A method for predicting the residual life of a coal-to-oil machine pump

Abstract

This invention discloses a method for predicting the remaining life of a coal-to-oil pump, belonging to the field of equipment health management and life prediction technology; it includes: S1: constructing a five-field coupled degradation mechanism model and establishing a damage evolution equation that can quantitatively describe the interaction of five factors: high temperature, high pressure, erosion, corrosion, and fatigue; S2: constructing a physical information neural network prediction model, embedding the five-field coupled model of the coal-to-oil pump constructed in S1 into the neural network loss function to improve the model's generalization ability under small sample conditions; S3: constructing a multi-scale, multi-physics field real-time coupled digital twin for generating virtual fault data and calibrating the online prediction model through virtual-real synchronous evolution; S4: constructing an edge-cloud collaborative computing architecture model to predict the real-time remaining life of the coal-to-oil pump and provide optimized maintenance decisions.

Description

Technical Field

[0001] This invention belongs to the field of equipment health management and life prediction technology, specifically relating to a method for predicting the remaining life of a coal-to-oil pump. Background Technology

[0002] Coal-to-oil pumps operate under extremely harsh conditions: the operating temperature of the coal-oil slurry pump can reach over 450℃, the pressure exceeds 20MPa, and the solid particle content of the medium is as high as 40%~50%, while also containing highly corrosive gases such as H2S and CO2. This special operating environment leads to extremely complex failure mechanisms for the pumps, resulting in frequent malfunctions such as erosion and wear of flow components, fatigue fracture of moving parts, mechanical seal leakage, and bearing failure. Statistics show that unplanned downtime of pumps in coal-to-oil plants accounts for more than 35% of total downtime, and the direct economic loss caused by a single critical pump failure can reach several million yuan per day.

[0003] Equipment remaining life prediction technology is a core technology for shifting from "reactive maintenance" to "predictive maintenance." It can predict equipment failures in advance, rationally schedule maintenance plans, and significantly reduce the risk of unplanned downtime and maintenance costs. However, current methods for predicting the remaining life of general-purpose pumps have the following limitations when applied to the special operating conditions of coal-to-oil pumps: 1) Severely Distorted Failure Mechanism Description: General pump residual life prediction models typically consider only one or a few failure modes, while the failure of coal-to-oil pumps is the result of the coupled effects of five fields: high temperature, high pressure, erosion, corrosion, and fatigue. Existing physical models do not consider the influence of material thermal softening and plastic strengthening effects at high temperatures on wear rates, and also lack mathematical models that can quantitatively describe the interaction between corrosion and fatigue, resulting in prediction errors generally exceeding 25%, which cannot meet the requirements of industrial applications.

[0004] 2) The problem of small sample size is extremely prominent: A single coal-to-oil pump is worth over ten million yuan, and downtime due to failure results in huge losses. Companies typically perform preventative replacements before serious equipment failures occur, leading to an extreme scarcity of fault data throughout the entire lifecycle. Simultaneously, the low reliability and poor data quality of sensors under high temperature and high pressure environments further exacerbate the data scarcity problem. Purely data-driven models exhibit poor generalization ability under small sample conditions and are prone to overfitting.

[0005] 3) Poor adaptability to dynamic operating conditions: During coal-to-oil production, the operating parameters of pumps, such as flow rate, pressure, and temperature, can fluctuate by more than 30% due to changes in raw coal quality and load adjustments. Most existing models are based on steady-state assumptions, and their prediction accuracy drops sharply when operating conditions change abruptly, sometimes even resulting in completely incorrect predictions. Summary of the Invention

[0006] This invention provides a method for predicting the remaining life of coal-to-oil pumps. Addressing the industry pain point of predicting the remaining life of coal-to-oil pumps under extreme operating conditions, this invention proposes a four-layer hybrid technical solution: a dedicated physical model as a constraint, a physical information neural network as the core, a multi-scale digital twin as a carrier, and edge-cloud collaboration as support. The solution first constructs a five-field coupled degradation mechanism model unique to coal-to-oil pumps, encompassing high temperature, high pressure, erosion, corrosion, and fatigue. This model is then embedded as a physical constraint into the neural network loss function, forming a small-sample adaptive physical information neural network. Simultaneously, a multi-scale, multi-physics field digital twin is built, ranging from microscopic material damage to macroscopic overall machine performance, generating virtual fault data to supplement insufficient real data. Finally, through an edge-cloud collaborative architecture, real-time data processing, online model updates, and intelligent maintenance decisions with confidence levels are achieved.

[0007] To achieve the above-mentioned technical objectives, the present invention is implemented through the following technical solution: A method for predicting the remaining life of a coal-to-oil pump includes: S1: A five-field coupled degradation mechanism model was constructed, and a damage evolution equation capable of quantitatively describing the interaction of five factors—high temperature, high pressure, erosion, corrosion, and fatigue—was established, including: The high-temperature modified erosion wear model is based on the traditional Archard wear model and is modified by introducing the material thermal softening coefficient and plastic strengthening coefficient to reflect the dynamic changes of material hardness and yield strength at high temperature. A corrosion-fatigue interactive damage model was developed, and an electrochemical corrosion-mechanical fatigue coupled damage evolution equation was established to distinguish between two mechanisms: corrosion-accelerated fatigue and fatigue-accelerated corrosion. The dynamic load fatigue cumulative damage model improves upon the traditional Miner linear cumulative damage theory by introducing a load sequence effect coefficient and a working condition fluctuation coefficient. The dynamic load spectrum is processed using the rainflow counting method, and combined with the temperature correction of the material SN curve, the fatigue damage accumulation rate under different working conditions is calculated. S2: Construct a physical information neural network prediction model, and embed the five-field coupling model of coal-to-oil pump constructed in S1 into the neural network loss function to improve the generalization ability of the model under small sample conditions. S3: Construct a multi-scale, multi-physics field real-time coupled digital twin, building a multi-scale digital twin from microscopic material damage to macroscopic overall machine performance, used for virtual-real synchronous evolution to generate virtual fault data and calibrate online prediction models; and provide optimized maintenance decisions; S4: Real-time lifespan prediction based on edge-cloud collaborative architecture. Construct an edge-cloud collaborative computing architecture model to predict the real-time remaining lifespan of coal-to-oil pumps.

[0008] Preferably, the high-temperature modified erosion wear model, based on the Archard wear model, considers the shape factor and concentration distribution of oil-coal slurry particles, and corrects the influence of particle impact angle and velocity on the wear rate. The corrected model yields the following formula, which is used to calculate the erosion wear rate W (mm / a) of the flow components of the coal-to-oil pump, such as impeller blades, pump casing inner wall, sealing ring, plunger, valve seat, etc., at any spatial position under specific operating conditions: Where K is the overall wear coefficient; Temperature-dependent material hardness; It is a function of the impact angle; Let be the particle concentration distribution function; This is the reference hardness at room temperature. The normal impact force of the particles is the normal component of a single oil-coal slurry particle when it vertically impacts the material surface. The unit is N, and it is obtained by CFD solid-liquid two-phase flow simulation calculation. The relative impact velocity of the particles; It represents the nominal hardness of the material, the same as the value of H0, and is used for dimensionless processing in formulas.

[0009] Preferably, the corrosion-fatigue interactive damage model clearly distinguishes and quantitatively describes two fundamentally different interaction mechanisms, namely: Mechanism A: Corrosion accelerates fatigue, contributing about 70% of the total damage as the dominant mechanism: corrosion pits generated by electrochemical corrosion act as stress concentration sources, accelerating crack initiation, while H2S / CO2 media enter the crack tip, accelerating crack propagation. Mechanism B: Fatigue accelerates corrosion as a feedback mechanism: Cyclic loading causes the corrosion product film to rupture periodically, exposing fresh metal surfaces and significantly increasing the local corrosion rate. The final output of the model is the total corrosion-fatigue coupled damage D_total at any time. When D_total reaches the critical damage threshold D_cr, which is usually taken as 0.8~0.9, the component is judged to have suffered corrosion fatigue failure. 1) The process of establishing the basic model is as follows: a. First, establish a pure corrosion rate model under stress-free conditions. in, The pure corrosion current density (A / m²) For exchange current density; The activation energy of the reaction is expressed in J / mol. and They are respectively as well as The partial pressure (MPa); a and b are the reaction order, which are usually taken as a = 0.5~0.8 and b = 0.3~0.6 under coal-to-oil conditions; Pure corrosion rate: in, The molar mass of the metal; The number of electrons in the reaction; It is Faraday's constant; The density of the material; b. Pure fatigue crack propagation model The fatigue crack propagation rate under corrosion-free conditions is described using the Paris equation: in, The crack propagation rate per cycle (m / cycle); Stress intensity factor amplitude (MPa·m) 0.5 ); and These are material constants, obtained through room temperature fatigue tests; 2) Realization of corrosion-accelerated fatigue mechanism a. Corrosion pit-induced crack initiation model Corrosion pits generate significant stress concentrations. The stress concentration factor K_t is related to the pit depth a and radius r as follows: Introducing an equivalent initial crack size a_eq for corrosion pits: The crack initiation life induced by corrosion pits was calculated using Miner's damage theory: ,in, The fatigue life corresponding to the initial crack size a_eq; b. Corrosive Media Accelerated Crack Propagation Model The corrosion fatigue factor (CFF) is introduced to quantitatively describe the accelerating effect of H2S / CO2 on crack propagation rate: Where A is the material-medium coupling coefficient, which is determined by high temperature and high pressure corrosion fatigue test; For reference corrosion current density; It is a correction function for stress intensity factor, temperature, and pressure; For typical coal-to-oil operating conditions, the specific form of the correction function is as follows: in, This is the corrosion fatigue threshold value; It is the activation energy for corrosion fatigue; , This is an experimental constant; Corrected corrosion fatigue crack propagation rate: ; 3) Realization of fatigue-accelerated corrosion mechanism a. Dynamic growth model of corrosion product film Under high temperature and pressure, a corrosion product film composed of FeS, FeCO3, etc., will form on the metal surface. The growth of the film follows parabolic dynamics. in, Let t be the film thickness (m) at time t; The steady-state film thickness; The membrane growth time constant; The protective effect of a membrane is represented by the membrane protection coefficient η: ,in The film density coefficient; Corrosion rate considering the protective effect of the film: ; b. Criteria and Repair Model for Membrane Rupture under Cyclic Loading When the strain generated by cyclic stress exceeds the fracture strain of the corrosion product film... e f At that time, the membrane ruptured: e max > e f After the film ruptures, the fresh metal surface is exposed, and the corrosion rate instantly recovers to near the pure corrosion rate v. corr,0 Then the film begins to regrow, and the corrosion rate gradually decreases. Introducing membrane rupture area fraction ϕ Describe the degree of damage to the membrane under cyclic loading: in, This represents the initial fracture area fraction. , This is an experimental constant; This represents the number of loop iterations. Corrected actual corrosion rate: .

