of values ​​of . , dimensionless, is calculated by this formula, and represents the proportion of the conservation violation component in the total residual. The closer its value is to 1, the more the deviation is biased towards local failure that violates the conservation relationship. The closer its value is to 0, the more the deviation is biased towards the conservation-allowed operating condition drift. The L2 norm of a vector; and The definition is the same as the aforementioned formula; To prevent small positive numbers with zero denominators, a scalar value of 0.0001 is used, which is dimensionless and set empirically to avoid calculation overflow when the residuals are close to zero. Based on the conservation violation degree, when the entire cross-physical quantity residual vector falls into the tangent space of the parameterized conservation constraint manifold of the operating condition, i.e., the conservation violation degree is lower than the preset criterion, it is determined as operating condition drift or calibration drift; when the cross-physical quantity residual vector has a non-zero component in the normal of the parameterized conservation constraint manifold of the operating condition, i.e., the conservation violation degree is higher than the preset criterion, it is determined as a local failure.

[0040] Considering that the shape of the conservation constraint manifold in the pumping station continuously drifts with operating conditions, and that transient changes in operating conditions can cause time misalignments in the electromechanical-hydraulic processes, using a fixed constraint Jacobian matrix would incorrectly project legitimate deviations from these transient changes onto the manifold normal, leading to misjudgments of conservation violations and false alarms. Therefore, this step uses the inverter speed, valve opening, and water tank level as parameters to perform online parameter identification of the conservation constraint manifold, enabling the constraint Jacobian matrix to adaptively update with changing operating conditions. The online parameter identification employs a recursive update with a forgetting factor. ,in: for The conserved constraint manifold parameter vector at time t is a column vector, with dimensionless components. It is recursively calculated using this formula and represents the values ​​of the coupling coefficients of the constraint Jacobian matrix under the current working condition. It serves as the basis for constructing the working condition-parameterized conserved constraint manifold; subscripts... The time points are numbered, taking consecutive discrete time points; The vector of conserved constraint manifold parameters from the previous moment has the same meaning. This is obtained from the previous recursion; Let be the covariance matrix of the previous time step. is a square matrix, dimensionless, maintained by the recursive process, and characterizes the uncertainty of the parameter estimation. for The regression vector at any given time is a column vector, dimensionless, constructed from the inverter speed, valve opening, and water tank level expanded using basis functions, representing the effect of the current operating condition on the manifold parameters; superscript Indicates vector transpose; The forgetting factor is a scalar with a value range of (0,1]. In this embodiment, it is set to 0.98. It is dimensionless and is set empirically. The smaller the value, the faster the historical data is forgotten and the more sensitive the tracking of changes in working conditions. If the value is too small, the estimation variance will increase. If the value is too large, the tracking will lag. Therefore, a compromise of 0.98 is adopted. for The conservation constraint target value, calculated from the health state data at any given time, is a scalar, dimensionless quantity. It is obtained by substituting the current spatiotemporally aligned sensing data into the conservation relation and represents the conservation constraint that must be satisfied at the current time. Simultaneously, the time misalignment of the cross-physical quantity residual vectors between the electrical-mechanical-hydraulic links during the operating condition switching period is phase-aligned, ensuring that the residuals of each link correspond to the same physical process during conservation decomposition. Phase alignment is achieved by finding the time offset that maximizes the cross-correlation between the residuals of each physical quantity and the drive control signal. ,in: For the first The optimal phase alignment delay of each physical quantity is a scalar, and its value range is [value range missing]. The unit is seconds (s), calculated by this formula, representing the time misalignment of the physical quantity residual relative to the drive control signal. It is used to eliminate the time misalignment during transients of operating condition switching before conservation decomposition. (Subscript) Number the physical quantities, from 1 to 5; superscript This indicates the value corresponding to the extreme value; Let be the candidate delay, be a scalar, and be within the interval . Internal search; The maximum search delay is denoted by , which is a scalar with a value of 2 and a unit of seconds. It is determined by the longest transition time when the pump station's operating conditions are switched, and it represents the search range for phase alignment. for The normalized drive control signal is a scalar, dimensionless quantity, determined by the rate of change of the inverter speed command, and characterizes the drive for switching operating conditions. For the first A physical quantity in The normalized residual at time has the same meaning as described above. Summation symbol This represents all moments within the phase alignment window. Summation. Furthermore, the drift of the identified parameters of the parameterized conserved constraint manifold under the aforementioned operating conditions over time is output as a calibration aging index. The calibration aging index reflects the slow shift of the conserved constraint manifold parameters relative to the initial calibration value after excluding the influence of operating conditions; it can be used independently to assess the calibration aging status of the sensors and the unit. ,in: The calibration aging index is a scalar with a range of non-negative real numbers and no dimension. It is calculated by this formula and represents the cumulative drift of the conserved constraint manifold parameter relative to the initial calibration value. The larger the value, the more significant the calibration aging. for The conservation constraint manifold parameter vector at time t is defined as before; This is the conserved constraint manifold parameter vector at the initial calibration time, with the same meaning as... This will be determined when the system is first commissioned and calibrated; L2 norm of a vector; subscript This is the initial calibration time.

