Multi-oil-cylinder synchronous equipment digital twin safety monitoring system and method based on physical information neural network

Abstract

The present application relates to a kind of based on physical information neural network multi-oil cylinder synchronous equipment digital twin safety monitoring system and method, system includes multi-source sensing acquisition module, physical-data double-drive twin modeling module, twin residual error driven state deduction module and hierarchical safety evaluation module;Physical-data double-drive twin modeling module is based on physical information neural network and constructs multi-oil cylinder digital twin model, and the model will the physical equation of hydraulic synchronous system as soft constraint term embed in the loss function of physical information neural network;State deduction module is used to calculate the twin residual error between predicted value and the measured value of corresponding physical quantity, to trigger parameter identification, update the scalar coefficient of physical equation in physical information neural network;Hierarchical safety evaluation module calculates comprehensive safety index, to evaluate safety state.Compared with prior art, the present application greatly improves the precision and reliability of state perception, risk prediction and safety control in multi-oil cylinder synchronous construction process.

Description

Technical Field

[0001] This invention relates to the field of multi-cylinder safety monitoring technology, and in particular to a digital twin safety monitoring system and method for multi-cylinder synchronous equipment based on physical information neural networks. Background Technology

[0002] In recent years, with the rapid development of infrastructure construction, marine engineering, and heavy equipment manufacturing, multi-cylinder collaborative drive technologies such as hydraulic synchronous lifting, synchronous jacking, and multi-cylinder linkage support have been widely applied in key construction scenarios such as the overall lifting of large-span steel structure roofs, the assembly of ultra-large bridge segments, the installation of offshore wind turbine jackets, and the horizontal movement of large components. The displacement synchronization accuracy, load balance, and health status of key components of the multi-cylinder synchronous system directly determine construction safety, component installation accuracy, and work efficiency. Any displacement deviation, pressure imbalance, or component deterioration in any cylinder may lead to serious accidents such as uneven loading of the overall structure, attitude instability, or even component fall. Therefore, constructing a high-fidelity, highly reliable, and interpretable digital twin safety monitoring system for multi-cylinder synchronous equipment has become a pressing technical challenge in the engineering field.

[0003] To address this need, existing technologies mainly follow three technical routes. The first route is a digital twin solution based on pure mechanistic modeling, represented by finite element simulation and AME Sim hydraulic system simulation. This involves constructing a virtual simulation model by performing detailed physical modeling of the cylinder structure, hydraulic circuit, and load constraints. While this approach offers strong physical interpretability, once the modeling parameters are fixed, it struggles to adapt to the evolving operating conditions of the cylinder over long-term operation, such as seal wear, changes in oil viscosity, and leakage coefficient drift. This leads to a gradual deviation between the virtual model and the physical entity, resulting in mapping failure. The second route is a purely data-driven intelligent monitoring solution, represented by long short-term memory networks, Transformer architecture, and the increasingly popular large language models. This approach uses massive amounts of historical operating data for training to construct an end-to-end state mapping model. While this approach possesses some adaptability, it is essentially a black-box modeling approach. Because model training heavily relies on large-scale labeled data, and construction sites, especially with fault samples, are extremely scarce, it easily leads to underfitting of the model. Furthermore, the model has weak extrapolation capabilities for operating conditions not covered by the training set, and its prediction accuracy drops sharply under extreme conditions such as sudden wind load changes, abnormal loads, and oil contamination. Finally, the model's decision-making process is uninterpretable, making it difficult to meet the safety review and accountability requirements of major engineering projects. The third approach is a hybrid scheme that simply superimposes mechanisms and data, typically using data models to post-process and correct the mechanism simulation results. However, in this type of scheme, mechanisms and data are independent and lack deep coupling, failing to fundamentally overcome the shortcomings of the above two methods.

[0004] In summary, existing digital twin monitoring technologies generally suffer from common defects such as distorted and fixed mechanistic models, uninterpretable black-box data models, lack of multi-cylinder coupling modeling, and inability of twin models to adaptively evolve. These defects fail to meet the engineering requirements of multi-cylinder synchronous equipment for accurate state perception, forward-looking risk prediction, and reliable safety control. There is an urgent need to develop a new digital twin safety monitoring solution that combines data-driven adaptability with physical interpretability. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a digital twin safety monitoring system and method for multi-cylinder synchronous equipment based on physical information neural network. This system fundamentally overcomes the technical bottlenecks of traditional digital twin schemes, such as mapping distortion, extrapolation failure, small sample inaccuracy, and lack of interpretability, and significantly improves the accuracy and reliability of state perception, risk prediction, and safety control during multi-cylinder synchronous construction.