[0010] Preferably, the dynamic load fatigue cumulative damage model finally outputs the fatigue cumulative damage D_fatigue at any time. When D_fatigue reaches the critical damage threshold D_cr, the threshold is usually taken as 0.7~0.8. Considering the coupling effect of corrosion and wear, the component is determined to have fatigue failure.

[0011] Preferably, the construction and implementation process of the dynamic load fatigue cumulative damage model is as follows: 1) Implementation of dynamic load spectrum preprocessing and rainflow counting a. Load spectrum acquisition and preprocessing: The fatigue load of the coal-to-oil pump mainly comes from: oil-coal slurry pressure pulsation, with a main frequency of 10~50Hz; periodic load generated by impeller rotation, with a main frequency equal to the rotation speed; and low-frequency load fluctuations caused by changes in operating conditions, 0.01~1Hz. Raw data were collected from pump outlet pressure sensor, bearing vibration sensor, and motor current sensor, with a sampling frequency ≥1kHz; high-frequency noise and electromagnetic interference were removed using wavelet threshold denoising; and the multi-source data were fused to convert it into an equivalent stress load spectrum. in, , , The stress conversion factor is calibrated through finite element analysis; b. Improved rainflow counting: The four-point rainflow counting method is used to process the non-stationary dynamic load spectrum and extract the amplitude, mean and cycle number of each load level; Improved for coal-to-oil: The working condition marking function is added to associate each load cycle with the corresponding temperature, pressure and flow rate working condition parameters to provide a basis for subsequent working condition fluctuation correction. 2) Temperature-corrected material SN curve model a. Basic SN curve The fatigue life of materials at room temperature follows a power function relationship: in, and m are material constants, determined by standard fatigue tests; b. High-temperature SN curve The operating temperature (350~450℃) of the coal-to-oil pump leads to a significant decrease in the fatigue strength of the material and generates thermal fatigue effects. Therefore, a temperature correction factor K_T is introduced. in, The fatigue limit at temperature T; This represents the fatigue limit at room temperature. The melting point of the material; The temperature sensitivity coefficient is determined through high-temperature fatigue testing. For commonly used wear-resistant alloy steels, Take a value of 0.3~0.5; Corrected high-temperature SN curve: c. Mean stress correction The effect of mean stress on fatigue life is corrected using the Goodman formula: in, The tensile strength of the material at temperature T; 3) Improved fatigue cumulative damage model a. Introduction of the load sequence effect coefficient Traditional Miner's Theory: in, Let be the number of cycles for the i-th level load; This represents the fatigue life under this load level. High loads generate residual compressive stress within the material, slowing down the accumulation of damage under subsequent low loads; while low loads generate microcracks, accelerating the propagation of damage under subsequent high loads. Introducing the load sequence effect coefficient λ ij Describe the effect of the i-th level load on the damage caused by the j-th level load: in, , The load sequence sensitivity coefficient was determined by variable amplitude fatigue test, with β=0.2~0.4 and γ=0.3~0.5. This represents the maximum stress in the load spectrum; b. Introducing a working condition fluctuation coefficient: During coal-to-oil production, frequent switching of working conditions can cause changes in the statistical characteristics of the load spectrum, accelerating fatigue damage; introducing a working condition fluctuation coefficient... : in, The standard deviation of the load amplitude; is the average load amplitude; f (times / hour) is the operating condition switching frequency; The sensitivity coefficient to operating condition fluctuations is calibrated using field data. =0.1~0.3; c. Improved cumulative damage calculation, taking into account both load sequence effect and working condition fluctuation effect, the improved fatigue cumulative damage formula is as follows: The calculation logic is as follows: Arrange the loads at each level in chronological order; for the i-th level load, calculate the cumulative influence coefficient of all previous loads on it. Multiply by the operating condition fluctuation coefficient The actual damage contribution of this load level is obtained; the damage contributions of all loads are summed to obtain the total cumulative fatigue damage. 4) Model Coupling and Solution Process a. Coupling with other damage models: This model needs to be bidirectionally coupled with the high-temperature modified erosion wear model and the corrosion-fatigue interactive damage model. Wear-fatigue coupling: Erosion wear leads to a reduction in component cross-section and an increase in stress, through Update the load spectrum; Corrosion-fatigue coupling: Corrosion pits act as stress concentration sources, through... Correct stress amplitude; Total damage calculation: ; b. Solution process Initialization: Set the initial time t=0, the initial damage D=0, and the initial cross-sectional area A=A0; Data acquisition: Collect pressure, vibration, current, and temperature data within the current time step Δt; Load spectrum generation: Converting multi-source data into an equivalent stress load spectrum σ(t); Rainflow counting: Rainflow counting is performed on σ(t) to obtain the amplitude, mean, and number of cycles of each load level; SN curve correction: Calculate the corrected SN curve N(T, σ_a) based on the current temperature T; Damage calculation: a. Calculate the nominal damage n_i / N_i for each load level. b. Calculate the load sequence effect coefficient λ_ji c. Calculate the operating condition fluctuation coefficient ξ d. Calculate the fatigue damage increment ΔD_fatigue at the current time step; Total damage update: ; Coupled update: Update component cross-sectional area and stress concentration factor based on wear and corrosion damage; Termination judgment: If D_total ≥ D_cr, output the remaining lifetime; otherwise, t = t + Δt, return to the data acquisition step.

[0012] Preferably, the neural network loss function in S2 is designed as a physical constraint loss function: The loss function consists of three parts: in, The data fitting loss is calculated based on real sensor data. The physical constraint loss is calculated based on a five-field coupling degradation mechanism model, forcing the model output to conform to physical laws. To smooth out losses and ensure the continuity and stability of prediction results; The design operating condition identification network identifies normal, fluctuating, and sudden operating conditions in real time; and automatically adjusts according to the operating condition type. as well as Weights: Under stable operating conditions, increase the weight of data loss. To improve prediction accuracy; Increase the weight of physical loss under fluctuating or abrupt operating conditions. This ensures that the prediction results conform to physical laws.

[0013] Preferably, the method for implementing cross-domain fine-tuning of the model is as follows: 1) Large-scale pre-training in the source domain a) Source domain dataset construction, including the following three categories: General industrial pump dataset, coal chemical general pump dataset, and similar type of pump dataset; data standardization processing is performed on each type of data. b. Basic Model Architecture The APINN (Adaptive Physical Information Neural Network) architecture, which is completely consistent with the target model, is used as the base model to ensure the effectiveness of the transfer. c. Pre-training task design A two-stage pre-training strategy of "self-supervised pre-training + supervised pre-training" is adopted: The self-supervised pre-training utilizes a large amount of unlabeled data to accomplish the following tasks: Task 1: Time series prediction task, predict sensor data for the next 10 time steps; Task 2: Mask Reconstruction Task. Reconstruct the masked data by randomly masking the data for 15% of the time steps. Task 3: Comparative learning tasks enhance similar samples from the same device and differentiate samples from different devices; The supervised pre-training utilizes labeled data: Task: Residual lifetime regression prediction; Loss function: MSE loss combined with physical constraint loss; Training rounds: 50 rounds, learning rate 1e-3; 2) Unsupervised domain adaptive implementation The data distribution differences between the source domain and the target domain mainly stem from: Differences in operating conditions: Coal-to-oil machine pumps have higher temperatures, pressures, and solid content; Equipment differences: Differences in structure and performance between pumps from different manufacturers and models; Sensor differences: Differences in accuracy and noise characteristics between different sensors; Therefore, a hybrid approach combining Adversarial Domain Adaptive (DANN) and Maximum Mean Difference (MMD) is adopted to simultaneously align marginal and conditional distributions. a. Domain-Adaptive Network Architecture Design Based on the basic APINN model, a neighborhood discriminator and an MMD loss calculation module are added, including a data input layer; an APINN feature extraction layer; branch 1: the RUL prediction head obtains the RUL prediction loss; branch 2: the neighborhood discriminator obtains the adversarial loss; branch 3: the MMD loss calculation module obtains the distribution alignment loss. b. Loss Function Design The total loss function consists of three parts: Among them, mission losses RUL prediction MSE loss calculated using only labeled data from the source domain; Combat losses Domain adversarial training is achieved using a gradient inversion layer (GRL). in For the domain discriminator, , These represent the number of samples in the source domain and the target domain, respectively. MMD loss : Calculate the distance between the features of the source and target domains in the reproducing kernel Hilbert space: in, The kernel mapping function uses a Gaussian kernel; c. Training process Freeze the APINN feature extraction layer and RUL prediction head, and train only the domain discriminator for 5 rounds; Freeze the neighborhood discriminator, train the APINN feature extraction layer and RUL prediction head, and iterate for 1 round; The two steps above are trained alternately for a total of 20 iterations; Learning rate setting: Feature extraction layer 1e -5 Domain discriminator 1e -4 ; 3) Parameter Efficient Fine-Tuning (PEFT) Implementation Based on the characteristics of the APINN model, LoRA (Low-Rank Adaptation) was chosen as the core PEFT technique for the following reasons: extremely few parameters: only 0.1% to 1% of the parameters of the original model need to be trained; no change to the original model structure: no additional overhead during inference; good adaptability: can be applied to convolutional layers, fully connected layers, and attention layers; preserves pre-trained knowledge: will not destroy the general features learned from the source domain. a. In the original weight matrix W0∈R d×k Add two low-rank matrices A∈R next to it r×k and B∈R d×r During training, only A and B are updated; during inference, BA is added to W0. W = W0 + BA; The insertion position of LoRA in the APINN model: All convolutional layers of a multi-scale TCN layer; A fully connected layer with physical constraint layers; Fully connected layer of the RUL prediction head; Key parameter settings in the model: rank r: set to 8 or 16 to balance parameter quantity and performance; scaling factor α: set to r to ensure that the contribution of LoRA is 0 during initialization; Dropout: set to 0.1 to prevent overfitting; The fine-tuning training process is as follows: a. Dataset partitioning: Training set: 3 months of normal operation data of the target coal-to-oil machine pump plus 1-2 historical fault data; Validation set: runtime data from the most recent month; Test set: Data from one reserved fault test; b. Training process Completely freeze all parameters of the pre-trained APINN model; Insert a LoRA matrix into the specified layer, and initialize A to a Gaussian distribution and B to 0; Only the LoRA matrix is ​​trained, while other parameters remain unchanged; Loss function: MSE loss combined with physical constraint loss; Training rounds: 20-30 rounds, early stop strategy; Learning rate: 1e -4 ; After training, the LoRA matrix BA is merged into the original weight matrix W0, resulting in a fine-tuned model with the same structure as the original model. There is no additional overhead during inference, maintaining the real-time performance of the original model.