[0041] Step S4: Conservative and consistent component reinjection twin self-calibration. For example... Figure 5 As shown, in step S4, the conserved consistency component is fed back into the hybrid twin model formed by the fusion of the mechanism model and the data-driven model, and the hybrid twin model is subjected to twin self-calibration.

[0042] The hybrid twin model is formed by fusing the mechanistic model and the data-driven model. The mechanistic model includes a geometric model constructed based on Building Information Modeling (BIM) and 3D laser scanning, a system-level reduced-order model, and a simplified computational fluid dynamics (CFD) model: the geometric model generates a high-precision 3D geometric mirror of the pump station by registering BIM and 3D laser scanning data; the system-level reduced-order model extracts the dominant flow modes by performing intrinsic orthogonal decomposition on the full-order CFD simulation results, forming a low-dimensional behavioral model that can be solved online in real time; the simplified CFD model provides hydraulic mechanism modeling capabilities for the internal flow field of the pump. The data-driven model includes a neural network model and a support vector machine (SVM) model, which learns the equipment's operating rules from the spatiotemporally aligned sensing data. The geometric modeling, reduced-order modeling, simplified CFD modeling, and training of the neural network and SVM are all known techniques in the field and can be implemented according to the actual structure and historical data of the pump station.

[0043] Since the conserved consistency component represents the allowable operating condition or calibration drift deviation of the conservation relationship, i.e., the reasonable offset of the physical entity relative to the current hybrid twin model, the conserved consistency component is fed back into the hybrid twin model for self-calibration of its parameters. This ensures that the hybrid twin model continuously approximates the physical entity as its operating condition and calibration state change, without contaminating the twin due to inadvertently incorporating degradation deviations into the calibration. When fusing the mechanistic model and the data-driven model, dynamic weights are assigned to the mechanistic model and the data-driven model based on the magnitude of the conserved consistency component: when the conserved consistency component is small, indicating a good fit between the mechanistic model and the measured values, the weight of the mechanistic model is increased; when the conserved consistency component is large, indicating a significant deviation of the operating condition from the mechanistic model's calibration point, the weight of the data-driven model is increased, thereby maintaining the mapping accuracy of the hybrid twin model under different operating conditions. The dynamic weight allocation is determined by linear interpolation of the values ​​of the conserved and consistent components between the mechanistic model and the data-driven model. This is a conventional weighted fusion method in the field, and its specific weighting coefficients will not be elaborated here.

[0044] After step S4, the hybrid twin model completes twin self-calibration based on the conserved consistent components. Its mapping fidelity is maintained continuously during the change of operating conditions. At the same time, since the degradation information has been decomposed and stripped to the conservation violation components, the twin self-calibration process will not absorb the degradation deviation into the model, thereby ensuring the purity of the conservation violation components on which subsequent fault perception is based.

[0045] Step S5: Conservation violation components drive fault perception and virtual-real bidirectional synchronization. For example... Figure 5 As shown, step S5 uses the conservation violation component as a device degradation fingerprint to drive fault perception, and updates the bidirectional synchronization between the motor, water pump, valve, pipeline and water tank and the hybrid twin model through the virtual and real bidirectional driving engine.

