Water pump residual life prediction system and method based on large model

Abstract

The application provides a water pump residual life prediction method based on a large model, and the method comprises the following steps: S1, collecting multi-source heterogeneous data of water pump operation in real time through a vibration sensor, a temperature sensor, a pressure sensor and a monitoring unit, wherein the vibration sensor captures axial and radial vibration spectrum, the temperature sensor monitors bearing and sealing cavity temperature gradient change, the pressure sensor records inlet and outlet pressure fluctuation characteristics, and the monitoring unit extracts three-phase current harmonic components; S2, performing lightweight pretreatment on the multi-source heterogeneous original sensing data on an edge computing node, including wavelet transform-based vibration signal denoising processing, temperature and pressure data calibration and normalization according to working condition load segmentation, and filtering transient abnormal data flow through a sliding time window; and S3, inputting the pretreated data flow into a cloud large model platform, and analyzing long-period dependence of the vibration signal through a Transformer encoder in a time sequence feature extraction module.

Description

Technical Field

[0001] The present invention belongs to the field of water pump life prediction, and in particular relates to a water pump remaining life prediction system and method based on a large model. Background Art

[0002] Currently, current industrial water pump remaining life prediction technologies primarily rely on threshold alarms and statistical models, which have significant limitations. At the data acquisition level, traditional methods typically use a single type of sensor (such as monitoring only vibration or temperature). This makes it impossible to fully capture the degradation characteristics of the water pump's multi-physics coupling. For example, the indicative effect of current harmonics on bearing electrical corrosion is ignored. Furthermore, sensor deployment is often limited to the exterior of the pump body, making it difficult to capture the true operating conditions of key areas such as the seal chamber.

[0003] At the data processing level, edge nodes often use fixed-threshold filtering algorithms, which cannot effectively separate water flow impact noise from true fault signals. In particular, transient abnormal data caused by operating condition fluctuations during variable frequency operation can contaminate training samples. Cloud-based models often directly apply general machine learning frameworks (such as SVM or random forests) and lack customized designs for the fluid dynamics characteristics of water pumps. At the model construction level, mainstream data-driven methods rely heavily on the scale of historical fault data. When samples are scarce, prediction accuracy drops sharply. Pure black-box models also have difficulty integrating water pump physical mechanisms (such as the cavitation critical pressure equation and the exponential decay law of bearing wear), resulting in prediction results that violate basic physical conservation laws. A typical case includes the misjudgment of the remaining life of a petrochemical plant's feedwater pump due to failure to consider changes in high-temperature fluid viscosity, leading to unplanned downtime.

[0004] At the feature fusion level, existing technologies often use simple feature splicing to process heterogeneous data such as vibration, temperature, and pressure. They fail to construct a knowledge graph based on the physical topology of the pump, resulting in insufficient temporal and spatial alignment accuracy of different sensor data. For example, the time-delay correlation between the motor current phase and the bearing vibration signal is often overlooked. At the maintenance decision-making level, traditional systems only output point estimates of remaining life without providing confidence intervals, and lack a closed-loop control chain between prediction results and operational optimization. This makes it difficult for maintenance personnel to dynamically adjust the speed based on the degree of wear to avoid resonance zones, resulting in the industry pain point of "accurate prediction but delayed response."

[0005] Furthermore, existing solutions separate edge and cloud functions. The edge performs only simple filtering and uploads raw data, resulting in significant network bandwidth pressure and insufficient real-time performance. This fails to meet the millisecond-level response requirements of high-risk scenarios like nuclear power plants. These shortcomings collectively restrict the large-scale application of predictive technology in industrial settings. Summary of the Invention

[0006] The present invention proposes a large-model-based water pump remaining life prediction system and method to solve the technical problems in water pump remaining life prediction, such as the difficulty in fusing multi-source heterogeneous data, the separation of physical mechanism and data-driven model, poor adaptability to dynamic working conditions, and lack of closed-loop optimization of maintenance decisions, thereby realizing high-precision life prediction and active operation control based on physical law constraints.

[0007] The technical solution of the present invention is achieved as follows: a method for predicting the remaining life of a water pump based on a large model, the method comprising the following steps:

[0008] S1: Multi-source heterogeneous data on water pump operation is collected in real time through vibration sensors, temperature sensors, pressure sensors, and a monitoring unit. The vibration sensor captures the axial and radial vibration spectra, the temperature sensor monitors the temperature gradient changes in the bearing and seal chamber, the pressure sensor records the inlet and outlet pressure fluctuation characteristics, and the monitoring unit extracts the harmonic components of the motor's three-phase current.