[0014] Preferably, the multi-scale digital twin adopts a three-layer nested architecture of micro-meso-macro, with bidirectional information transmission. The information transmission mechanism between scales is as follows: Macroscopic operating parameters serve as mesoscopic boundary conditions; mesoscopic stress and temperature distributions serve as microscopic load conditions. Microscale material damage parameters update mesoscale material properties; mesoscale component wear updates macroscale overall machine performance parameters; The macroscopic scale serves as a digital twin of overall system performance degradation, used for real-time state synchronization between physical devices and virtual models, providing global visualization and rapid lifespan prediction. The overall performance model is to establish a characteristic curve model of the pump and calculate the pump's efficiency, head, shaft power and other performance indicators based on the current operating parameters. The performance degradation model is based on the component wear amount output at the mesoscale, which is used to correct the characteristic curve of the pump and simulate the performance degradation process over time. The mesoscale data serves as a multi-physics coupled digital twin of key components, used for five-field coupled calculations involving flow field, temperature field, stress field, wear field, and corrosion field, to accurately predict the damage distribution of key components. The multiphysics coupled solver design adopts a hybrid coupling strategy of "weak coupling as the main method and strong coupling in key regions": Flow field solution: Based on large eddy simulation (LES), the solid-liquid two-phase flow field of oil-coal slurry is solved to obtain the particle velocity, concentration, and impact angle distribution; Temperature field solution: Solving for the temperature distribution of the flow field and the temperature distribution of the solid component based on the energy equation; Stress field solution: The stress and strain distribution of the component is solved based on the finite element method, taking into account temperature stress and fluid dynamic pressure; Wear field solution: Use the high-temperature modified erosion wear model to calculate the wear rate and cumulative wear amount of each grid cell; Corrosion field solution: Call the corrosion-fatigue interactive damage model to calculate the corrosion rate and corrosion damage of each grid cell; To meet real-time requirements, the high-fidelity finite element model is reduced in order: the orthogonal decomposition (POD) method is used to project the high-dimensional physical field data into a low-dimensional subspace; and a mapping relationship between the reduced-order model and the operating parameters is established. As wear progresses, the geometry of the component changes. The computational mesh is dynamically updated using the Arbitrary Lagrange-Euler (ALE) method. Each update cycle adjusts the mesh node positions based on the amount of wear. When the mesh deformation exceeds the threshold, the mesh is automatically re-divided; Achieve bidirectional coupling between damage evolution and geometric changes; The microscale, serving as a digital twin of material damage evolution, is used to simulate the microscopic damage evolution process within the material, providing material property parameters for the mesoscale; it includes the following models: Material microstructure modeling: Based on scanning electron microscope (SEM) images, a polycrystalline microstructure model of the material is constructed, including features such as grains, grain boundaries, and second phases; Crystal plasticity model: Establish a crystal plasticity constitutive model that considers temperature effects to simulate the plastic deformation and dislocation evolution of materials under cyclic loading; Electrochemical corrosion model: Establish a corrosion pit initiation and propagation model to simulate the electrochemical corrosion process under the action of H2S / CO2 medium; Microcrack propagation model: Based on the extended finite element method (XFEM), the initiation, propagation and coalescence of microcracks are simulated; Offline pre-calculation combined with online interpolation: Due to the extremely large amount of computation at the microscale, offline pre-calculation is used to calculate the material response under different working conditions and damage levels to establish a database; during online runtime, material property parameters are quickly obtained through interpolation.

[0015] Preferably, the virtual-real synchronous evolution is implemented as follows: a) Real-time access and fusion of multi-source heterogeneous data The collected data includes multi-source heterogeneous data such as process data, status data, offline data, and environmental data; A spatiotemporal alignment algorithm is used to unify data with different sampling frequencies and timestamps to the same time base; a federated filtering algorithm is used to fuse multi-sensor data to improve data reliability and accuracy; and a data quality assessment model is established to detect sensor drift, missing data, and outliers in real time. b. Asynchronous updating and dynamic calibration of multi-scale models At a macro scale, it achieves second-level status synchronization and updates the virtual model's operating parameters (pressure, temperature, flow rate, speed) and overall performance indicators (efficiency, head, shaft power) in real time; it adopts a proportional-integral-derivative (PID) controller to adjust the input parameters of the virtual model in real time based on sensor measurements, ensuring that the output of the virtual model is consistent with the physical equipment; At the mesoscale, damage synchronization is achieved within minutes, updating wear distribution, stress distribution, temperature distribution, and corrosion damage of key components; the extended Kalman filter (EKF) algorithm is used in conjunction with sensor data for online calibration of the multiphysics model; based on the arbitrary Lagrange-Euler (ALE) method, the mesh node positions of the virtual model are adjusted in real time according to the calculated wear amount, accurately simulating the geometric changes of the flow components caused by wear; when the mesh deformation exceeds 20%, mesh reconstruction is automatically triggered to ensure calculation accuracy; At the microscale, the system synchronizes the properties of materials at the time of application, updating mechanical property parameters such as hardness, yield strength, and fatigue limit, as well as micro-damage parameters such as pit depth and microcrack density. The system employs Bayesian parameter estimation methods, combining offline material test data and online operational data, to update the parameters of the micro-damage model. c. Virtual-real bidirectional mapping and feedback control Forward mapping: Changes in the state of physical devices are transmitted to the digital twin in real time, driving the evolution of the virtual model; Reverse mapping: The simulation results and prediction information of the digital twin are fed back to the physical system to guide production operation and maintenance decisions; When an abnormal operating condition is detected, a warning signal is sent to the control system, suggesting adjustments to the operating parameters; After maintenance is completed, the damage status of the digital twin is updated according to the maintenance record, and the lifespan of the corresponding components is reset.

[0016] Preferably, the edge-cloud collaborative computing architecture model includes: Application layer: As an intelligent operation and maintenance visualization platform, it specifically implements equipment status display, lifespan prediction display, maintenance work order management, and report statistics; Cloud layer: As the global computing and decision-making center, it specifically implements data management, model training and version management, high-precision life assessment, multi-scale digital twin engine, and generates intelligent maintenance decisions; Network layer: as a communication system for the Industrial Internet of Things; Edge layer: As a real-time computing and control unit on site, it specifically implements multi-source data acquisition, real-time data preprocessing, real-time operating condition identification, lightweight APINN inference module, local hierarchical early warning, and edge-cloud data interaction; Perception layer: Generates a multi-mode sensor network; The edge terminal achieves real-time prediction of remaining service life, as follows: 1) Input data: Includes eight types of core sensor data, namely pump outlet pressure, inlet pressure, flow rate, motor current, speed, pump body temperature, bearing horizontal vibration, and bearing vertical vibration; 2) Real-time operating condition identification: Call the lightweight TCN operating condition identification network to output the current operating condition probability distribution; determine the operating condition type based on the maximum probability: normal p>0.9, fluctuation 0.7≤p≤0.9, sudden change p<0.7; 3) Lightweight APINN model inference Model structure: The multi-scale TCN feature extraction layer includes 3 parallel branches with convolutional kernel sizes of 3 / 5 / 7; a physical constraint layer; and 2 fully connected prediction heads. Model compression: After pruning, INT8 quantization, and knowledge distillation, the number of parameters was compressed from 12MB to 860KB, and the inference speed was improved by 6 times; Inference output: Remaining life prediction (RUL_pred): in days; prediction confidence: a value between 0 and 1; risk level: normal (RUL>30 days, confidence>0.7), warning (7 days≤RUL≤30 days), severe (RUL<7 days); 4) Local early warning and data upload When the risk level is "severe", a local audible and visual alarm is immediately triggered, and the PLC is linked to reduce the operating load of the pump. Data upload strategy: Under normal operating conditions: average data and prediction results are uploaded every 10 minutes; Fluctuating operating conditions: Data is uploaded every 1 minute; Abnormal operating conditions: Upload raw data and early warning information in real time; The cloud-based hourly high-precision lifespan calibration method is as follows: 1) Data reception and fusion Receive real-time data and prediction results uploaded from the edge device; Integrate offline data (equipment design parameters, material properties, historical maintenance records, failure analysis reports); Conduct data quality assessments, remove outliers, and complete missing data. 2) High-precision life assessment High-precision lifetime prediction is achieved by calling the full APINN model. Input features include: Sensor data uploaded from the edge; Multiphysics characteristics (stress distribution, temperature distribution, wear distribution) output by digital twin; Medium characteristic parameters (H2S / CO2 partial pressure, solids content of oil-coal slurry, particle size distribution); The Monte Carlo simulation method is used to quantify data uncertainty, model uncertainty, and operating condition uncertainty, and output the 95% confidence interval of the remaining lifetime. Generate a cloud map of the overall machine's lifespan distribution to accurately locate the most vulnerable locations and failure modes; 3) Model Updates and Pushes The APINN model is fine-tuned weekly using newly collected operational data and virtual data generated by the digital twin. Incremental updates are used to transmit only the parameters that change in the model (usually <100KB), reducing bandwidth consumption. A canary release strategy is adopted, first pushing to 10% of edge nodes for verification, and then pushing to the full population after there are no problems. The edge device employs a dual-model hot-switching mechanism, ensuring that the update process does not affect normal operation.