[0046] First, a physical topology graph is constructed based on the connection relationships of the motor, pump, valve, pipeline, and water tank. Each node of the physical topology graph corresponds to a physical entity and its sensing location, and each edge corresponds to the hydraulic and mechanical connection relationships between physical entities. The conservation violation components are then used to construct a conservation violation residual space field at each node of the physical topology graph. Since the actual deployed sensors are sparse at the topology nodes, the conservation violation residual space field at all nodes of the physical topology graph needs to be reconstructed from sparse observations. The reconstruction employs graph signal smoothing regularization. ,in: For the reconstructed conservation violation of the residual space field, for The dimensionless column vector, whose components are all dimensionless, is calculated by this formula and represents the degree of conservation violation on all nodes of the physical topology graph. It serves as the basis for subsequent root cause inversion and synchronization frequency allocation. The total number of nodes in the physical topology graph is a positive integer, determined by the number of physical entities of the pump station and the number of sensing locations; Here is the sensor sampling matrix, and here is... The matrix has elements of 0 or 1 and is dimensionless. Each row has 1 at the node corresponding to the deployed sensor and 0 at the rest. It is determined by the sensor deployment location and represents the sampling relationship of sparse observations over the entire field. The number of nodes with deployed sensors is a positive integer, and Less than Characterizing the sparsity of observations; superscript Indicates matrix transpose; superscript This represents finding the inverse of a matrix. The smoothing regularization coefficient is a scalar with a range of positive real numbers. In this embodiment, it is set to 0.5. It is dimensionless and is set empirically. The larger the value, the smoother the reconstructed field. If the value is too large, it will blur the root cause location; if the value is too small, it will amplify the observation noise. Therefore, a compromise of 0.5 is adopted. Let be the Laplace matrix of the physical topology graph, and be... The matrix, dimensionless, is calculated from the adjacency relationship of the physical topology graph and represents the diffusion relationship of conservation violation along the topology; For the conservation violation observations on the deployed sensor nodes, The column vector, with dimensionless components, is obtained from the conservation violation components at their corresponding nodes, representing the degree of conservation violation observed sparsely. Furthermore, the root cause location corresponding to the device degradation fingerprint is obtained by inverting the conservation violation residual space field along the propagation direction of the physical topology. Since local failures act as the source of conservation violations, their influence spreads node by node along the conservation coupling relationship in the physical topology. Applying the diffusion inverse operation of the topology to the conservation violation residual space field can restore the source distribution; the node where the source distribution reaches its maximum is the root cause location. ,in: For the conservation violation source distribution obtained by inversion, The dimensional column vector, with each component being dimensionless, is calculated using this formula and characterizes the strength of each node as a source of conservation violation. Let Laplace matrix be the physical topology graph, defined as before; For the reconstructed conservation violation residual space field, the definition is the same as before; The node corresponding to the root cause is determined by this formula, and represents the physical entity node with the greatest conservation violation source strength. For the node variables of the physical topology graph, traverse the node set. ; The set of nodes in the physical topology graph; For nodes The conservation at a certain point violates the source strength, which is a scalar, dimensionless, and is... The One component; Indicates taking the absolute value; This indicates the node that maximizes the subsequent expression. Considering the undersampling of the conservation violation residual space field in the region between sensors under sparse sensing conditions, this embodiment further selects a soft-measurement supplementary point location at the location with the maximum field strength gradient of the conservation violation residual space field to supplement the sensing of the undersampling region. The field strength gradient of the conservation violation residual space field is measured on the physical topology graph by the difference in field value between each node and its neighboring nodes, and the candidate node with the largest field strength gradient is selected as the soft-measurement supplementary point location: ,in: For nodes The field strength gradient at a given point is a scalar, taking values ​​of non-negative real numbers, dimensionless, and calculated using this formula. It characterizes the degree of drastic change in the residual space field relative to adjacent nodes due to the violation of conservation at that node; subscript... Number the nodes; This indicates that for all nodes... Connected edges Summation; The set of edges of the physical topology graph; For nodes Adjacent nodes connected to each other; For nodes With nodes The edge weights between the nodes are scalars, taking values ​​of positive real numbers, dimensionless, and determined by the tightness of the hydraulic and mechanical connection between the two nodes, characterizing the strength of the conserved coupling. and They are nodes With nodes The conservation at a certain point violates the residual space field value, which is a scalar, dimensionless, and is... The corresponding components; The node corresponding to the soft measurement supplementary point location is determined by this formula; The set of candidate replacement nodes is the set of nodes without deployed physical sensors; superscript This indicates the node corresponding to the extreme value. At the soft-sensor compensation point, key performance parameters are reconstructed based on the electromechanical-hydraulic energy-mass conservation relationship, and the reconstructed key performance parameters are fed back into the calculation of the cross-physical quantity residual vector, thereby improving the residual calculation accuracy in the undersampled region without adding physical sensors. The soft-sensor reconstruction of the key performance parameters is itself based on the inverse solution of the conservation relationship, which can be obtained by independently measuring physical quantities. This is a known soft-sensor method in the art. Its innovation lies in the fact that the compensation point position is determined by the field strength gradient of the residual space field that violates the conservation relationship.