[0009] S2: Lightweight preprocessing of multi-source heterogeneous raw sensor data at edge computing nodes, including vibration signal noise reduction based on wavelet transform, temperature and pressure data calibration and normalization based on load conditions, and filtering of transient abnormal data streams using sliding time windows.

[0010] S3: The preprocessed data stream is fed into the cloud-based large model platform. The Transformer encoder in the time series feature extraction module analyzes the long-term dependencies of the vibration signal. The multi-source data fusion module constructs a knowledge graph based on the physical topology of the water pump, achieving spatiotemporal alignment of the four-dimensional features of vibration, temperature, pressure, and current.

[0011] S4: Deeply integrate physical mechanisms and data-driven approaches in the remaining life prediction module: The water pump cavitation characteristic equation and bearing wear dynamics model are pre-embedded in the weight calculation process of the LSTM-attention hybrid network. Degradation-sensitive features are dynamically screened using an attention score matrix constrained by the physical equations. The cavitation characteristic equation is used to calculate the critical pressure fluctuation threshold for cavitation at different speeds, and the bearing wear dynamics model is used to quantify the exponential decay relationship between vibration energy and remaining life. This fusion mechanism enables the model to simultaneously optimize data fitting errors and physical law matching during the training phase.

[0012] S5: Based on the degradation state probability distribution output in step S4, a confidence interval prediction value of the remaining life is generated. When the prediction value is lower than the preset threshold, a maintenance work order is triggered, and the optimal operating speed range is recommended based on the degree of bearing wear.

[0013] There are five core flaws in existing technologies:

[0014] First, data acquisition relies on a single type of sensor (e.g., vibration monitoring only), which is unable to simultaneously capture key degradation characteristics such as the impact of current harmonics on bearing corrosion and the impact of seal chamber temperature on mechanical seal failure. Furthermore, traditional deployment methods make it difficult to obtain the true state of the flow field within the pump.

[0015] Second, edge preprocessing uses fixed threshold filtering, which cannot distinguish between water flow impact noise and early fault signals under variable frequency operation conditions, resulting in transient abnormal data contaminating the model input;

[0016] Third, the cloud-based model mechanically applies general algorithms (such as support vector machines) that are not tailored to the fluid dynamics of the pump. In particular, it ignores the dynamic relationship between cavitation critical pressure and speed and the exponential decay law of bearing wear vibration energy, causing the prediction results to violate the laws of conservation of mass and energy.

[0017] Fourth, feature fusion relies on simple data splicing and lacks a spatiotemporal alignment mechanism based on the pump's physical topology (impeller, shaft, and seal chamber). This results in the loss of the time-delayed correlation between the motor current phase and the bearing vibration signal, and the inability to coordinate multi-source data.

[0018] Fifth, maintenance decisions only output a point estimate of remaining life, without providing a confidence interval to quantify prediction uncertainty, and without establishing a closed-loop "prediction-control" link, making it impossible to dynamically adjust the speed to avoid the resonance zone based on the degree of bearing wear. This solution covers all degradation paths through a four-dimensional vibration / temperature / pressure / current sensing network. Edge-side wavelet noise reduction and adaptive calibration of working conditions suppress transient noise. A physical topology knowledge graph is constructed in the cloud to achieve multi-source feature phase-sensitive delay compensation. More innovatively, the cavitation equation and bearing wear model are embedded in the LSTM-attention network weight calculation. Degradation-sensitive features are screened through a physically constrained attention matrix, and finally a confidence interval is output to guide the speed optimization closed loop. This overcomes the three major industry pain points of integrating physical laws with data-driven development, dynamic working condition adaptation, and delayed maintenance decisions.

[0019] As a preferred implementation, after recommending the optimal operating speed range, step S5 updates the key data of the water pump model through an incremental learning mechanism, freezes the shared feature coding layer when a new water pump model is added, and performs zero-sample data migration through parameter-isolated incremental learning.