[0017] The beneficial effects of this invention are: By constructing a five-field coupled degradation mechanism model of coal-to-oil pump with high temperature, high pressure, erosion, corrosion and fatigue, the damage evolution process of multiple factors interacting under complex working conditions is accurately characterized, and the core problem of the distortion of failure mechanism description in existing general models is effectively solved. By embedding a dedicated physical model as a constraint into a neural network and combining it with high-fidelity virtual fault data generated by multi-scale digital twins, the dependence on real fault data is greatly reduced, and a predictive model with good generalization ability can still be trained even under conditions where fault data is extremely scarce. Through real-time operating condition identification and dynamic weight adjustment mechanisms, the model can adapt to different operating states. Even when the flow, pressure, and temperature fluctuate frequently or even change abruptly during the production process, it can still maintain stable predictive performance, avoiding the defect of a sharp decline in accuracy of traditional models. A multi-scale, multi-physics digital twin was constructed, ranging from microscopic material damage to macroscopic overall system performance. Through asynchronous updates and online calibration mechanisms, the entire lifecycle of physical devices and virtual models was synchronized, transforming the digital twin from a static display tool into a dynamic computing engine that supports high-precision life prediction. The edge-cloud collaborative computing architecture enables rapid real-time prediction at the edge to ensure on-site safety, while the cloud performs high-precision global assessment and model updates. At the same time, the constraints of the physical model give the prediction results clear physical meaning, making them easy for operation and maintenance personnel to understand and apply. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0021] A method for predicting the remaining life of a coal-to-oil pump includes: S1: A five-field coupled degradation mechanism model was constructed, and a damage evolution equation capable of quantitatively describing the interaction of five factors—high temperature, high pressure, erosion, corrosion, and fatigue—was established, including: The high-temperature modified erosion wear model is based on the traditional Archard wear model and is modified by introducing the material thermal softening coefficient and plastic strengthening coefficient to reflect the dynamic changes of material hardness and yield strength at high temperature. The high-temperature modified erosion wear model, based on the Archard wear model, considers the shape factor and concentration distribution of oil-coal slurry particles, and corrects the influence of particle impact angle and velocity on the wear rate. The corrected model yields the following formula, which is used to calculate the erosion wear rate W (mm / a) of the flow components of the coal-to-oil pump, such as impeller blades, pump casing inner wall, sealing ring, plunger, and valve seat, at any spatial location under specific operating conditions: Where K is the overall wear coefficient; Temperature-dependent material hardness; It is a function of the impact angle; Let be the particle concentration distribution function; This is the reference hardness at room temperature. The normal impact force of the particles is the normal component of a single oil-coal slurry particle when it vertically impacts the material surface. The unit is N, and it is obtained by CFD solid-liquid two-phase flow simulation calculation. The relative impact velocity of the particles; It represents the nominal hardness of the material, which is the same as the value of H0, and is used for dimensionless processing in the formula; A corrosion-fatigue interactive damage model was developed, and an electrochemical corrosion-mechanical fatigue coupled damage evolution equation was established to distinguish between two mechanisms: corrosion-accelerated fatigue and fatigue-accelerated corrosion. Mechanism A: Corrosion accelerates fatigue, contributing about 70% of the total damage as the dominant mechanism: corrosion pits generated by electrochemical corrosion act as stress concentration sources, accelerating crack initiation, while H2S / CO2 media enter the crack tip, accelerating crack propagation. Mechanism B: Fatigue accelerates corrosion as a feedback mechanism: Cyclic loading causes the corrosion product film to rupture periodically, exposing fresh metal surfaces and significantly increasing the local corrosion rate. The final output of the model is the total corrosion-fatigue coupled damage D_total at any time. When D_total reaches the critical damage threshold D_cr, which is usually taken as 0.8~0.9, the component is judged to have suffered corrosion fatigue failure. 1) The process of establishing the basic model is as follows: a. First, establish a pure corrosion rate model under stress-free conditions. in, The pure corrosion current density (A / m²) For exchange current density; The activation energy of the reaction is expressed in J / mol. and They are respectively as well as The partial pressure (MPa); a and b are the reaction order, which are usually taken as a = 0.5~0.8 and b = 0.3~0.6 under coal-to-oil conditions; Pure corrosion rate: in, The molar mass of the metal; The number of electrons in the reaction; It is Faraday's constant; The density of the material; b. Pure fatigue crack propagation model The fatigue crack propagation rate under corrosion-free conditions is described using the Paris equation: in, The crack propagation rate per cycle (m / cycle); Stress intensity factor amplitude (MPa·m) 0.5 ); and These are material constants, obtained through room temperature fatigue tests; 2) Realization of corrosion-accelerated fatigue mechanism a. Corrosion pit-induced crack initiation model Corrosion pits generate significant stress concentrations. The stress concentration factor K_t is related to the pit depth a and radius r as follows: Introducing an equivalent initial crack size a_eq for corrosion pits: The crack initiation life induced by corrosion pits was calculated using Miner's damage theory: ,in, The fatigue life corresponding to the initial crack size a_eq; b. Corrosive Media Accelerated Crack Propagation Model The corrosion fatigue factor (CFF) is introduced to quantitatively describe the accelerating effect of H2S / CO2 on crack propagation rate: Where A is the material-medium coupling coefficient, which is determined by high temperature and high pressure corrosion fatigue test; For reference corrosion current density; It is a correction function for stress intensity factor, temperature, and pressure; For typical coal-to-oil operating conditions, the specific form of the correction function is as follows: in, This is the corrosion fatigue threshold value; It is the activation energy for corrosion fatigue; , This is an experimental constant; Corrected corrosion fatigue crack propagation rate: ; 3) Realization of fatigue-accelerated corrosion mechanism a. Dynamic growth model of corrosion product film Under high temperature and pressure, a corrosion product film composed of FeS, FeCO3, etc., will form on the metal surface. The growth of the film follows parabolic dynamics. in, Let t be the film thickness (m) at time t; The steady-state film thickness; The membrane growth time constant; The protective effect of a membrane is represented by the membrane protection coefficient η: ,in The film density coefficient; Corrosion rate considering the protective effect of the film: ; b. Criteria and Repair Model for Membrane Rupture under Cyclic Loading When the strain generated by cyclic stress exceeds the fracture strain of the corrosion product film... e f At that time, the membrane ruptured: e max > e f After the film ruptures, the fresh metal surface is exposed, and the corrosion rate instantly recovers to near the pure corrosion rate v. corr,0 Then the film begins to regrow, and the corrosion rate gradually decreases. Introducing membrane rupture area fraction ϕ Describe the degree of damage to the membrane under cyclic loading: in, This represents the initial fracture area fraction. , This is an experimental constant; This represents the number of loop iterations. Corrected actual corrosion rate: ; Two-way coupled solution process: The corrosion-fatigue interaction is a dynamic, two-way coupled process, which is solved using a time-stepping iterative method. 1) Initialization: Set the initial time t=0, initial damage D=0, and initial film thickness h=0; 2) Corrosion Calculation: Calculate the current corrosion rate based on the current film thickness and the fraction of ruptured area. v corr ; 3) Corrosion damage update: Calculate the corrosion damage increment ΔD_corr within time step Δt, and update the corrosion pit size; 4) Stress calculation: Based on the current corrosion pit size, calculate the stress concentration factor and stress intensity factor amplitude ΔK; 5) Fatigue calculation: Calculate the crack propagation rate da / dN based on the current ΔK and corrosion fatigue damage factor CFF; 6) Fatigue damage update: Calculate the fatigue damage increment ΔD_fatigue within time step Δt; 7) Total damage update: D` total =D total +ΔD corr +ΔD fatigue ; 8) Membrane condition update: Update the membrane rupture area fraction and membrane thickness based on the current cyclic stress; 9) Determine the termination condition: If D total ≥D cr Output the remaining lifetime; otherwise, t = t + Δt, return to step 2. The dynamic load fatigue cumulative damage model improves upon the traditional Miner linear cumulative damage theory by introducing a load sequence effect coefficient and a working condition fluctuation coefficient. The dynamic load spectrum is processed using the rainflow counting method, and combined with the temperature correction of the material SN curve, the fatigue damage accumulation rate under different working conditions is calculated. The final output of the dynamic load fatigue cumulative damage model is the fatigue cumulative damage D_fatigue at any time. When D_fatigue reaches the critical damage threshold D_cr, the threshold is usually taken as 0.7~0.8. Considering the coupling effect of corrosion and wear, the component is determined to have fatigue failure. The construction and implementation process of the dynamic load fatigue cumulative damage model is as follows: 1) Implementation of dynamic load spectrum preprocessing and rainflow counting a. Load spectrum acquisition and preprocessing: The fatigue load of the coal-to-oil pump mainly comes from: oil-coal slurry pressure pulsation, with a main frequency of 10~50Hz; periodic load generated by impeller rotation, with a main frequency equal to the rotation speed; and low-frequency load fluctuations caused by changes in operating conditions, 0.01~1Hz. Raw data were collected from pump outlet pressure sensor, bearing vibration sensor, and motor current sensor, with a sampling frequency ≥1kHz; high-frequency noise and electromagnetic interference were removed using wavelet threshold denoising; and the multi-source data were fused to convert it into an equivalent stress load spectrum. in, , , The stress conversion factor is calibrated through finite element analysis; b. Improved rainflow counting: The four-point rainflow counting method is used to process the non-stationary dynamic load spectrum and extract the amplitude, mean and cycle number of each load level; Improved for coal-to-oil: The working condition marking function is added to associate each load cycle with the corresponding temperature, pressure and flow rate working condition parameters to provide a basis for subsequent working condition fluctuation correction. 2) Temperature-corrected material SN curve model a. Basic SN curve The fatigue life of materials at room temperature follows a power function relationship: in, and m are material constants, determined by standard fatigue tests; b. High-temperature SN curve The operating temperature (350~450℃) of the coal-to-oil pump leads to a significant decrease in the fatigue strength of the material and generates thermal fatigue effects. Therefore, a temperature correction factor K_T is introduced. in, The fatigue limit at temperature T; This represents the fatigue limit at room temperature. The melting point of the material; The temperature sensitivity coefficient is determined through high-temperature fatigue testing. For commonly used wear-resistant alloy steels, Take a value of 0.3~0.5; Corrected high-temperature SN curve: c. Mean stress correction The effect of mean stress on fatigue life is corrected using the Goodman formula: in, The tensile strength of the material at temperature T; 3) Improved fatigue cumulative damage model a. Introduction of the load sequence effect coefficient Traditional Miner's Theory: in, Let be the number of cycles for the i-th level load; This represents the fatigue life under this load level. High loads generate residual compressive stress within the material, slowing down the accumulation of damage under subsequent low loads; while low loads generate microcracks, accelerating the propagation of damage under subsequent high loads. Introducing the load sequence effect coefficient λ ij Describe the effect of the i-th level load on the damage caused by the j-th level load: in, , The load sequence sensitivity coefficient was determined by variable amplitude fatigue test, with β=0.2~0.4 and γ=0.3~0.5. This represents the maximum stress in the load spectrum; b. Introducing a working condition fluctuation coefficient: During coal-to-oil production, frequent switching of working conditions can cause changes in the statistical characteristics of the load spectrum, accelerating fatigue damage; introducing a working condition fluctuation coefficient... : in, The standard deviation of the load amplitude; is the average load amplitude; f (times / hour) is the operating condition switching frequency; The sensitivity coefficient to operating condition fluctuations is calibrated using field data. =0.1~0.3; c. Improved cumulative damage calculation, taking into account both load sequence effect and working condition fluctuation effect, the improved fatigue cumulative damage formula is as follows: The calculation logic is as follows: Arrange the loads at each level in chronological order; for the i-th level load, calculate the cumulative influence coefficient of all previous loads on it. Multiply by the operating condition fluctuation coefficient The actual damage contribution of this load level is obtained; the damage contributions of all loads are summed to obtain the total cumulative fatigue damage. 4) Model Coupling and Solution Process a. Coupling with other damage models: This model needs to be bidirectionally coupled with the high-temperature modified erosion wear model and the corrosion-fatigue interactive damage model. Wear-fatigue coupling: Erosion wear leads to a reduction in component cross-section and an increase in stress, through Update the load spectrum; Corrosion-fatigue coupling: Corrosion pits act as stress concentration sources, through... Correct stress amplitude; Total damage calculation: ; b. Solution process Initialization: Set the initial time t=0, the initial damage D=0, and the initial cross-sectional area A=A0; Data acquisition: Collect pressure, vibration, current, and temperature data within the current time step Δt; Load spectrum generation: Converting multi-source data into an equivalent stress load spectrum σ(t); Rainflow counting: Rainflow counting is performed on σ(t) to obtain the amplitude, mean, and number of cycles of each load level; SN curve correction: Calculate the corrected SN curve N(T, σ_a) based on the current temperature T; Damage calculation: a. Calculate the nominal damage n_i / N_i for each load level. b. Calculate the load sequence effect coefficient λ_ij c. Calculate the operating condition fluctuation coefficient ξ d. Calculate the fatigue damage increment ΔD_fatigue at the current time step; Total damage update: ; Coupled update: Update component cross-sectional area and stress concentration factor based on wear and corrosion damage; Termination judgment: If D_total ≥ D_cr, output the remaining lifetime; otherwise, t = t + Δt, return to the data acquisition step; S2: Construct a physical information neural network prediction model, and embed the five-field coupling model of coal-to-oil pump constructed in S1 into the neural network loss function to improve the generalization ability of the model under small sample conditions. The neural network loss function is designed as a physically constrained loss function, consisting of three parts: in, The data fitting loss is calculated based on real sensor data. The physical constraint loss is calculated based on a five-field coupling degradation mechanism model, forcing the model output to conform to physical laws. To smooth out losses and ensure the continuity and stability of prediction results; The design operating condition identification network identifies normal, fluctuating, and sudden operating conditions in real time; and automatically adjusts according to the operating condition type. as well as Weights: Under stable operating conditions, increase the weight of data loss. To improve prediction accuracy; Increase the weight of physical loss under fluctuating or abrupt operating conditions. This ensures that the prediction results conform to physical laws; The hierarchical structure of the working condition identification network is as follows: Input layer: As a multi-sensor timing window, it adopts a sliding window mechanism with a window size of window_size=64 (corresponding to 6.4s of data, sampling frequency of 10Hz) and a stride=1. Data preprocessing layer: Normalizes, denoises, and fills in missing values ​​for the sensor-acquired data; Multi-scale TCN feature extraction layer: Three parallel TCN branches are used to capture the operating conditions at different time scales. Branch 1 captures the rapid change features of abrupt operating conditions; Branch 2 captures the medium-speed change features of fluctuating operating conditions; and Branch 3 captures the slow change features of normal operating conditions. The outputs of the three branches are concatenated along the channel dimension to obtain the fused feature map; Attention mechanism layer: includes channel attention module, used to adaptively learn the importance weights of different sensors for identifying working conditions; The temporal attention module adaptively learns the importance weights of different time steps for identifying the current working condition, focusing on the most recent time step and the point of change. The classification head and output layer include: global average pooling: compressing temporal features into a fixed-length feature vector; and outputting the probability distribution and scores of the three working