[0006] The objective of this invention can be achieved through the following technical solutions: A digital twin safety monitoring system for multi-cylinder synchronous equipment based on physical information neural network includes a multi-source sensor acquisition module, a physical-data dual-drive twin modeling module, a twin residual-driven state inference module, and a hierarchical safety assessment module. The multi-source sensor acquisition module is deployed at the end of each cylinder in the multi-cylinder synchronous construction site to collect data on the extension displacement, working pressure, oil temperature and flow rate of each cylinder in real time. The physical-data dual-drive twin modeling module constructs a multi-cylinder digital twin model based on a physical information neural network. This multi-cylinder digital twin model embeds the physical equations of the hydraulic synchronization system as soft constraints into the loss function of the physical information neural network. The soft constraints include physical residual loss and multi-cylinder coupling constraint loss. The physical residual loss is obtained by back-solving the dynamic equation residuals of the rigid body of the hydraulic cylinder and the continuity equation residuals of the fluid in the hydraulic cavity based on the output of the physical information neural network, and combining them. The multi-cylinder coupling constraint loss includes a constraint term to force the displacement of each cylinder to remain synchronous and consistent, and a constraint term to force the sum of the bearing capacities of each cylinder to equal the total weight of the component. The twin residual-driven state deduction module is used to calculate the twin residual between the predicted value of the multi-cylinder digital twin model and the measured value of the corresponding physical quantity, so as to trigger parameter identification and update the scalar coefficients of the physical equation in the physical information neural network. The graded safety assessment module is used to calculate a comprehensive safety index based on the multi-cylinder state vector output by the multi-cylinder digital twin model, so as to assess the safety status of the multi-cylinder synchronous construction site.

[0007] Furthermore, the physical equations include the hydraulic cylinder rigid body dynamics equation for calculating the residuals of the dynamics equations, and the hydraulic cavity fluid continuity equation for calculating the residuals of the continuity equations. The physical parameters covered by the rigid body dynamics equation of the hydraulic cylinder include: equivalent mass of the cylinder piston, damping coefficient, equivalent stiffness, effective working area of ​​the two chambers, and external load force. The physical parameters covered by the fluid continuity equation of the hydraulic cavity include: cavity volume, bulk modulus of the oil, instantaneous flow rate, and leakage coefficient. The physical residual loss is obtained by normalizing the residuals of the dynamic equation and the residuals of the continuity equation, and then performing a mean-square weighted summation. The residuals of the dynamic equations are normalized to the square of the rated load force, and the residuals of the continuity equations are normalized to the square of the rated pressure.

[0008] Furthermore, the calculation expression for the multi-cylinder coupling constraint loss is as follows: In the formula, For multi-cylinder coupling constraint loss, For parameters of the multi-cylinder digital twin model, The total number of hydraulic cylinders, superscript Indicates the first hydraulic cylinder, and The first Predicted values ​​of displacement and pressure of the hydraulic cylinder. For the first The displacement value of the hydraulic cylinder. The effective working area of ​​the rodless chamber of the hydraulic cylinder. The total weight of the lifted component, The number of time sampling points, The relative weights of the load conservation terms.

[0009] Furthermore, the loss function of the physical information neural network also includes data fitting loss and boundary condition loss; The data fitting loss is the mean square error of the predicted cylinder state output by the multi-cylinder digital twin model and the corresponding measured value output by the multi-source sensor acquisition module over all sampling points. The boundary condition loss includes an initial condition part and a boundary condition part. The initial condition part corresponds to the initial displacement, initial pressure and initial temperature before the cylinder is lifted. The boundary condition part corresponds to the oil inlet flow boundary and pressure boundary of the hydraulic pump station. The loss function of the physical information neural network is a weighted sum of data fitting loss, physical residual loss, boundary condition loss, and multi-cylinder coupling constraint loss.

[0010] Furthermore, the multi-source sensing acquisition module includes a laser displacement sensor, an oil pressure sensor, a temperature sensor, and a flow sensor. The laser displacement sensor is installed at the end of the piston rod of each oil cylinder or at the lifting point of the component to collect the extension displacement of the oil cylinder in real time, so as to reflect the lifting or pushing displacement state of the component. The hydraulic pressure sensors are installed at the oil circuit interfaces of the rodless and rod chambers of each hydraulic cylinder to collect the working pressure of the hydraulic cylinder in real time, so as to reflect the load distribution status. The temperature sensor is installed in the oil circuit of the cylinder or at the oil outlet of the hydraulic station to collect the oil temperature in real time, so as to help evaluate the working status of the hydraulic system. The flow sensor is installed in the main oil inlet of each cylinder to collect the oil flow rate in real time, so as to assist in solving the physical equation residuals. The data collected by the laser displacement sensor, oil pressure sensor, temperature sensor, and flow sensor are calibrated with a unified timestamp to achieve spatiotemporal alignment for subsequent processing.

[0011] Furthermore, the physical information neural network includes an input embedding layer, a deep fully connected backbone network, and a physical constraint calculation layer connected in series. The input embedding layer is used to encode the spatiotemporal coordinates of the input in a high dimension using Fourier feature embedding technology. The spatiotemporal coordinates include time coordinates and spatial coordinates, and the spatial coordinates correspond to the cylinder number and piston position. The deep fully connected backbone network is used to learn the spatiotemporal evolution of the displacement field, pressure field, and temperature field of multiple hydraulic cylinders. The physical constraint calculation layer is used to solve the dynamic equation residuals and continuity equation residuals of the hydraulic system's physical equations from the output of the neural network based on automatic differentiation technology.