[0047] Finally, the virtual-real bidirectional drive engine updates the bidirectional synchronization between the motor, pump, valve, pipeline, and water tank and the hybrid twin model. The virtual-real bidirectional drive engine provides a bidirectional synchronization channel between the physical entity and the hybrid twin model, enabling the real-time mapping of the physical entity's operating state to the hybrid twin model, while simultaneously sending the optimization results of the hybrid twin model to the physical entity's execution mechanisms in real time, forming a closed loop. Considering that applying a uniform high-frequency synchronization to all nodes under limited edge-cloud bandwidth would waste resources, this embodiment allocates synchronization frequencies to each node of the physical topology graph based on the field strength distribution of the conservation violation residual space field. Nodes with field strengths higher than a first threshold have their synchronization frequencies increased, while nodes with field strengths lower than a second threshold have their synchronization frequencies decreased, where the first threshold is greater than the second threshold. The synchronization frequency allocation is performed using the following formula:

[0048] ,in: To be assigned to a node The synchronization frequency is a scalar, and its value range is... The unit is Hz, calculated by this formula, representing the synchronization frequency between the physical entity of the node and the hybrid twin model; subscript Number the nodes; The minimum synchronization frequency is a scalar with a value of 1 and a unit of Hz. It is determined by the minimum refresh requirement of the system for low field strength nodes and represents the lower limit of the synchronization frequency. The maximum synchronization frequency is a scalar with a value of 1000 in Hz. It is determined by the system's highest real-time requirements for critical nodes and the upper limit of edge-cloud bandwidth, and represents the upper limit of the synchronization frequency. For nodes The conservation at this point violates the residual space field value, as defined above; Let be the maximum value of the conservation violation residual space field at all nodes, be a scalar, dimensionless, determined by the conservation violation residual space field at the current moment, and be used to normalize the field value to . ; This is a limiting function, indicating that... Limited to the range inside, when Less than Time to take Greater than Time to take Otherwise take The superscript 2 is used to ensure that the allocated synchronization frequency does not exceed the allowable range; the superscript 2 indicates the square. In actual allocation, the above formula is normalized as a whole according to the constraint that the sum of the synchronization frequencies allocated to each node does not exceed the edge cloud bandwidth budget. This ensures that synchronization resources are concentrated on nodes with high field strength that violate the conservation of the residual space field, i.e., nodes that carry the most degraded information. This balances the real-time response of key nodes with the fidelity of global mapping under limited bandwidth.

[0049] After processing in step S5, the conservation violation component is transformed into a device degradation fingerprint and used to drive fault perception. The result of the fault perception is used to generate at least one decision service among device health assessment, fault diagnosis and root cause analysis, remaining life prediction, economic operation scheduling, and energy efficiency optimization. Based on the decision service, a locally deployed industrial big data model generates human-machine interaction information and operation and maintenance work orders. The industrial big data model is locally deployed at the edge of the pumping station, and organizes the structured results of the decision service into natural language to generate interactive information and work orders for operation and maintenance personnel. All outputs are strictly limited to technical analysis and intelligent decision-making results of pumping station operation data, which are used by downstream water conservancy engineers and pumping station operation and maintenance managers in conjunction with engineering specifications.