conditions through fully connected layer 1 and fully connected layer 2. The basic model is pre-trained using a large amount of data from general pumps and other coal chemical equipment; the parameter efficient fine-tuning (PEFT) technique is used to fine-tune the model using only a small amount of normal operation data and historical fault data of the target coal-to-oil machine pump; and a domain adaptive algorithm is introduced to reduce the data distribution differences between the source domain and the target domain. The implementation method for cross-domain fine-tuning of the model is as follows: 1) Large-scale pre-training in the source domain a) Source domain dataset construction, including the following three categories: General industrial pump datasets: publicly available datasets such as NASA C-MAPSS, IEEE PHM 2012, and XJTU-SY bearing dataset; Coal Chemical Industry General Pump Dataset: Publicly available operating data of centrifugal pumps and reciprocating pumps from coal chemical enterprises; Data sets of similar pumps: full life cycle data of high-temperature and high-pressure pumps and slurry pumps in the petrochemical industry; Perform data standardization processing on each type of data separately; b. Basic Model Architecture The APINN (Adaptive Physical Information Neural Network) architecture, which is completely consistent with the target model, is used as the base model to ensure the effectiveness of the transfer. c. Pre-training task design A two-stage pre-training strategy of "self-supervised pre-training + supervised pre-training" is adopted: Self-supervised pre-training utilizes large amounts of unlabeled data to accomplish the following tasks: Task 1: Time series prediction task, predict sensor data for the next 10 time steps; Task 2: Mask Reconstruction Task. Reconstruct the masked data by randomly masking the data for 15% of the time steps. Task 3: Comparative learning tasks enhance similar samples from the same device and differentiate samples from different devices; Supervised pre-training utilizes labeled data: Task: Residual lifetime regression prediction; Loss function: MSE loss combined with physical constraint loss; Training rounds: 50 rounds, learning rate 1e-3; 2) Unsupervised domain adaptive implementation The differences in data distribution between the source and target domains mainly stem from: Differences in operating conditions: Coal-to-oil machine pumps have higher temperatures, pressures, and solid content; Equipment differences: Differences in structure and performance between pumps from different manufacturers and models; Sensor differences: Differences in accuracy and noise characteristics between different sensors; Therefore, a hybrid approach combining Adversarial Domain Adaptive (DANN) and Maximum Mean Difference (MMD) is adopted to simultaneously align marginal and conditional distributions. a. Domain-Adaptive Network Architecture Design Based on the basic APINN model, a neighborhood discriminator and an MMD loss calculation module are added, including a data input layer; an APINN feature extraction layer; branch 1: the RUL prediction head obtains the RUL prediction loss; branch 2: the neighborhood discriminator obtains the adversarial loss; branch 3: the MMD loss calculation module obtains the distribution alignment loss. b. Loss Function Design The total loss function consists of three parts: Among them, mission losses RUL prediction MSE loss calculated using only labeled data from the source domain; Combat losses Domain adversarial training is achieved using a gradient inversion layer (GRL). in For the domain discriminator, , These represent the number of samples in the source domain and the target domain, respectively. MMD loss : Calculate the distance between the features of the source and target domains in the reproducing kernel Hilbert space: in, The kernel mapping function uses a Gaussian kernel; c. Training process Freeze the APINN feature extraction layer and RUL prediction head, and train only the domain discriminator for 5 rounds; Freeze the neighborhood discriminator, train the APINN feature extraction layer and RUL prediction head, and iterate for 1 round; The two steps above are trained alternately for a total of 20 iterations; Learning rate setting: Feature extraction layer 1e -5 Domain discriminator 1e -4 ; 3) Parameter Efficient Fine-Tuning (PEFT) Implementation Based on the characteristics of the APINN model, LoRA (Low-Rank Adaptation) was chosen as the core PEFT technique for the following reasons: extremely few parameters: only 0.1% to 1% of the parameters of the original model need to be trained; no change to the original model structure: no additional overhead during inference; good adaptability: can be applied to convolutional layers, fully connected layers, and attention layers; preserves pre-trained knowledge: will not destroy the general features learned from the source domain. a. In the original weight matrix W0∈R d×k Add two low-rank matrices A∈R next to it r×k and B∈R d×r During training, only A and B are updated; during inference, BA is added to W0. W = W0 + BA; The insertion position of LoRA in the APINN model: All convolutional layers of a multi-scale TCN layer; A fully connected layer with physical constraint layers; Fully connected layer of the RUL prediction head; Key parameter settings in the model: rank r: set to 8 or 16 to balance parameter quantity and performance; scaling factor α: set to r to ensure that the contribution of LoRA is 0 during initialization; Dropout: set to 0.1 to prevent overfitting; The fine-tuning training process is as follows: a. Dataset partitioning: Training set: 3 months of normal operation data of the target coal-to-oil machine pump plus 1-2 historical fault data; Validation set: runtime data from the most recent month; Test set: Data from one reserved fault test; b. Training process Completely freeze all parameters of the pre-trained APINN model; Insert a LoRA matrix into the specified layer, and initialize A to a Gaussian distribution and B to 0; Only the LoRA matrix is ​​trained, while other parameters remain unchanged; Loss function: MSE loss combined with physical constraint loss; Training rounds: 20-30 rounds, early stop strategy; Learning rate: 1e -4 ; After training, the LoRA matrix BA is merged into the original weight matrix W0, resulting in a fine-tuned model with the same structure as the original model. There is no additional overhead during inference, maintaining the real-time performance of the original model. S3: Construct a multi-scale, multi-physics field real-time coupled digital twin, building a multi-scale digital twin from microscopic material damage to macroscopic overall machine performance, used for virtual-real synchronous evolution to generate virtual fault data and calibrate online prediction models; and provide optimized maintenance decisions; The multi-scale digital twin adopts a three-layer nested architecture of micro-meso-macro, with bidirectional information transmission. The information transmission mechanism between scales is as follows: Macroscopic operating parameters serve as mesoscopic boundary conditions; mesoscopic stress and temperature distributions serve as microscopic load conditions. Microscale material damage parameters update mesoscale material properties; mesoscale component wear updates macroscale overall machine performance parameters; The macroscopic scale serves as a digital twin of overall system performance degradation, used for real-time state synchronization between physical devices and virtual models, providing global visualization and rapid lifespan prediction; Among them, the overall performance model is to establish the characteristic curve model of the pump and calculate the pump's efficiency, head, shaft power and other performance indicators based on the current operating parameters. The performance degradation model is based on the component wear amount output at the mesoscale, which corrects the characteristic curve of the pump and simulates the performance degradation process over time. At the mesoscale, a multi-physics coupled digital twin of key components is used for five-field coupled calculations of flow field, temperature field, stress field, wear field, and corrosion field to accurately predict the damage distribution of key components. The multiphysics coupled solver design adopts a hybrid coupling strategy of "weak coupling as the main method and strong coupling in key regions": Flow field solution: Based on large eddy simulation (LES), the solid-liquid two-phase flow field of oil-coal slurry is solved to obtain the particle velocity, concentration, and impact angle distribution; Temperature field solution: Solving for the temperature distribution of the flow field and the temperature distribution of the solid component based on the energy equation; Stress field solution: The stress and strain distribution of the component is solved based on the finite element method, taking into account temperature stress and fluid dynamic pressure; Wear field solution: Use the high-temperature modified erosion wear model to calculate the wear rate and cumulative wear amount of each grid cell; Corrosion field solution: Call the corrosion-fatigue interactive damage model to calculate the corrosion rate and corrosion damage of each grid cell; To meet real-time requirements, the high-fidelity finite element model is reduced in order: the orthogonal decomposition (POD) method is used to project the high-dimensional physical field data into a low-dimensional subspace; and a mapping relationship between the reduced-order model and the operating parameters is established. As wear progresses, the geometry of the component changes. The computational mesh is dynamically updated using the Arbitrary Lagrange-Euler (ALE) method. Each update cycle adjusts the mesh node positions based on the amount of wear. When the mesh deformation exceeds the threshold, the mesh is automatically re-divided; Achieve bidirectional coupling between damage evolution and geometric changes; The microscale serves as a digital twin of material damage evolution, used to simulate the microscopic damage evolution process within materials and to provide material property parameters for the mesoscale; it includes the following models: Material microstructure modeling: Based on scanning electron microscope (SEM) images, a polycrystalline microstructure model of the material is constructed, including features such as grains, grain boundaries, and second phases; Crystal plasticity model: Establish a crystal plasticity constitutive model that considers temperature effects to simulate the plastic deformation and dislocation evolution of materials under cyclic loading; Electrochemical corrosion model: Establish a corrosion pit initiation and propagation model to simulate the electrochemical corrosion process under the action of H2S / CO2 medium; Microcrack propagation model: Based on the extended finite element method (XFEM), the initiation, propagation and coalescence of microcracks are simulated; Offline pre-calculation combined with online interpolation: Due to the extremely large computational load at the microscale, offline pre-calculation of material response under different working conditions and damage levels is used to establish a database; during online runtime, material property parameters are quickly obtained through interpolation. The implementation process of virtual-real synchronous evolution based on digital twins is as follows: a) Real-time access and fusion of multi-source heterogeneous data The collected data includes multi-source heterogeneous data such as process data, status data, offline data, and environmental data; A spatiotemporal alignment algorithm is used to unify data with different sampling frequencies and timestamps to the same time base; a federated filtering algorithm is used to fuse multi-sensor data to improve data reliability and accuracy; and a data quality assessment model is established to detect sensor drift, missing data, and outliers in real time. b. Asynchronous updating and dynamic calibration of multi-scale models At a macro scale, it achieves second-level status synchronization and updates the virtual model's operating parameters (pressure, temperature, flow rate, speed) and overall performance indicators (efficiency, head, shaft power) in real time; it adopts a proportional-integral-derivative (PID) controller to adjust the input parameters of the virtual model in real time based on sensor measurements, ensuring that the output of the virtual model is consistent with the physical equipment; At the mesoscale, damage synchronization is achieved within minutes, updating wear distribution, stress distribution, temperature distribution, and corrosion damage of key components; the extended Kalman filter (EKF) algorithm is used in conjunction with sensor data for online calibration of the multiphysics model; based on the arbitrary Lagrange-Euler (ALE) method, the mesh node positions of the virtual model are adjusted in real time according to the calculated wear amount, accurately simulating the geometric changes of the flow components caused by wear; when the mesh deformation exceeds 20%, mesh reconstruction is automatically triggered to ensure calculation accuracy; At the microscale, the system synchronizes the properties of materials at the time of application, updating mechanical property parameters such as hardness, yield strength, and fatigue limit, as well as micro-damage parameters such as pit depth and microcrack density. The system employs Bayesian parameter estimation methods, combining offline material test data and online operational data, to update the parameters of the micro-damage model. c. Virtual-real bidirectional mapping and feedback control Forward mapping: Changes in the state of physical devices are transmitted to the digital twin in real time, driving the evolution of the virtual model; Reverse mapping: The simulation results and prediction information of the digital twin are fed back to the physical system to guide production operation and maintenance decisions; When an abnormal operating condition is detected, a warning signal is sent to the control system, suggesting adjustments to the operating parameters; After maintenance is completed, the damage status of the digital twin is updated according to the maintenance record, and the lifespan of the corresponding components is reset. S4: Real-time lifespan prediction based on edge-cloud collaborative architecture, constructing an edge-cloud collaborative computing architecture model to predict the real-time remaining lifespan of coal-to-oil pumps; Edge-cloud collaborative computing architecture models include: Application layer: As an intelligent operation and maintenance visualization platform, it specifically implements equipment status display, lifespan prediction display, maintenance work order management, and report statistics; Cloud layer: As the global computing and decision-making center, it specifically implements data management, model training and version management, high-precision life assessment, multi-scale digital twin engine, and generates intelligent maintenance decisions; Network layer: as a communication system for the Industrial Internet of Things; Edge layer: As a real-time computing and control unit on site, it specifically implements multi-source data acquisition, real-time data preprocessing, real-time operating condition identification, lightweight APINN inference module, local hierarchical early warning, and edge-cloud data interaction; Perception layer: Generates a multi-mode sensor network; The edge terminal achieves real-time prediction of remaining service life, as follows: 1) Input data: Includes eight types of core sensor data, namely pump outlet pressure, inlet pressure, flow rate, motor current, speed, pump body temperature, bearing horizontal vibration, and bearing vertical vibration; 2) Real-time operating condition identification: Call the lightweight TCN operating condition identification network to output the current operating condition probability distribution; determine the operating condition type based on the maximum probability: normal p>0.9, fluctuation 0.7≤p≤0.9, sudden change p<0.7; 3) Lightweight APINN model inference Model structure: The multi-scale TCN feature extraction layer includes 3 parallel branches with convolutional kernel sizes of 3 / 5 / 7; a physical constraint layer; and 2 fully connected prediction heads. Model compression: After pruning, INT8 quantization, and knowledge distillation, the number of parameters was compressed from 12MB to 860KB, and the inference speed was improved by 6 times; Inference output: Remaining life prediction (RUL_pred): in days; prediction confidence: a value between 0 and 1; risk level: normal (RUL>30 days, confidence>0.7), warning (7 days≤RUL≤30 days), severe (RUL<7 days); 4) Local early warning and data upload When the risk level is "severe", a local audible and visual alarm is immediately triggered, and the PLC is linked to reduce the operating load of the pump. Data upload strategy: Under normal operating conditions: average data and prediction results are uploaded every 10 minutes; Fluctuating operating conditions: Data is uploaded every 1 minute; Abnormal operating conditions: Upload raw data and early warning information in real time; The cloud-based hourly high-precision lifespan calibration method is as follows: 1) Data reception and fusion Receive real-time data and prediction results uploaded from the edge device; Integrate offline data (equipment design parameters, material properties, historical maintenance records, failure analysis reports); Conduct data quality assessments, remove outliers, and complete missing data. 2) High-precision life assessment High-precision lifetime prediction is achieved by calling the full APINN model. Input features include: Sensor data uploaded from the edge; Multiphysics characteristics (stress distribution, temperature distribution, wear distribution) output by digital twin; Medium characteristic parameters (H2S / CO2 partial pressure, solids content of oil-coal slurry, particle size distribution); The Monte Carlo simulation method is used to quantify data uncertainty, model uncertainty, and operating condition uncertainty, and output the 95% confidence interval of the remaining lifetime. Generate a cloud map of the overall machine's lifespan distribution to accurately locate the most vulnerable locations and failure modes; 3) Model Updates and Pushes The APINN model is fine-tuned weekly using newly collected operational data and virtual data generated by the digital twin. Incremental updates are used to transmit only the parameters that change in the model (usually <100KB), reducing bandwidth consumption. A canary release strategy is adopted, first pushing to 10% of edge nodes for verification, and then pushing to the full population after there are no problems. The edge device employs a dual-model hot-switching mechanism, ensuring that the update process does not affect normal operation.