[0012] Furthermore, the twin residual-driven state deduction module includes a real-time residual calculation submodule, an accumulated threshold judgment submodule, an online parameter identification submodule, and a parameter back-injection update submodule; The residual real-time calculation submodule is used to calculate the twin residual between the predicted value of the multi-cylinder digital twin model and the measured value of the corresponding physical quantity in real time; The cumulative threshold judgment submodule is used to calculate the root mean square value of the twin residual norm output by the real-time residual calculation submodule under each sliding window, and use it as the cumulative residual value. The online parameter identification submodule is used to trigger when the cumulative residual value exceeds a preset threshold, and to identify the damping coefficient, equivalent stiffness, leakage coefficient and oil volumetric elastic modulus online based on the recursive least squares method. The parameter back-injection update submodule is used to back-inject the parameter identification results output by the online parameter identification submodule into the scalar coefficients of the physical equations in the physical information neural network.

[0013] Furthermore, the process of calculating the comprehensive safety index by the graded safety assessment module includes: calculating the displacement synchronization error, pressure equalization error and twin residual error based on the multi-cylinder state vector output by the multi-cylinder digital twin model, and then performing weighted fusion to obtain the comprehensive safety index; The displacement synchronization error is the normalized standard deviation of the real-time displacement of each cylinder relative to the average displacement of multiple cylinders. The pressure equalization error is the ratio of the standard deviation of the multi-cylinder pressure to the pressure mean. The twin residual error is the normalized result of the cumulative residual value of the twin residual output by the state inference module driven by the twin residual, and the corresponding normalization benchmark is the residual trigger threshold. The assessment of the safety status of the multi-cylinder synchronous construction site includes: comparing the comprehensive safety index with the preset safety threshold and danger threshold; if the comprehensive safety index is not higher than the safety threshold, it is in normal operation status; if the comprehensive safety index is higher than the safety threshold but not higher than the danger threshold, it is in a status warning status; if the comprehensive safety index is higher than the danger threshold, it is in a high-risk shutdown status.

[0014] Furthermore, the system also includes a B / S visualization interaction module, used for displaying multi-cylinder displacement synchronization curves, pressure balance bar charts, oil temperature trend curves, twin residual heat maps, real-time mapping of three-dimensional virtual models, safety level color and audio-visual prompts, historical data storage, query and playback, and remote adjustment of sensor parameters and safety thresholds.

[0015] This invention also provides a digital twin safety monitoring method for multi-cylinder synchronous equipment based on physical information neural networks, comprising the following steps: Real-time acquisition of data on the extension displacement, working pressure, oil temperature, and flow rate of each hydraulic cylinder during synchronous construction of multiple hydraulic cylinders; The collected data is input into a multi-cylinder digital twin model constructed based on a physical information neural network, and the predicted values ​​of the multi-cylinder state vectors are output to complete the real-time synchronous mapping between the physical entity and the virtual model. During the training process of the multi-cylinder digital twin model, the physical equations of the hydraulic synchronization system are embedded as soft constraints into the loss function of the physical information neural network, and offline pre-training is performed based on historical operating data. Calculate the twin residuals between the predicted values ​​of the multi-cylinder digital twin model and the measured values ​​of the corresponding physical quantities to trigger parameter identification and update the scalar coefficients of the physical equations in the physical information neural network; The comprehensive safety index is calculated based on the multi-cylinder state vector output by the multi-cylinder digital twin model to assess the safety status of the multi-cylinder synchronous construction site.

[0016] Compared with the prior art, the present invention has the following advantages: (1) This invention is the first to use the hydraulic cylinder dynamics equation, fluid continuity equation and multi-cylinder coupling constraint equation as the loss function of soft constraint deep embedding neural network, and constructs a physical information neural network twin model with deep coupling of mechanism and data, so that the twin model has both the flexible adaptability of data-driven and the rigorous interpretability of physical mechanism, and significantly improves the fidelity and engineering credibility of twin mapping.

[0017] (2) Based on the inherent constraint of physical equations, this invention can still maintain the continuity and stability of twin mapping by relying on physical mechanism deduction in the event of abnormal situations such as temporary failure of some sensors, reduced sampling rate, and short-term data loss. This forms a highly robust monitoring capability with dual insurance of sensing and mechanism deduction, significantly improving the engineering reliability of the system.

[0018] (3) The physical information neural network architecture of the present invention follows the laws of fluid mechanics and dynamics in depth. It has a significantly better extrapolation prediction capability than pure data-driven models for extreme working conditions, new load conditions and abnormal disturbance scenarios not covered by the training set. This enables the monitoring system to provide reliable state perception and risk prediction under rare but high-risk working conditions such as sudden wind load changes, abnormal loads and oil pollution.