[0020] As a preferred embodiment, the data acquisition in step S1 is performed by adding the following monitoring units for deeply buried high-pressure pipelines: a distributed fiber optic acoustic sensor array is deployed circumferentially on the outer surface of the pipe wall to capture the micro-strain of the pipe structure through phase-sensitive optical time-domain reflection technology; a piezoelectric bolt preload monitoring gasket is installed at the flange connection to track the risk of sealing failure in real time; and an infrasonic sensor array is set in the valve well to identify the precursors of water hammer events based on the low-frequency sound wave spectrum characteristics.

[0021] As a preferred embodiment, the lightweight preprocessing of step S2 performs formation disturbance compensation, adopts an inertial measurement unit and a rigid connection between the pipe body to form a kinematic constraint system to solve the three-dimensional deformation trajectory of the pipeline in real time; separates the water flow impact vibration and the ground pulsation noise caused by the surrounding subway operation through wavelet packet decomposition; and a settlement predictor based on Kalman filtering automatically marks the pipe section as a high-risk monitoring area when an uneven settlement trend is detected.

[0022] As a preferred implementation, the triggering maintenance work order of step S5 is associated with the city's public safety network. When the predicted remaining life is lower than a threshold, the distributed optical fiber settlement monitoring network of the adjacent road is automatically activated, and the surface deformation field above the pipeline is constructed based on the optical frequency shift. If the deformation field shows shear mutation characteristics, the traffic signal system is linked to implement road closure, and the trenchless repair robot positioning program is started simultaneously.

[0023] A large-scale model-based water pump remaining life prediction system, including a multi-source sensing unit, an edge pre-processing device, a cloud-based physical embedded prediction platform, and a maintenance decision actuator;

[0024] The multi-source sensing unit consists of a vibration sensor for collecting axial and radial vibration spectra, a temperature sensor for monitoring the temperature gradient of the bearing and seal chamber, a pressure sensor for recording inlet and outlet pressure fluctuations, and a monitoring unit for extracting the harmonic components of the motor's three-phase current.

[0025] The edge preprocessing device is deployed on the explosion-proof industrial controller on the pump equipment side. It has a built-in wavelet transform noise reduction module that performs frequency domain filtering on vibration signals. It integrates a load segmentation calibration algorithm to achieve adaptive normalization of temperature and pressure data, and configures a sliding time window filter to eliminate transient abnormal data streams.

[0026] The cloud-based physical embedding prediction platform includes a Transformer encoder in the time series feature extraction module to analyze the long-period dependencies of vibration signals. The multi-source data fusion module constructs a knowledge graph based on the physical topology of the water pump and achieves spatiotemporal alignment of the four-dimensional features of vibration, temperature, pressure, and current. The remaining life prediction module embeds the water pump cavitation characteristic equation and the bearing wear dynamics model into the weight calculation process of the LSTM-attention hybrid network. The attention score matrix constrained by the physical equations dynamically screens degradation-sensitive features. The cavitation characteristic equation calculates the critical pressure fluctuation threshold for cavitation at different speeds, and the bearing wear dynamics model quantifies the exponential decay relationship between vibration energy and remaining life.

[0027] The maintenance decision executor receives the degradation state probability distribution output by the remaining life prediction module and generates a confidence interval prediction value. When the prediction value is lower than the preset threshold, it automatically triggers the maintenance work order system and simultaneously outputs the optimal operating speed range control instruction matching the bearing wear degree to the water pump inverter.

[0028] After adopting the above technical solution, the beneficial effects of the present invention are as follows: This solution creates four core values: First, prediction accuracy is significantly improved. The physical mechanism embedding reduces the model error in small sample scenarios (for example, eliminating misjudgments caused by viscosity changes in high-temperature conditions). The bearing wear dynamics constraint ensures that the life decay curve conforms to the exponential law. The confidence interval output quantifies the prediction reliability and avoids the mismatch of maintenance resources caused by single-point estimation.

[0029] Secondly, its adaptability to dynamic operating conditions has been significantly enhanced. Edge-side sliding time window filtering and segmented normalization of operating conditions effectively suppress transient interference caused by variable-frequency operation. Knowledge-graph-driven multi-source spatiotemporal alignment accurately captures the phase coupling relationship between current and vibration signals. The Transformer encoder's ability to extract long-period degradation features is superior to traditional LSTM.