Claims

1. A method for predicting the remaining life of a coal-to-oil pump, characterized in that, include: S1: A five-field coupled degradation mechanism model was constructed, and a damage evolution equation capable of quantitatively describing the interaction of five factors—high temperature, high pressure, erosion, corrosion, and fatigue—was established, including: The high-temperature modified erosion wear model is based on the traditional Archard wear model and is modified by introducing the material thermal softening coefficient and plastic strengthening coefficient to reflect the dynamic changes of material hardness and yield strength at high temperature. A corrosion-fatigue interactive damage model was developed, and an electrochemical corrosion-mechanical fatigue coupled damage evolution equation was established to distinguish between two mechanisms: corrosion-accelerated fatigue and fatigue-accelerated corrosion. The dynamic load fatigue cumulative damage model improves upon the traditional Miner linear cumulative damage theory by introducing a load sequence effect coefficient and a working condition fluctuation coefficient. The dynamic load spectrum is processed using the rainflow counting method, and combined with the temperature correction of the material SN curve, the fatigue damage accumulation rate under different working conditions is calculated. S2: Construct a physical information neural network prediction model, and embed the five-field coupling model of coal-to-oil pump constructed in S1 into the neural network loss function to improve the generalization ability of the model under small sample conditions. S3: Construct a multi-scale, multi-physics field real-time coupled digital twin, building a multi-scale digital twin from microscopic material damage to macroscopic overall machine performance, used for virtual-real synchronous evolution to generate virtual fault data and calibrate online prediction models; and provide optimized maintenance decisions; S4: Real-time lifespan prediction based on edge-cloud collaborative architecture. Construct an edge-cloud collaborative computing architecture model to predict the real-time remaining lifespan of coal-to-oil pumps.

2. The method for predicting the remaining life of a coal-to-oil pump according to claim 1, characterized in that, The high-temperature modified erosion wear model is based on the Archard wear model and modified to obtain the following formula. The erosion wear rate W (mm / a) of the flow-through components of the coal-to-oil pump at any spatial location under specific operating conditions can be calculated using the following formula: Where K is the overall wear coefficient; Temperature-dependent material hardness; It is a function of the impact angle; Let be the particle concentration distribution function; This is the reference hardness at room temperature. The normal impact force of the particles is the normal component of a single oil-coal slurry particle when it vertically impacts the material surface. The unit is N, and it is obtained by CFD solid-liquid two-phase flow simulation calculation. The relative impact velocity of the particles; It represents the nominal hardness of the material, the same as the value of H0, and is used for dimensionless processing in formulas.