[0019] (4) The system architecture of the present invention has strong versatility and can be flexibly adapted to various multi-cylinder collaborative construction scenarios such as synchronous lifting of large-span steel structures, synchronous jacking of ultra-large bridges, and multi-cylinder linkage support of offshore platforms. It can be quickly deployed without significant changes to the hardware configuration and has wide engineering applicability and promotion value. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of a digital twin safety monitoring system for multi-cylinder synchronous equipment based on a physical information neural network, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the workflow of the residual feedback evolution mechanism in a twin residual-driven state deduction module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the comprehensive safety index calculation and three-level safety level determination process in a graded safety assessment module provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0024] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed during use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0025] Example 1 This embodiment provides a digital twin safety monitoring system for multi-cylinder synchronous equipment based on physical information neural networks, including a multi-source sensor acquisition module, a physical-data dual-drive twin modeling module, a twin residual-driven state deduction module, and a hierarchical safety assessment module; The multi-source sensor acquisition module is deployed at the end of each cylinder in the multi-cylinder synchronous construction site to collect data on the extension displacement, working pressure, oil temperature and flow rate of each cylinder in real time. The physics-data dual-drive twin modeling module constructs a multi-cylinder digital twin model based on a physical information neural network. This multi-cylinder digital twin model embeds the physical equations of the hydraulic synchronization system as soft constraints into the loss function of the physical information neural network. The soft constraints include physical residual loss and multi-cylinder coupling constraint loss. The physical residual loss is obtained by back-solving the dynamic equation residuals of the hydraulic cylinder rigid body and the continuity equation residuals of the hydraulic cavity fluid based on the output of the physical information neural network, and combining them. The multi-cylinder coupling constraint loss includes a constraint term to force the displacement of each cylinder to remain synchronous and consistent, and a constraint term to force the sum of the bearing capacities of each cylinder to equal the total weight of the component. The twin residual-driven state deduction module is used to calculate the twin residual between the predicted value of the multi-cylinder digital twin model and the measured value of the corresponding physical quantity, so as to trigger parameter identification and update the scalar coefficients of the physical equation in the physical information neural network. The graded safety assessment module is used to calculate comprehensive safety indicators based on the multi-cylinder state vectors output by the multi-cylinder digital twin model, in order to assess the safety status of the multi-cylinder synchronous construction site.

[0026] Preferably, this system also includes a B / S visualization interaction module for displaying multi-cylinder displacement synchronization curves, pressure balance bar charts, oil temperature trend curves, twin residual heat maps, real-time mapping of 3D virtual models, safety level color and audio-visual prompts, historical data storage, query and playback, and remote adjustment of sensor parameters and safety thresholds.

[0027] Specifically, the multi-source sensor acquisition module is deployed at the end of each cylinder in the multi-cylinder synchronous construction site. It consists of a laser displacement sensor, an oil pressure sensor, a temperature sensor, and a flow sensor. The laser displacement sensor is installed at the end of the piston rod of each cylinder or at the lifting point of the component to collect the cylinder extension displacement in real time to reflect the lifting or pushing displacement of the component. The oil pressure sensor is installed at the oil circuit interface of the rodless and rod chambers of each cylinder to collect the cylinder working pressure in real time to reflect the load distribution. The temperature sensor is installed at the oil circuit of the cylinder or the oil outlet of the hydraulic station to collect the oil temperature in real time to help evaluate the working status of the hydraulic system. The flow sensor is installed at the main oil inlet of each cylinder to collect the oil flow in real time to help solve the residual of the physical equation. The data collected by each sensor is collected to the local industrial control computer through an industrial bus. After being calibrated with a unified timestamp, the spatiotemporal alignment of the multi-source heterogeneous data is completed to ensure strict consistency of the multi-cylinder data on the time axis.

[0028] The physics-data dual-driven twin modeling module, deployed on the data monitoring center server, is the core innovative module of this invention. It employs a physical information neural network architecture to construct a multi-cylinder digital twin model. The core idea of ​​this module is to deeply embed the key physical equations of the hydraulic synchronization system as soft constraints into the loss function of the neural network, enabling the twin model to strictly adhere to the physical laws of fluid mechanics and rigid body dynamics while fitting sensor measurement data.

[0029] The physical information neural network of this module consists of three concatenated parts: an input embedding layer, a deep fully connected backbone network, and a physical constraint calculation layer. The network input is spatiotemporal coordinates (time t and spatial coordinate x, where x corresponds to the cylinder number and piston position), and the network output is a multi-physics prediction vector, including real-time predictions of four physical quantities: cylinder displacement, rodless chamber pressure, rod chamber pressure, and oil temperature. The input embedding layer uses Random Fourier Feature (RFF) technology to perform high-dimensional nonlinear encoding on the low-dimensional spatiotemporal coordinates. Its core mathematical expression is: in For the input coordinate vector, Let be the frequency projection matrix, whose elements have a mean of zero and a standard deviation of... Independent sampling is performed within a Gaussian distribution; this embodiment takes... Embedded Dimension This mapping transforms two-dimensional spatiotemporal coordinates into a 128-dimensional feature space. This encoding method overcomes the spectral bias problem inherent in traditional neural networks when learning high-frequency pressure pulsations and displacement vibration signals, significantly improving the model's ability to fit the transient high-frequency response of hydraulic systems.