[0030] Secondly, maintenance costs are revolutionized. Wear-based speed recommendations link the inverter to avoid resonance zones, extending bearing life. A maintenance work order trigger mechanism, combined with a spare parts inventory system, enables precise resource scheduling, reducing downtime compared to scheduled maintenance. Finally, the technology's universal applicability breaks through industry barriers. The physically constrained attention mechanism can be applied to all pump types, including centrifugal and positive displacement pumps. The edge-cloud collaborative architecture meets the millisecond-level response requirements of nuclear power plants, and the federated learning framework supports cross-site knowledge sharing. Typical cases demonstrate a reduction in false alarm rates for high-pressure water injection pumps in petrochemical plants, with unplanned downtime reduced to zero. Speed ​​optimization of cooling pumps in nuclear power plants saves over one million kWh of electricity annually.

[0031] In summary, this solution takes the "data-physics" dual drive as its core, opens up the perception-decision-control closed loop, and provides a standardized paradigm for intelligent operation and maintenance of process industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0033] Figure 1 This is a flow chart of the system of the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0035] Example:

[0036] like Figure 1 As shown, a method for predicting the remaining life of a water pump based on a large model includes the following steps:

[0037] S1: Multi-source heterogeneous data on water pump operation is collected in real time through vibration sensors, temperature sensors, pressure sensors, and a monitoring unit. The vibration sensor captures the axial and radial vibration spectra, the temperature sensor monitors the temperature gradient changes in the bearing and seal chamber, the pressure sensor records the inlet and outlet pressure fluctuation characteristics, and the monitoring unit extracts the harmonic components of the motor's three-phase current.

[0038] S2: Lightweight preprocessing of multi-source heterogeneous raw sensor data at edge computing nodes, including vibration signal noise reduction based on wavelet transform, temperature and pressure data calibration and normalization based on load conditions, and filtering of transient abnormal data streams using sliding time windows.

[0039] S3: The preprocessed data stream is fed into the cloud-based large model platform. The Transformer encoder in the time series feature extraction module analyzes the long-term dependencies of the vibration signal. The multi-source data fusion module constructs a knowledge graph based on the physical topology of the water pump, achieving spatiotemporal alignment of the four-dimensional features of vibration, temperature, pressure, and current.

[0040] S4: Deeply integrate physical mechanisms and data-driven approaches in the remaining life prediction module: The water pump cavitation characteristic equation and bearing wear dynamics model are pre-embedded in the weight calculation process of the LSTM-attention hybrid network. Degradation-sensitive features are dynamically screened using an attention score matrix constrained by the physical equations. The cavitation characteristic equation is used to calculate the critical pressure fluctuation threshold for cavitation at different speeds, and the bearing wear dynamics model is used to quantify the exponential decay relationship between vibration energy and remaining life. This fusion mechanism enables the model to simultaneously optimize data fitting errors and physical law matching during the training phase.

[0041] S5: Based on the degradation state probability distribution output in step S4, a confidence interval prediction value of the remaining life is generated. When the prediction value is lower than the preset threshold, a maintenance work order is triggered, and the optimal operating speed range is recommended based on the degree of bearing wear.

[0042] In the municipal water supply network's secondary booster pump station scenario, this solution uses a sensor array deployed on the high-pressure centrifugal pump group to capture the multi-physics operating status in real time. Axial and radial vibration sensors are magnetically mounted orthogonally to the pump bearing seat, synchronously collecting vibration spectrum characteristics within the 10Hz-10kHz frequency band to detect potential impeller dynamic balance degradation and bearing raceway spalling. Temperature sensors, embedded in the bearing outer ring and mechanical seal flushing chamber via pre-embedded thermocouples, monitor the gradient temperature rise caused by high-temperature carbonization of grease and dry friction on the seal surface. Pressure sensors, using a diaphragm-isolated structure at the pump inlet and outlet flanges, record pressure pulsations and water hammer impacts caused by fluctuations in water consumption in the pipeline network. The monitoring unit, integrated into the motor control cabinet, extracts the 5th and 7th harmonic components of the stator current using Rogowski coils to identify electromagnetic torque pulsations caused by rotor bar fracture.

[0043] The edge computing node uses an IP67-rated industrial gateway. Its built-in wavelet packet decomposition algorithm divides the raw vibration signal into eight frequency bands, eliminating high-frequency noise (>5kHz) caused by water flow impact caused by sudden changes in pipe network pressure. It also dynamically divides the operating load range based on the speed signal output by the water pump inverter. It performs segmented normalization on temperature and pressure data to eliminate dimensional differences caused by peaks and valleys in water consumption during the day and night. A sliding time window mechanism is used to filter out transient current anomalies caused by grid voltage sags.