3. The method for predicting the remaining life of a coal-to-oil machine pump according to claim 1, characterized in that, The corrosion-fatigue interactive damage model clearly distinguishes and quantitatively describes two fundamentally different interaction mechanisms, namely: Mechanism A: Corrosion accelerates fatigue, serving as the dominant mechanism; Mechanism B: Fatigue accelerates corrosion as a feedback mechanism; The final output of the model is the total corrosion-fatigue coupled damage D_total at any time. When D_total reaches the critical damage threshold D_cr, the component is determined to have suffered corrosion fatigue failure. 1) The process of establishing the basic model is as follows: a. First, establish a pure corrosion rate model under stress-free conditions. in, The pure corrosion current density (A / m²) For exchange current density; The activation energy of the reaction is expressed in J / mol. and They are respectively as well as The partial pressure (MPa); a and b are the reaction order, which are usually taken as a = 0.5~0.8 and b = 0.3~0.6 under coal-to-oil conditions; Pure corrosion rate: in, The molar mass of the metal; The number of electrons in the reaction; It is Faraday's constant; The density of the material; b. Pure fatigue crack propagation model The fatigue crack propagation rate under corrosion-free conditions is described using the Paris equation: in, The crack propagation rate per cycle (m / cycle); Stress intensity factor amplitude (MPa·m) 0.5 ); and These are material constants, obtained through room temperature fatigue tests; 2) Realization of corrosion-accelerated fatigue mechanism a. Corrosion pit-induced crack initiation model Corrosion pits generate significant stress concentrations. The stress concentration factor K_t is related to the pit depth a and radius r as follows: Introducing an equivalent initial crack size a_eq for corrosion pits: The crack initiation life induced by corrosion pits was calculated using Miner's damage theory: ,in, The fatigue life corresponding to the initial crack size a_eq; b. Corrosive Media Accelerated Crack Propagation Model The corrosion fatigue damage factor (CFF) is introduced to quantitatively describe the accelerating effect of H2S / CO2 on crack propagation rate: Where A is the material-medium coupling coefficient, which is determined by high temperature and high pressure corrosion fatigue test; For reference corrosion current density; It is a correction function for stress intensity factor, temperature, and pressure; For typical coal-to-oil operating conditions, the specific form of the correction function is as follows: in, This is the corrosion fatigue threshold value; It is the activation energy for corrosion fatigue; , This is an experimental constant; Corrected corrosion fatigue crack propagation rate: ； 3) Realization of fatigue-accelerated corrosion mechanism a. Dynamic growth model of corrosion product film Under high temperature and pressure, a corrosion product film composed of FeS, FeCO3, etc., will form on the metal surface. The growth of the film follows parabolic dynamics. in, Let t be the film thickness (m) at time t; The steady-state film thickness; The membrane growth time constant; The protective effect of a membrane is represented by the membrane protection coefficient η: ,in The film density coefficient; Corrosion rate considering the protective effect of the film: ； b. Criteria and Repair Model for Membrane Rupture under Cyclic Loading When the strain generated by cyclic stress exceeds the fracture strain of the corrosion product film... ε f At that time, the membrane ruptured: ε max > ε f After the film ruptures, the fresh metal surface is exposed, and the corrosion rate instantly recovers to near the pure corrosion rate v. corr,0 Then the film begins to regrow, and the corrosion rate gradually decreases. Introducing membrane rupture area fraction ϕ Describe the degree of damage to the membrane under cyclic loading: in, This represents the initial fracture area fraction. , This is an experimental constant; This represents the number of loop iterations. Corrected actual corrosion rate: 。 4. The method for predicting the remaining life of a coal-to-oil pump according to claim 1, characterized in that, The dynamic load fatigue cumulative damage model ultimately outputs the fatigue cumulative damage D_fatigue at any time. When D_fatigue reaches the critical damage threshold D_cr, which is usually taken as 0.7~0.8, the coupling effect of corrosion and wear is considered to determine that the component has fatigued.

5. The method for predicting the remaining life of a coal-to-oil pump according to claim 4, characterized in that, The construction and implementation process of the dynamic load fatigue cumulative damage model is as follows: 1) Implementation of dynamic load spectrum preprocessing and rainflow counting a. Load spectrum acquisition and preprocessing: The fatigue load of the coal-to-oil pump comes from: pressure pulsation of coal-oil slurry, with a main frequency of 10~50Hz; and periodic load generated by impeller rotation, with the main frequency equal to the rotation speed. Low-frequency load fluctuations caused by changes in operating conditions, 0.01~1Hz; Raw data were collected from pump outlet pressure sensor, bearing vibration sensor, and motor current sensor, with a sampling frequency ≥1kHz; high-frequency noise and electromagnetic interference were removed using wavelet threshold denoising; and the multi-source data were fused to convert it into an equivalent stress load spectrum. in, , , The stress conversion factor is calibrated through finite element analysis; b. Improved rainflow counting: The four-point rainflow counting method is used to process the non-stationary dynamic load spectrum and extract the amplitude, mean and cycle number of each load level; Improved for coal-to-oil: The working condition marking function is added to associate each load cycle with the corresponding temperature, pressure and flow rate working condition parameters to provide a basis for subsequent working condition fluctuation correction. 2) Temperature-corrected material SN curve model a. Basic SN curve The fatigue life of materials at room temperature follows a power function relationship: in, and m are material constants, determined by standard fatigue tests; b. High-temperature SN curve To address the issue that the operating temperature of coal-to-oil pumps can significantly reduce material fatigue strength and induce thermal fatigue effects, a temperature correction factor K_T is introduced: in, The fatigue limit at temperature T; This represents the fatigue limit at room temperature. The melting point of the material; The temperature sensitivity coefficient is determined through high-temperature fatigue testing. For commonly used wear-resistant alloy steels, Take a value of 0.3~0.5; Corrected high-temperature SN curve: c. Mean stress correction The effect of mean stress on fatigue life is corrected using the Goodman formula: in, The tensile strength of the material at temperature T; 3) Improved fatigue cumulative damage model a. Introduction of the load sequence effect coefficient Traditional Miner's Theory: in, Let be the number of cycles for the i-th level load; This represents the fatigue life under this load level. High loads generate residual compressive stress within the material, slowing down the accumulation of damage under subsequent low loads; while low loads generate microcracks, accelerating the propagation of damage under subsequent high loads. Introducing the load sequence effect coefficient λ ij Describe the effect of the i-th level load on the damage caused by the j-th level load: in, , The load sequence sensitivity coefficient was determined by variable amplitude fatigue test, with β=0.2~0.4 and γ=0.3~0.

5. This represents the maximum stress in the load spectrum; b. Introducing a working condition fluctuation coefficient: During coal-to-oil production, frequent switching of working conditions can cause changes in the statistical characteristics of the load spectrum, accelerating fatigue damage; introducing a working condition fluctuation coefficient... : in, The standard deviation of the load amplitude; is the average load amplitude; f (times / hour) is the operating condition switching frequency; The sensitivity coefficient to operating condition fluctuations is calibrated using field data. =0.1~0.3; c. Improved cumulative damage calculation, taking into account both load sequence effect and working condition fluctuation effect, the improved fatigue cumulative damage formula is as follows: The calculation logic is as follows: Arrange the loads at each level in chronological order; for the i-th level load, calculate the cumulative influence coefficient of all previous loads on it. Multiply by the operating condition fluctuation coefficient The actual damage contribution of this load level is obtained; the damage contributions of all loads are summed to obtain the total cumulative fatigue damage. 4) Model Coupling and Solution Process a. Coupling with other damage models: This model needs to be bidirectionally coupled with the high-temperature modified erosion wear model and the corrosion-fatigue interactive damage model. Wear-fatigue coupling: Erosion wear leads to a reduction in component cross-section and an increase in stress, through Update the load spectrum; Corrosion-fatigue coupling: Corrosion pits act as stress concentration sources, through... Correct stress amplitude; Total damage calculation: ; b. Solution process Initialization: Set the initial time t=0, the initial damage D=0, and the initial cross-sectional area A=A0; Data acquisition: Collect pressure, vibration, current, and temperature data within the current time step Δt; Load spectrum generation: Converting multi-source data into an equivalent stress load spectrum σ(t); Rainflow counting: Rainflow counting is performed on σ(t) to obtain the amplitude, mean, and number of cycles of each load level; SN curve correction: Calculate the corrected SN curve N(T, σ_a) based on the current temperature T; Damage calculation: a. Calculate the nominal damage n_i / N_i for each load level. b. Calculate the load sequence effect coefficient λ_ji c. Calculate the operating condition fluctuation coefficient ξ d. Calculate the fatigue damage increment ΔD_fatigue at the current time step; Total damage update: ; Coupled update: Update component cross-sectional area and stress concentration factor based on wear and corrosion damage; Termination judgment: If D_total ≥ D_cr, output the remaining lifetime; otherwise, t = t + Δt, return to the data acquisition step.

6. The method for predicting the remaining life of a coal-to-oil machine pump according to claim 1, characterized in that, The neural network loss function in S2 is designed as a physical constraint loss function. The loss function consists of three parts: in, The data fitting loss is calculated based on real sensor data. The physical constraint loss is calculated based on a five-field coupling degradation mechanism model, forcing the model output to conform to physical laws. To smooth out losses and ensure the continuity and stability of prediction results; The design operating condition identification network identifies normal, fluctuating, and sudden operating conditions in real time; and automatically adjusts according to the operating condition type. as well as Weights: Under stable operating conditions, increase the weight of data loss. To improve prediction accuracy; Increase the weight of physical loss under fluctuating or abrupt operating conditions. This ensures that the prediction results conform to physical laws; The hierarchical structure of the working condition identification network is as follows: Input layer: As a multi-sensor timing window, it adopts a sliding window mechanism with window size window_size=64 and stride=1; Data preprocessing layer: Normalizes, denoises, and fills in missing values ​​for the sensor-acquired data; Multi-scale TCN feature extraction layer: Three parallel TCN branches are used to capture the operating conditions at different time scales. Branch 1 captures the rapid change features of abrupt operating conditions; Branch 2 captures the medium-speed change features of fluctuating operating conditions; and Branch 3 captures the slow change features of normal operating conditions. The outputs of the three branches are concatenated along the channel dimension to obtain the fused feature map; Attention mechanism layer: includes channel attention module, used to adaptively learn the importance weights of different sensors for identifying working conditions; The temporal attention module adaptively learns the importance weights of different time steps for identifying the current working condition, focusing on the most recent time step and the point of change. The classification head and output layer include: global average pooling: compressing temporal features into a fixed-length feature vector; and outputting the probability distribution and scores of the three working conditions through fully connected layer 1 and fully connected layer 2. The basic model is pre-trained using a large amount of data from general pumps and other coal chemical equipment; a parameter-efficient fine-tuning technique is adopted, using only a small amount of normal operation data and historical fault data of the target coal-to-oil machine pump for fine-tuning; a domain-adaptive algorithm is introduced to reduce the data distribution difference between the source domain and the target domain.