[0030] The deep fully connected backbone network consists of eight stacked fully connected layers, each containing 128 neurons. Hyperbolic tangent activation functions are used between layers, and the network weights are initialized using a Xavier normal distribution strategy to ensure training stability. The backbone network is responsible for learning the spatiotemporal evolution of the cylinder displacement, pressure, and temperature fields, and the last layer is a linear output layer.

[0031] The physical constraint computation layer, based on Automatic Differentiation (AD), solves the residuals of the hydraulic cylinder rigid body dynamics equation and the hydraulic cavity fluid continuity equation from the neural network output, and uses these residuals as physical loss terms in network training. Specifically, the dynamics equation residuals include parameters such as the equivalent mass of the cylinder piston, damping coefficient, equivalent stiffness, effective working area of ​​the two cavities, and external load force; the continuity equation residuals include parameters such as cavity volume, oil volumetric elastic modulus, instantaneous flow rate, and leakage coefficient. All partial derivative terms of the residuals of both types of equations are directly solved through the automatic differentiation interface of the PyTorch framework, without manual derivation. The mean square value of the residuals on the physical collocation set is used as the physical loss in training. In this embodiment, 5000 physical collocations are used, uniformly distributed in the spatiotemporal domain using a Latin hypercube sampling strategy.

[0032] Specifically, the physical equations embedded in the physics-data dual-drive twin modeling module include the rigid body dynamics equation of the hydraulic cylinder and the fluid continuity equation of the hydraulic cavity. The equations cover key physical parameters such as cylinder piston mass, damping coefficient, equivalent stiffness, effective cavity area, cavity pressure, cavity volume, oil volumetric elastic modulus, flow rate and leakage coefficient. The training data of the twin modeling module integrates hydraulic system simulation data, historical engineering operation data and typical fault case data.

[0033] For the The rigid body dynamics equation of a hydraulic cylinder can be expressed as: in, For the first The piston rod of the hydraulic cylinder extends and shifts. and These are the piston rod velocity and acceleration, respectively. The mass of the hydraulic cylinder piston and its equivalent moving parts. This is the equivalent damping coefficient. For equivalent stiffness, and These represent the effective working areas of the rodless cavity and the rod cavity, respectively. and These are the pressures in the rodless chamber and the rod chamber, respectively. For the first The external load force borne by the hydraulic cylinder.

[0034] The fluid continuity equation for hydraulic chambers describes the relationship between chamber flow rate, piston motion, fluid compressibility, and leakage. For the first... The continuity equation for the rodless chamber of a hydraulic cylinder can be expressed as: For a rod-shaped cavity, its continuity equation can be expressed as: in, and These represent the pressure change rates in the rodless chamber and the rod chamber, respectively. This is the equivalent bulk modulus of elasticity of the oil. and These are the instantaneous volumes of the rodless cavity and the rod cavity, respectively. and These represent the flow rates in the rodless cavity and the rod cavity, respectively. Let be the internal leakage coefficient of the hydraulic cylinder. The instantaneous volume of the cavity can be expressed as follows, based on the hydraulic cylinder structural parameters and piston displacement: in, and These are the initial dead space volumes of the rodless cavity and the rod cavity, respectively. This is the effective stroke of the hydraulic cylinder.

[0035] The total loss function of this module is composed of four weighted components: data fitting loss, physical residual loss, boundary condition loss, and multi-cylinder coupling constraint loss. The data fitting loss constraint model outputs values ​​close to the actual sensor readings, defined using the mean square error of the predicted and measured values ​​across all sampling points. The measured data comes from real-time acquisition results from displacement, oil pressure, temperature, and flow sensors. The physical residual loss constraint model conforms to the partial differential equations of the hydraulic system and is composed of the mean square weighted sum of the residuals from the dynamic equations and the continuity equations at the physical collocation points. Specifically, the physical information neural network... Hydraulic cylinder and timing The output includes piston displacement prediction values. Predicted value of rodless chamber pressure Predicted value of rod chamber pressure and oil temperature prediction value The result is obtained by inversely calculating the output of the network based on automatic differentiation. , , and Then, substitute the values ​​into the above physical equations to calculate the residuals. The residual of the dynamic equation of the hydraulic cylinder is defined as: When the model predicts a state that fully satisfies the rigid body dynamics equations of the hydraulic cylinder, the residual approaches zero. The continuity equation residuals for the rodless and rod-type chambers of the hydraulic cylinder are defined as follows: When the model predicts a state that satisfies the fluid continuity equation of the hydraulic cavity, the residual of the continuity equation approaches zero. Considering the significant differences in physical dimensions and numerical magnitudes between the residuals of the kinetic equation and the continuity equation, this invention introduces a normalization coefficient to normalize the two types of residuals to the same order of magnitude. Specifically, the residual of the kinetic equation is expressed as the square of the rated load force. As a normalized baseline, the residuals of the continuity equation are expressed as the square of the rated pressure. This serves as the normalization benchmark. Therefore, the... The dynamic residual loss of the hydraulic cylinder can be expressed as: in, For the number of physical points, For the first The time coordinates corresponding to each physical point. The physical residual loss of a multi-cylinder system is further expressed as: in, The total number of hydraulic cylinders. and These are the weighting coefficients for dynamic residual loss and continuity residual loss, respectively.