[0044] The preprocessed data stream is transmitted to the cloud platform via an industrial ring network. The time series feature extraction module uses a stacked Transformer encoder to analyze the long-term degradation trend of the vibration signal and a self-attention mechanism to capture the correlation pattern of the slowly increasing characteristic frequency amplitude during bearing wear. The multi-source data fusion module constructs a knowledge graph centered on the pump rotor dynamics topology, aligns the motor current phase signal with the bearing vibration pulse for time delay compensation, and simultaneously constructs a thermal-mechanical coupling feature vector by correlating the seal chamber temperature gradient with the inlet and outlet pressure difference, achieving spatiotemporal collaborative representation of four-dimensional heterogeneous data. The remaining life prediction module innovatively integrates the physical constraints of the municipal water supply system. The NPSHr (Net Positive Net Shock Head) calculation formula in the pump cavitation characteristic equation is embedded in the weight update process of the forget gate of the LSTM network. The critical threshold for cavitation occurrence is dynamically corrected based on the real-time speed and inlet pressure. When the pressure fluctuation amplitude exceeds the threshold, the attention score of the corresponding time step is forcibly increased to focus on the cavitation damage characteristics. The exponential decay relationship between vibration energy and life in the bearing wear dynamics model is simultaneously converted into a physical regularization term in the loss function. During the backpropagation process, the predicted life value is constrained to conform to the material fatigue crack growth law.

[0045] The final output of the degradation state probability distribution generates a remaining life confidence interval prediction value. When the upper limit of the interval falls below the preset threshold, the work order system is automatically triggered and pushed to the municipal operation and maintenance mobile terminal. At the same time, based on the resonant frequency band distribution in the bearing wear characteristic spectrum, the optimal operating range instruction is generated to avoid the critical speed (such as 1.2 times the operating frequency). It is sent to the water pump variable frequency actuator via the OPC UA protocol to realize closed-loop speed control.

[0046] During daily operation and maintenance, when a new water pump model is put into operation, the system automatically freezes the Transformer feature encoding layer weights shared in the cloud platform, and only unfreezes the physical constraint sub-network specific to the equipment (including cavitation equations and bearing wear model parameters). The degraded feature vector of the basic model is called as the initial value through the parameter isolation mechanism, and migration training is performed based on the vibration / current baseline data of the new pump during the no-load test phase. During this period, the edge controller continuously collects 72 hours of running-in period operation data and transmits it to the cloud in encrypted form. The resonance frequency band distribution and cavitation sensitive interval characteristics of the new pump are extracted through the knowledge distillation framework. The updated model can output life prediction under zero-fault sample conditions. Water personnel verify through mobile terminals that the prediction curve is consistent with the theoretical fatigue life in the equipment manual and then enable formal monitoring. The entire process does not require stopping the pump or manual labeling.

[0047] For the municipal water main buried 20 meters underground, during the construction phase, distributed fiber-optic acoustic sensor cables were circumferentially fused to the pipe wall. Phase-sensitive optical time-domain reflectometry technology was used to capture micro-strain changes in the pipe body at the 0.1με level in real time. Piezoelectric ceramic preload monitoring gaskets were embedded in the bolt holes on the flange connection surfaces to convert the contact pressure of the sealing surface into a charge signal, which was wirelessly transmitted to the wellhead hub. An infrasound sensor array with a bandwidth of 10-50Hz was installed on the sidewall of the valve well to identify the precursor characteristics of water hammer pressure waves based on the propagation time difference of low-frequency sound waves in the fluid. During routine inspections, operations and maintenance personnel use handheld readers to verify whether the flange bolt stress values ​​are within the design range of 120-150kN. When the cloud platform alarm indicates that the infrasound energy in a certain pipe section exceeds the threshold, personnel are immediately dispatched to verify whether the pressure reducing valve opening is abnormal.