7. The method for predicting the remaining life of a coal-to-oil pump according to claim 6, characterized in that, The method for implementing cross-domain fine-tuning of the model is as follows: 1) Large-scale pre-training in the source domain a) Source domain dataset construction, including the following three categories: General industrial pump dataset; General coal chemical pump dataset; Similar pump dataset; Perform data standardization processing on each type of data separately; b. Basic Model Architecture The APINN architecture, which is completely consistent with the target model, is used as the base model to ensure the effectiveness of the transfer. c. Pre-training task design A two-stage pre-training strategy of self-supervised pre-training + supervised pre-training is adopted: The self-supervised pre-training utilizes a large amount of unlabeled data to accomplish the following tasks: Task 1: Time series prediction task, predict sensor data for the next 10 time steps; Task 2: Mask Reconstruction Task. Reconstruct the masked data by randomly masking the data for 15% of the time steps. Task 3: Comparative learning tasks enhance similar samples from the same device and differentiate samples from different devices; The supervised pre-training utilizes labeled data: Task: Residual lifetime regression prediction; Loss function: MSE loss combined with physical constraint loss; Training rounds: 50 rounds, learning rate 1e-3; 2) Unsupervised domain adaptive implementation The data distribution differences between the source domain and the target domain mainly stem from: Differences in operating conditions: Coal-to-oil machine pumps have higher temperatures, pressures, and solid content; Equipment differences: Differences in structure and performance between pumps from different manufacturers and models; Sensor differences: Differences in accuracy and noise characteristics between different sensors; Therefore, a hybrid approach combining adversarial domain adaptation with maximum mean difference is adopted to align marginal and conditional distributions. a. Domain-Adaptive Network Architecture Design Based on the basic APINN model, a domain discriminator and an MMD loss calculation module are added, including a data input layer and an APINN feature extraction layer; Branch 1: The RUL prediction head obtains the RUL prediction loss; Branch 2: The neighborhood discriminator obtains the adversarial loss; Branch 3: The MMD loss calculation module obtains the distribution alignment loss. b. Loss Function Design The total loss function consists of three parts: Among them, mission losses RUL prediction MSE loss calculated using only labeled data from the source domain; Combat losses Domain adversarial training is achieved using a gradient inversion layer. in For the domain discriminator, , These represent the number of samples in the source domain and the target domain, respectively. MMD loss : Calculate the distance between the features of the source and target domains in the reproducing kernel Hilbert space: in, The kernel mapping function uses a Gaussian kernel; c. Training process Freeze the APINN feature extraction layer and RUL prediction head, and train only the domain discriminator for 5 rounds; Freeze the neighborhood discriminator, train the APINN feature extraction layer and RUL prediction head, and iterate for 1 round; The two steps above are trained alternately for a total of 20 iterations; Learning rate setting: Feature extraction layer 1e -5 Domain discriminator 1e -4 ; 3) Parameters can be finely tuned efficiently. Based on the characteristics of the APINN model, LoRA was selected as the core PEFT technique; a. In the original weight matrix W0∈R d×k Add two low-rank matrices A∈R next to it r×k and B∈R d×r During training, only A and B are updated; during inference, BA is added to W0. W = W0 + BA; The insertion positions of LoRA in the APINN model are as follows: All convolutional layers of a multi-scale TCN layer; A fully connected layer with physical constraint layers; Fully connected layer of the RUL prediction head; Key parameter settings in the model: rank r: set to 8 or 16 to balance parameter quantity and performance; scaling factor α: set to r to ensure that the contribution of LoRA is 0 during initialization; Dropout: set to 0.1 to prevent overfitting; The fine-tuning training process is as follows: a. Dataset partitioning: Training set: 3 months of normal operation data of the target coal-to-oil machine pump plus 1-2 historical fault data; Validation set: runtime data from the most recent month; Test set: Data from one reserved fault test; b. Training process Completely freeze all parameters of the pre-trained APINN model; Insert a LoRA matrix into the specified layer, and initialize A to a Gaussian distribution and B to 0; Only the LoRA matrix is ​​trained, while other parameters remain unchanged; Loss function: MSE loss combined with physical constraint loss; Training rounds: 20-30 rounds, early stop strategy; Learning rate: 1e -4 ; After training, the LoRA matrix BA is merged into the original weight matrix W0, resulting in a fine-tuned model with the same structure as the original model.

8. The method for predicting the remaining life of a coal-to-oil machine pump according to claim 1, characterized in that, The multi-scale digital twin adopts a three-layer nested architecture of micro-meso-macro, with bidirectional information transmission: The macroscopic scale serves as a digital twin of overall system performance degradation, used for real-time state synchronization between physical devices and virtual models, providing global visualization and rapid lifespan prediction. The overall performance model is a characteristic curve model of the pump, and the performance indicators of the pump are calculated based on the current operating parameters. The performance degradation model is based on the component wear amount output at the mesoscale, which is used to correct the characteristic curve of the pump and simulate the performance degradation process over time. The mesoscale data serves as a multi-physics coupled digital twin of key components, used for five-field coupled calculations involving flow field, temperature field, stress field, wear field, and corrosion field, to accurately predict the damage distribution of key components. The multiphysics coupled solver design adopts a hybrid coupling strategy with weak coupling as the main component and strong coupling in key regions: Flow field solution: The solid-liquid two-phase flow field of oil-coal slurry is solved based on large eddy simulation to obtain the particle velocity, concentration, and impact angle distribution; Temperature field solution: Solving for the temperature distribution of the flow field and the temperature distribution of the solid component based on the energy equation; Stress field solution: The stress and strain distribution of the component is solved based on the finite element method, taking into account temperature stress and fluid dynamic pressure; Wear field solution: Use the high-temperature modified erosion wear model to calculate the wear rate and cumulative wear amount of each grid cell; Corrosion field solution: Call the corrosion-fatigue interactive damage model to calculate the corrosion rate and corrosion damage of each grid cell; To meet real-time requirements, the high-fidelity finite element model is reduced in order: an orthogonal decomposition method is used to project the high-dimensional physical field data onto a low-dimensional subspace; and a mapping relationship is established between the reduced-order model and the operating parameters. As wear progresses, the geometry of the component changes. The computational mesh is dynamically updated using an arbitrary Lagrange-Euler method, as follows: Each update cycle adjusts the mesh node positions based on the amount of wear. When the mesh deformation exceeds the threshold, the mesh is automatically re-divided; Achieve bidirectional coupling between damage evolution and geometric changes; The microscale, serving as a digital twin of material damage evolution, is used to simulate the microscopic damage evolution process within the material, providing material property parameters for the mesoscale; it includes the following models: Material microstructure modeling, crystal plasticity model, electrochemical corrosion model, microcrack propagation model; Offline pre-calculation combined with online interpolation: Due to the extremely large amount of computation at the microscale, offline pre-calculation is used to calculate the material response under different working conditions and damage levels to establish a database; during online runtime, material property parameters are quickly obtained through interpolation.

9. The method for predicting the remaining life of a coal-to-oil machine pump according to claim 1, characterized in that, The virtual-real synchronous evolution process is implemented as follows: a) Real-time access and fusion of multi-source heterogeneous data The collected data includes multi-source heterogeneous data such as process data, status data, offline data, and environmental data; A spatiotemporal alignment algorithm is used to unify data with different sampling frequencies and timestamps to the same time base; a federated filtering algorithm is used to fuse multi-sensor data to improve data reliability and accuracy. Establish a data quality assessment model to detect sensor drift, missing data, and outliers in real time; b. Asynchronous updating and dynamic calibration of multi-scale models At a macro scale, it achieves second-level state synchronization and updates the virtual model's operating parameters and overall performance indicators in real time; it adopts a proportional-integral-derivative controller to adjust the virtual model's input parameters in real time based on sensor measurements, ensuring that the virtual model's output is consistent with the physical device; At the mesoscale, damage synchronization is achieved within minutes, updating the wear distribution, stress distribution, temperature distribution, and corrosion damage of key components; An extended Kalman filter algorithm is used in conjunction with sensor data to perform online calibration of the multiphysics model; based on the arbitrary Lagrange-Euler method, the mesh node positions of the virtual model are adjusted in real time according to the calculated wear amount to accurately simulate the geometric changes of the flow components caused by wear; when the mesh deformation exceeds 20%, mesh reconstruction is automatically triggered to ensure calculation accuracy. At the microscale, the system synchronizes the properties of materials at the space level, updates the mechanical performance parameters and micro-damage parameters of the materials, and uses Bayesian parameter estimation methods to update the parameters of the micro-damage model by combining offline material test data and online operation data. c. Virtual-real bidirectional mapping and feedback control Forward mapping: Changes in the state of physical devices are transmitted to the digital twin in real time, driving the evolution of the virtual model; Reverse mapping: The simulation results and prediction information of the digital twin are fed back to the physical system to guide production operation and maintenance decisions; When an abnormal operating condition is detected, a warning signal is sent to the control system, suggesting adjustments to the operating parameters; After maintenance is completed, the damage status of the digital twin is updated according to the maintenance record, and the lifespan of the corresponding components is reset.

10. The method for predicting the remaining life of a coal-to-oil pump according to claim 1, characterized in that, The edge-cloud collaborative computing architecture model includes: Application layer: As an intelligent operation and maintenance visualization platform, it specifically implements equipment status display, lifespan prediction display, maintenance work order management, and report statistics; Cloud layer: As the global computing and decision-making center, it specifically implements data management, model training and version management, high-precision life assessment, multi-scale digital twin engine, and generates intelligent maintenance decisions; Network layer: as a communication system for the Industrial Internet of Things; Edge layer: As a real-time computing and control unit on site, it specifically implements multi-source data acquisition, real-time data preprocessing, real-time operating condition identification, lightweight APINN inference module, local hierarchical early warning, and edge-cloud data interaction; Perception layer: Generates a multi-mode sensor network; The edge terminal achieves real-time prediction of remaining service life, as follows: 1) Input data: Includes data from eight core sensors; 2) Real-time operating condition identification: Call the lightweight TCN operating condition identification network to output the current operating condition probability distribution; determine the operating condition type based on the maximum probability: normal p>0.9, fluctuation 0.7≤p≤0.9, sudden change p<0.7; 3) Lightweight APINN model inference Model structure: The multi-scale TCN feature extraction layer includes 3 parallel branches with convolutional kernel sizes of 3 / 5 / 7; a physical constraint layer; and 2 fully connected prediction heads. Inference Output: Remaining Life Prediction: in days; Prediction Confidence: a value between 0 and 1; Risk Level: Normal, RUL > 30 days, confidence > 0.7; Warning, 7 days ≤ RUL ≤ 30 days; Critical, RUL < 7 days; 4) Local early warning and data upload When the risk level is severe, a local audible and visual alarm is immediately triggered, and the PLC is linked to reduce the operating load of the pump. Data upload strategy: Under normal operating conditions: average data and prediction results are uploaded every 10 minutes; Fluctuating operating conditions: Data is uploaded every 1 minute; Abnormal operating conditions: Upload raw data and early warning information in real time; The cloud-based hourly high-precision lifespan calibration method is as follows: 1) Data reception and fusion Receive real-time data and prediction results uploaded from the edge device; Integrate offline data; Conduct data quality assessments, remove outliers, and complete missing data. 2) High-precision life assessment High-precision lifetime prediction is achieved by calling the full APINN model. Input features include: Sensor data uploaded from the edge; Multiphysics characteristics of digital twin output; Medium characteristic parameters; The Monte Carlo simulation method is used to quantify data uncertainty, model uncertainty, and operating condition uncertainty, and output the 95% confidence interval of the remaining lifetime. Generate a cloud map of the overall machine's lifespan distribution to accurately locate the most vulnerable locations and failure modes; 3) Model Updates and Pushes The APINN model is fine-tuned weekly using newly collected operational data and virtual data generated by the digital twin. Incremental updates are used, transmitting only the parameters that change in the model, thus reducing bandwidth consumption; A canary release strategy is adopted, first pushing to 10% of edge nodes for verification, and then pushing to the full population after there are no problems. The edge device employs a dual-model hot-switching mechanism, ensuring that the update process does not affect normal operation.

Images