[0036] The boundary condition loss constraint model satisfies both initial and boundary conditions, comprising two parts: initial condition loss and boundary condition loss. The initial condition part corresponds to the initial displacement, initial pressure, and initial temperature before the cylinder is lifted, used to ensure that the output of the twin model at the initial moment is consistent with the initial state obtained from sensor acquisition or system calibration. For the first... The initial condition loss of the hydraulic cylinder can be expressed as: in, , , and The first The initial displacement, initial pressure in the rodless chamber, initial pressure in the rod chamber, and initial oil temperature are collected by the multi-source sensor acquisition module or obtained through system calibration before the hydraulic cylinder is lifted.

[0037] The boundary conditions section corresponds to the inlet flow rate and pressure boundaries of the hydraulic pump station, used to ensure that the model's predictions at boundary locations such as the pump station's inlet, supply port, or return port are consistent with the corresponding boundary conditions. These boundary conditions can originate from measured boundary values ​​acquired by the multi-source sensor acquisition module, or from flow rate and pressure setpoints given by the hydraulic control system, or from physical boundary values ​​determined by the hydraulic system structure. Taking the pump station's supply flow rate and supply pressure boundaries as an example, the boundary condition loss can be expressed as: in, The number of boundary sampling points, and These are the oil supply flow rate and oil supply pressure predicted by the model at the boundary points or derived from the model state, respectively. and These represent the corresponding boundary flow rate and boundary pressure values, respectively. When the boundary quantities are directly acquired by the sensor... and These are measured values; when the boundary quantities are set by the control system, they are control setpoints; when the boundary quantities are determined by the system structure, they are the corresponding physical boundary constants or constraint values. Both initial condition losses and boundary condition losses participate in training in the form of mean squared errors, thereby ensuring that the twin model meets the actual operating conditions at both the initial state and the hydraulic boundary.

[0038] In summary, the total loss function of this module can be expressed as: in, For data fitting loss, For physical residual loss, Loss due to initial conditions For boundary condition loss, For multi-cylinder coupling constraint loss, , , and These are the weighting coefficients for the corresponding loss terms.

[0039] The weights of the four loss classes are dynamically adjusted using an adaptive balancing mechanism based on the Neural Tangent Kernel (NTK) theory, avoiding the influence of manual parameter tuning on the training results. The core idea is to dynamically adjust the weights based on the current gradient sensitivity of each loss term—loss terms with smaller gradient magnitudes automatically receive higher weights, while those with larger gradient magnitudes automatically receive lower weights, achieving dynamic self-balancing during training. The model training employs a two-stage optimization strategy, with specific parameter configurations shown in Table 1.

[0040] Table 1 Simultaneously, a Curriculum Learning strategy is employed. In the early stages of training, the physical loss weights are linearly warmed up, and the strength of physical constraints is gradually increased. This prevents the network from falling into local minima due to strong physical constraints before it has developed basic fitting capabilities. Specifically, during the warm-up phase of the first 2000 iterations, the physical loss weights linearly increase from 0 to the maximum allowable value given by the NTK mechanism, allowing the network to prioritize data fitting. Physical constraints are then gradually introduced after the network has developed preliminary prediction capabilities, effectively ensuring training convergence and stability.

[0041] Based on the above design, the training process of this module can be summarized as follows: First, initialize the network parameters and Fourier frequency projection matrix. Then, in each training iteration, perform the following sequentially: sample mini-batches from the measured dataset, physical collocation set, and boundary condition set; obtain multi-physics predictions through Fourier embedding and forward computation of the backbone network; calculate the residuals of the two types of equations through automatic differentiation; calculate four types of loss terms; update the loss weights every 100 iterations based on the NTK mechanism and adjust the physical loss weights according to the course learning strategy; obtain the total loss through weighted summation; use the Adam optimizer for the first 10,000 iterations, and then switch to the L-BFGS optimizer to update the network parameters; terminate training when the convergence criterion or the maximum number of iterations is reached. The trained Siamese model can then be used for real-time inference, providing a high-fidelity state prediction foundation for subsequent residual feedback evolution and hierarchical safety assessment.

[0042] The twin residual-driven state derivation module constructs a residual feedback closed loop between the physical entity and the twin virtual model. Its working process consists of four sequentially connected stages: real-time residual calculation, cumulative threshold judgment, online parameter identification, and parameter back-injection update. Specifically, it includes an input embedding layer, a deep fully connected backbone network, and a physical constraint calculation layer connected in series.

[0043] The details are as follows: In the real-time residual calculation stage, this module calculates the multidimensional twin residuals between sensor observations and twin model predictions in real time, according to a cycle synchronized with the twin model inference. The residuals are defined as follows: in For the sensor in Multidimensional measured vectors (including displacement, pressure, and temperature) are collected in real time. This represents the prediction vector of the twin model at the same time point. To eliminate single-point noise interference, this module further introduces a sliding window cumulative residual, that is, taking the root mean square value of the residual norm within the most recent time window as the cumulative residual evaluation metric to reflect the degree of steady-state deviation between the twin model and the physical entity.