[0048] In the water supply pipe section passing through the subway tunnel, the MEMS inertial measurement unit is rigidly fixed to the pipe wall surface. The high-frequency jitter of the pipe body caused by subway vibration (>30Hz) is separated from the low-frequency deformation caused by actual settlement (<1Hz) through six-degree-of-freedom motion solution; the wavelet packet decomposition module built into the edge node divides the vibration signal into 32 sub-bands, actively suppressing the 63Hz characteristic noise generated by the subway wheel-rail impact; the Kalman filter settlement predictor integrates the data of the earth pressure sensor around the pipe to construct a deformation dynamics model. When the differential settlement rate at the pipe joint exceeds 0.1mm / week, it is automatically marked as a red high-risk area on the digital twin platform and triggers a high-definition pipe endoscope review.

[0049] When the remaining life of a deep buried pipe section is predicted to be less than 3 years, the system automatically wakes up the Brillouin scattering optical fiber monitoring network buried along the pipeline, generates the ground deformation contour map by analyzing the laser frequency shift amount; if the deformation field shows that there is a 5mm sudden settlement tank within 2m above the pipeline, immediately send regional control instructions to the traffic signal control system through the OPC UA protocol, start the bypass sign 1 hour before the morning and evening peak and adjust the red and green light timing; send GIS coordinates and pipe diameter parameters to the trenchless repair robot at the same time, the robot autonomously navigates to the work well site and expands the memory alloy repair air bag to execute local reinforcement, and the whole process data is synchronized to the municipal emergency command large screen.

[0050] In the daily operation of the municipal water supply pump station, the explosion-proof edge controller continuously receives four-dimensional sensor original signals: the vibration sensor collects the frequency sideband modulation characteristics of the impeller through the magnetic base, the temperature sensor monitors whether the temperature difference of the flushing liquid in the sealed cavity exceeds the 5℃ warning line, the pressure sensor captures the 0.2s pulse width shock wave caused by water hammer, and the monitoring unit analyzes the amplitude and phase shift of the stator third harmonic; the edge layer executes wavelet denoising to eliminate power grid flicker interference, normalizes temperature and pressure data according to the speed gear of the frequency converter, and filters current burr caused by lightning through a sliding time window; after preprocessing, the data is uploaded to the cloud through the industrial ring network, the Transformer encoder extracts the bearing wear trend for up to 90 days from the vibration signal, the knowledge graph aligns the millisecond level time delay of the current harmonic phase and the bearing vibration pulse; the physical embedded prediction core converts the cavitation NPSHr equation into the LSTM forget gate constraint condition, strengthens the corresponding frequency spectrum weight when the inlet pressure is lower than the critical value of the current speed, and the bearing wear model reversely constrains the life attenuation curve through the vibration energy envelope integral value; the maintenance decision executor outputs the remaining life confidence interval, if the lower limit value touches the threshold, it pushes the "second warning" instruction to the work order system and links the SCADA system to adjust the speed to avoid the resonance band, and generates a spare parts procurement list synchronized to the material management platform.

[0051] The above merely describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting the remaining life of a water pump based on a large model, characterized in that: The method comprises the following steps: S1: Multi-source heterogeneous data on water pump operation is collected in real time through vibration sensors, temperature sensors, pressure sensors, and a monitoring unit. The vibration sensor captures the axial and radial vibration spectra, the temperature sensor monitors the temperature gradient changes in the bearing and seal chamber, the pressure sensor records the inlet and outlet pressure fluctuation characteristics, and the monitoring unit extracts the harmonic components of the motor's three-phase current. S2: Lightweight preprocessing of multi-source heterogeneous raw sensor data at edge computing nodes. This includes vibration signal noise reduction based on wavelet transform, temperature and pressure data calibration and normalization based on load conditions, and filtering of transient current anomalies caused by grid voltage sags using a sliding time window mechanism. S3: The preprocessed data stream is fed into the cloud-based large model platform. The Transformer encoder in the time series feature extraction module analyzes the long-term dependencies of the vibration signal. The multi-source data fusion module constructs a knowledge graph based on the physical topology of the water pump, achieving spatiotemporal alignment of the four-dimensional features of vibration, temperature, pressure, and current. S4: Deeply integrate physical mechanisms and data-driven approaches in the remaining life prediction module. The water pump cavitation characteristic equation and bearing wear dynamics model are pre-embedded in the weight calculation process of the LSTM-attention hybrid network. Degradation-sensitive features are dynamically screened using an attention score matrix constrained by the physical equations. The cavitation characteristic equation is used to calculate the critical pressure fluctuation threshold for cavitation at different speeds, and the bearing wear dynamics model is used to quantify the exponential decay relationship between vibration energy and remaining life. This deep integration of physical mechanisms and data-driven approaches enables the model to simultaneously optimize data fitting errors and physical law matching during the training phase. S5: Based on the degradation state probability distribution output in step S4, a confidence interval prediction value of the remaining life is generated. When the prediction value is lower than the preset threshold, a maintenance work order is triggered, and the optimal operating speed range is recommended based on the degree of bearing wear.