[0050] To verify the effectiveness of this method under varying operating conditions, it was applied during a switchover from two-pump to three-pump operation at the urban water supply pumping station. During the switchover, the inverter speed increased from 80% to 100% of its rated speed, the valve opening was adjusted from 60% to 90%, and the water level in the tank decreased from 3.5m to 3.2m. During the transient transition between operating conditions, if a fixed conserved constraint manifold is used, the residual vector across physical quantities generated during the transition will be incorrectly projected onto the manifold normal due to the misalignment of response times in each stage. The conservation violation rate jumps instantaneously from 0.12 in the steady state to 0.86, triggering a false alarm. After using the method described in this paper to identify the online parameters of the conserved constraint manifold and perform phase alignment on the residuals, the conservation violation rate during the transition is suppressed to below 0.18, and no false alarm is triggered. However, under a subsequently introduced pump impeller wear condition, the conservation violation rate steadily rises to 0.74, and is accurately located to the corresponding pump node through inversion of the conservation violation residual space field. This embodiment demonstrates that the method can stably distinguish between operating condition drift deviation and equipment degradation deviation under varying operating conditions.

[0051] In addition to its application in urban water supply pumping stations, this method can also be adapted to various application scenarios, such as intelligent operation and maintenance of urban drainage and flood control pumping stations, energy-saving optimization of agricultural irrigation pumping stations, and the construction of digital twins for large-scale water diversion projects. Under different application scenarios, the specific specifications of the motors, pumps, valves, pipelines, and water tanks, the sampling frequency of the sensors, and the number of links in the conserved constraint manifold can be adjusted according to the actual situation of the pumping station. However, all follow the core technical concept of decomposing the conserved constraint manifold parameterized along the operating condition using the residual vector across physical quantities into conserved consistent components and conserved violation components, enabling the conserved consistent components to undergo self-calibration through reinjection, and driving fault perception and virtual-real synchronization using the conserved violation components.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing intelligent sensing for digital twin pumping stations, characterized in that, include: The method is used in a pumping station including a motor, a water pump, valves, pipelines, and a water tank; vibration sensors are installed at the bearing seats of the motor and the water pump to collect vibration signals; temperature sensors are installed at the windings and bearings of the motor to collect temperature signals; pressure sensors are installed at the inlet and outlet of the water pump to collect pressure signals; flow sensors are installed at the main pipe of the water supply pipeline to collect flow signals; and current sensors are installed at the power supply circuit of the motor to collect current signals; the vibration signals, temperature signals, pressure signals, flow signals, and current signals are combined to form a multi-source heterogeneous sensing signal, and the multi-source heterogeneous sensing signal is spatiotemporally aligned to obtain spatiotemporally aligned sensing data. A mechanistic model is constructed based on the electromechanical-hydraulic energy-mass conservation relationship of a pumping station. This relationship includes electromechanical energy conversion, hydraulic work, pipeline head loss, and pool mass balance. The electromechanical energy conversion involves converting the motor's power supply into pump shaft power via electromagnetic conversion and mechanical transmission. The hydraulic work involves converting the pump shaft power into hydraulic power represented by the product of water head and flow rate. The pipeline head loss involves the head loss caused by friction and local resistance in the pipeline. The pool mass balance involves ensuring mass conservation between the inflow and outflow rates and water level changes in the forebay. Mechanism prediction values ​​for each physical quantity are generated from the spatiotemporally aligned sensing data using the mechanistic model, and the cross-physical quantity residual vector between the mechanism prediction values ​​and the spatiotemporally aligned sensing data is calculated. Based on the current operating condition parameters, construct a parameterized conservation constraint manifold. Project the cross-physical quantity residual vector onto the parameterized conservation constraint manifold and decompose it to obtain the conservation consistency component located in the tangent space of the parameterized conservation constraint manifold and the conservation violation component located in the normal of the parameterized conservation constraint manifold. The conserved consistency component is fed back into the hybrid twin model formed by the fusion of the mechanism model and the data-driven model, and the hybrid twin model is subjected to twin self-calibration; the conserved violation component is used as the device degradation fingerprint to drive fault perception, and the bidirectional synchronization between the motor, water pump, valve, pipeline and water tank and the hybrid twin model is updated through the virtual and real bidirectional driving engine.