[0044] In the cumulative threshold judgment stage, this module presets residual trigger thresholds for different physical quantities such as displacement and pressure. The thresholds are calibrated based on sensor accuracy, allowable construction deviations, and engineering experience. When the cumulative residual exceeds the corresponding threshold, it is determined that there is a working condition drift or parameter mismatch between the twin model and the physical entity, and the online parameter identification program is immediately triggered. To avoid false triggering caused by instantaneous sensor jitter, this module also sets a criterion for the number of consecutive threshold exceedances, that is, identification is only formally triggered when the residual exceeds the threshold for multiple consecutive sampling periods.

[0045] In the online parameter identification stage, this module uses Recursive Least Squares with Forgetting Factor (RLS) to identify key physical parameters in the hydraulic system in real time, such as damping coefficient, equivalent stiffness, leakage coefficient, and bulk modulus of hydraulic fluid. Specifically, the hydraulic cylinder dynamics equations and fluid continuity equations are first rewritten into a vector of parameters to be identified. The standard linear regression form for the unknowns is then used, and the solution is obtained online based on the recursive update rule: in for Parameter identification results at time points. and These are the observed variables and regression vectors constructed from the measured states, respectively. This is the gain vector calculated recursively using the covariance matrix. The algorithm introduces a forgetting factor mechanism, giving higher weight to recent data, thus enabling rapid response to operating condition drift; the entire identification process is completed within seconds, without affecting the real-time performance of the system's main control loop.

[0046] In the parameter back-injection and update phase, after identification, the new parameters, after undergoing rationality verification (comparing whether the changes before and after identification are within the physically reliable range), are injected into the equation coefficients of the physical information neural network through the parameter hot-replacement interface. This updates the specific forms of the residuals of the dynamic equations and the continuity equations, enabling the twin model to adaptively track and evolve to the current real state of the physical entity. The back-injection process only replaces the scalar coefficients in the physical equations and does not trigger a full retraining of the network weights, thus avoiding impacting the real-time performance of online inference. Through the above four-stage closed-loop feedback mechanism, this module can effectively overcome the mapping distortion problem caused by long-term operating condition drift due to factors such as seal aging, oil contamination, and load changes in traditional static twin models, ensuring that the twin model maintains a high-fidelity state throughout the entire equipment lifecycle.

[0047] The graded safety assessment module calculates three core safety indicators—displacement synchronization error, pressure equalization error, and twin residual error—based on the multi-cylinder state vector output by the physical information neural network twin model. Among these, displacement synchronization error... The degree of displacement deviation between cylinders is reflected by the normalized standard deviation of the real-time displacement of each cylinder relative to the average displacement of multiple cylinders. in The total number of hydraulic cylinders. For the first Measured displacement of the hydraulic cylinder This is the average displacement of multiple cylinders. To determine the maximum allowable synchronous deviation during construction, this embodiment uses 5 mm.

[0048] Pressure equalization error It reflects the uniformity of load distribution in multiple cylinders and is defined by the coefficient of variation (CV) of the working pressure of each cylinder, which is the ratio of the standard deviation of the pressure of multiple cylinders to the pressure mean. This index is dimensionless and insensitive to the pressure magnitude, making it suitable for horizontal comparison under different working conditions.

[0049] Twin residual error Reflecting the degree to which the current state of a physical entity deviates from a healthy baseline, the cumulative residuals output by the aforementioned twin residual extrapolation module are directly used via a sliding window. After normalization, the normalization benchmark is the residual trigger threshold. ,make The dimensions are consistent with the other two types of errors.

[0050] This module uses the analytic hierarchy process (AHP) to weight and fuse three types of error indicators to obtain a comprehensive safety index. Based on the safety and danger thresholds set in the construction safety specifications, it completes a three-level judgment of the system's operating status: Under normal operating conditions, the system's multi-cylinder synchronization, load balance, and twin mapping stability allow for continuous construction; under the status warning state, if an error indicator exceeds the safety threshold but does not reach the danger level, the system automatically prompts maintenance personnel to check the corresponding cylinder or hydraulic circuit; under the high-risk shutdown state, if the error indicator exceeds the danger threshold, the system immediately triggers the safety protection program, and the linkage control terminal executes the shutdown command to prevent structural eccentric loading, attitude instability, or sudden component failure.

[0051] The B / S visualization and interaction module is deployed on the display terminal of the data monitoring center, providing a human-machine interface based on a browser-server architecture. This module can display multi-dimensional monitoring information in real time, including multi-cylinder displacement synchronization curves, pressure balance bar charts, oil temperature trend curves, and twin residual heat maps. It also uses a 3D virtual model to reproduce the attitude state of multiple cylinders and the lifted components in real time, dynamically indicating the current safety level through color coding and audible / visual alarms. The module also supports full-cycle storage, querying, playback, and export of construction process data, facilitating post-event analysis and accident tracing. Furthermore, it allows maintenance personnel to adjust sensor acquisition parameters, modify safety thresholds, and view equipment operation logs through the interface, enabling remote interactive management and control.