2. The method for predicting the remaining life of a water pump based on a large model according to claim 1, wherein: After recommending the optimal operating speed range, step S5 updates the key data of the water pump model through an incremental learning mechanism, freezes the shared feature encoding layer when a new water pump model is added, and performs zero-sample data migration through parameter-isolated incremental learning.

3. The method for predicting the remaining life of a water pump based on a large model according to claim 1, wherein: The data acquisition in step S1 is performed by adding the following monitoring units for deeply buried high-pressure pipelines: a distributed fiber optic acoustic sensor array is deployed circumferentially on the outer surface of the pipe wall to capture the micro-strain of the pipe structure through phase-sensitive optical time-domain reflectometry technology; a piezoelectric bolt preload monitoring gasket is installed at the flange connection to track the risk of sealing failure in real time; and an infrasonic sensor array is set in the valve well to identify the precursors of water hammer events based on the low-frequency sound wave spectrum characteristics.

4. The method for predicting the remaining life of a water pump based on a large model according to claim 1, characterized in that: The lightweight preprocessing in step S2 performs ground disturbance compensation, adopts a kinematic constraint system formed by a rigid connection between an inertial measurement unit and the pipe body to solve the three-dimensional deformation trajectory of the pipeline in real time; separates the water flow impact vibration from the ground pulsation noise caused by the surrounding subway operation through wavelet packet decomposition; and uses a Kalman filter-based settlement predictor to automatically mark a pipe section as a high-risk monitoring area when an uneven settlement trend is detected.

5. The method for predicting the remaining life of a water pump based on a large model according to claim 1, characterized in that: The maintenance work order triggered in step S5 is linked to the city's public safety network. When the predicted remaining life is lower than a threshold, the distributed optical fiber settlement monitoring network of the adjacent road is automatically activated, and a surface deformation field above the pipeline is constructed based on the optical frequency shift. If the deformation field shows shear mutation characteristics, the traffic signal system is linked to implement road closure and the trenchless repair robot positioning program is simultaneously started.

6. The water pump remaining life prediction system based on the large model is characterized by: It includes a multi-source sensing unit, an edge pre-processing device, a cloud-based physical embedded prediction platform, and a maintenance decision executor; The multi-source sensing unit consists of a vibration sensor for collecting axial and radial vibration spectra, a temperature sensor for monitoring the temperature gradient of the bearing and seal chamber, a pressure sensor for recording inlet and outlet pressure fluctuations, and a monitoring unit for extracting the harmonic components of the motor's three-phase current. The edge preprocessing device is deployed on an explosion-proof industrial controller on the pump equipment side. It uses a built-in wavelet transform noise reduction module to perform frequency-domain filtering on vibration signals. It integrates a load-segmented calibration algorithm for operating conditions to achieve adaptive normalization of temperature and pressure data, and uses a sliding time window mechanism to filter transient current anomalies caused by grid voltage sags. The cloud-based physical embedding prediction platform includes a Transformer encoder in the time series feature extraction module to analyze the long-period dependencies of vibration signals. The multi-source data fusion module constructs a knowledge graph based on the physical topology of the water pump and achieves spatiotemporal alignment of the four-dimensional features of vibration, temperature, pressure, and current. The remaining life prediction module embeds the water pump cavitation characteristic equation and the bearing wear dynamics model into the weight calculation process of the LSTM-attention hybrid network. The attention score matrix constrained by the physical equations dynamically screens degradation-sensitive features. The cavitation characteristic equation calculates the critical pressure fluctuation threshold for cavitation at different speeds, and the bearing wear dynamics model quantifies the exponential decay relationship between vibration energy and remaining life. The maintenance decision executor receives the degradation state probability distribution output by the remaining life prediction module and generates a confidence interval prediction value. When the prediction value is lower than the preset threshold, it automatically triggers the maintenance work order system and simultaneously outputs the optimal operating speed range control instruction matching the bearing wear degree to the water pump inverter.

Images