2. The intelligent sensing construction method for digital twin pumping stations according to claim 1, characterized in that, The cross-physical quantity residual vectors are interconnected according to the coupling coefficients of the electromechanical energy conversion link, the hydraulic work link, the pipeline head loss link, and the pool mass balance link; when the cross-physical quantity residual vectors fall entirely into the tangent space of the operating condition parameterized conservation constraint manifold, it is determined to be an operating condition drift or calibration drift; when the cross-physical quantity residual vectors have a non-zero component in the normal direction of the operating condition parameterized conservation constraint manifold, it is determined to be a local failure.

3. The intelligent sensing construction method for digital twin pumping stations according to claim 2, characterized in that, The operating parameters include inverter speed, valve opening, and water tank level. Constructing the parameterized conservation constraint manifold for the operating conditions includes: online parameter identification of the conservation constraint manifold using the inverter speed, valve opening, and water tank level as parameters, and phase alignment of the cross-physical quantity residual vector during operating condition switching to obtain the parameterized conservation constraint manifold for the operating conditions that matches the operating parameters; and outputting the drift of the identified parameters of the parameterized conservation constraint manifold over time as a calibration aging index.

4. The intelligent sensing construction method for digital twin pumping stations according to claim 3, characterized in that, Also includes: A physical topology diagram is constructed based on the connection relationships of the motor, water pump, valve, pipeline and water tank. The conservation violation components are used to form a conservation violation residual space field at each node of the physical topology diagram. The root cause of the device degradation fingerprint is obtained by inverting the conservation violation residual space field along the propagation direction of the physical topology diagram.

5. The intelligent sensing construction method for digital twin pumping stations according to claim 4, characterized in that, Also includes: Calculate the field intensity gradient of the conservation violation residual space field, select a soft measurement point at the maximum field intensity gradient, and reconstruct the key performance parameters at the soft measurement point based on the electromechanical-hydraulic energy-mass conservation relationship. Then, feed the reconstructed key performance parameters back into the calculation of the cross-physical quantity residual vector.

6. The intelligent sensing construction method for digital twin pumping stations according to claim 5, characterized in that, The update of bidirectional synchronization via the virtual-real bidirectional driving engine includes: allocating synchronization frequencies to each node of the physical topology graph according to the field strength distribution of the conservation violation residual space field; increasing the synchronization frequency for nodes with field strength higher than a first threshold; and decreasing the synchronization frequency for nodes with field strength lower than a second threshold, wherein the first threshold is greater than the second threshold.

7. The intelligent sensing construction method for digital twin pumping stations according to claim 1, characterized in that, The method is deployed in an architecture that integrates a physical entity layer, an edge IoT layer, a digital twin model layer, and an intelligent decision service layer. The physical entity layer includes the motor, water pump, valve, pipeline, and water tank, as well as the vibration sensor, temperature sensor, pressure sensor, flow sensor, and current sensor. The edge IoT layer performs the spatiotemporal alignment. The digital twin model layer carries the hybrid twin model. The intelligent decision service layer outputs the fault perception results.

8. The intelligent sensing construction method for digital twin pumping stations according to claim 1, characterized in that, Before performing spatiotemporal alignment on the multi-source heterogeneous sensing signals, data cleaning, anomaly detection, and missing value imputation are sequentially performed on the multi-source heterogeneous sensing signals at the edge computing node; when the result of the anomaly detection meets the preset local response conditions, the edge computing node performs a local closed-loop response on the motor, water pump, valve, pipeline, and water tank.

9. The intelligent sensing construction method for digital twin pumping stations according to claim 1, characterized in that, The mechanistic model includes a geometric model, a system-level reduced-order model, and a computational fluid dynamics simplified model constructed based on building information model and 3D laser scanning. The data-driven model includes a neural network model and a support vector machine model. When fusing the mechanistic model and the data-driven model, dynamic weights are assigned to the mechanistic model and the data-driven model according to the magnitude of the conserved consistent components.

10. The intelligent sensing construction method for digital twin pumping stations according to claim 1, characterized in that, The results of the fault perception are used to generate at least one decision service among equipment health assessment, fault diagnosis and root cause analysis, remaining life prediction, economic operation scheduling and energy efficiency optimization, and the locally deployed industrial big data model generates human-machine interaction information and operation and maintenance work orders based on the decision service.

Images