[0052] The above system solution is illustrated below using a construction scenario of synchronous lifting of a large steel structure roof by four hydraulic cylinders as an example. This example does not constitute a limitation on the scope of protection of this invention.

[0053] In this embodiment, four lifting cylinders with a rated tonnage of 350 tons are configured on-site during synchronous construction, and are respectively arranged at the four lifting points of the steel structure roof. The multi-cylinder synchronous equipment digital twin safety monitoring system of this invention is used for online monitoring of the entire process. Figure 1 As shown, the system of the present invention consists of a multi-source sensor acquisition module, a physical-data dual-drive twin modeling module, a twin residual-driven state deduction module, a hierarchical security assessment module, and a B / S visualization interaction module, which are connected in sequence. The five modules are arranged hierarchically from bottom to top according to the data flow direction. Data uplink transmission and control command downlink response are realized through industrial Ethernet, forming a complete closed-loop monitoring system of "perception-modeling-deduction-assessment-interaction".

[0054] During the system hardware deployment phase, high-precision laser displacement sensors are installed at the piston rod ends of the four lifting cylinders, with lenses aimed at the component lifting points, to collect the lifting displacement of each cylinder in real time. High-range oil pressure sensors are installed at the oil circuit interfaces of the rodless and rod chambers of each cylinder to collect the working pressure of the two chambers in real time. Platinum resistance temperature sensors are installed on the main oil outlet of the hydraulic station to collect the oil temperature in real time. Gear flow sensors are installed on the main oil inlet of each cylinder to collect the oil flow rate in real time. Each cylinder is equipped with an industrial control computer as a local data acquisition terminal. The industrial control computer communicates stably with each sensor through an industrial bus to complete local data acquisition, caching, and preliminary packaging. The data monitoring center is equipped with a high-performance industrial server, equipped with a graphics processing unit accelerator card for training and real-time inference of the physical information neural network. The server deploys the physical-data dual-driven twin modeling module, the twin residual-driven state inference module, and the hierarchical safety assessment module of this invention, and builds a time-series database to store multi-cylinder acquired data, twin prediction results, error indicators, and safety assessment records. The monitoring center is also equipped with an industrial-grade touch display terminal as a human-machine interface, deploys a B / S visualization monitoring system, and builds an industrial Ethernet network to achieve stable communication between the industrial control computers of each hydraulic cylinder and the central server.

[0055] In the system initialization and twin model pre-training phase, the basic physical parameters of the steel structure roof, including geometric dimensions, mass distribution, lifting cylinder model parameters, stroke range, and rated pressure, are first input into the server to construct a geometric skeleton that maps one-to-one with the physical system. Then, the hyperparameters of the physical information neural network are configured, with the network backbone set to an eight-layer fully connected structure, each layer containing 128 neurons. The activation function is the hyperbolic tangent function, and the input layer uses a Fourier feature embedding layer for high-frequency encoding. The physical loss term incorporates the hydraulic cylinder dynamics equation and the fluid continuity equation, and the residuals are solved backwards from the network output using automatic differentiation technology. The training data consists of three parts: first, multi-condition simulation data generated by hydraulic system simulation software for initial training; second, historical operating data from similar lifting projects undertaken by the construction company for condition adaptation; and third, typical fault case data for abnormal state learning. The training process employs a two-stage strategy. The first stage uses an adaptive moment estimation optimizer with a high learning rate for coarse tuning through tens of thousands of iterations. The second stage switches to a quasi-Newton optimizer for fine tuning until convergence. Throughout the training process, the physical loss weights are gradually increased based on a curriculum learning strategy to prevent the model from getting trapped in undesirable local minima of physical constraints in the early stages of training. The dynamic balancing of the loss term weights is adaptively adjusted based on the neural tangent kernel theory.

[0056] During the system operation parameter configuration phase, the sensor sampling frequency was set to 100 times per second, the industrial control computer data preprocessing cycle to 10 milliseconds, and the server twin inference cycle to 100 milliseconds. Based on the safety specifications for steel structure roof construction, the allowable displacement synchronization deviation was set to no more than 5 mm, and the allowable pressure deviation to no more than 10% of the rated pressure. A judgment matrix was constructed based on the analytic hierarchy process (AHP), determining the weights for displacement synchronization error (0.5), pressure equalization error (0.4), and twin residual error (0.1). Safety and danger thresholds for the comprehensive safety indicators were also set. The twin residual trigger threshold for the residual feedback evolution module was set based on a combination of sensor accuracy and allowable construction deviations to ensure a balance between sensitivity and stability. Finally, the accuracy of each sensor was calibrated, and the clock references of the industrial control computers for each cylinder were unified to ensure that the timestamp synchronization error did not exceed one millisecond.

[0057] During the construction and operation phase, after the lifting operation is initiated, the industrial control computer of each hydraulic cylinder synchronously collects displacement, pressure, temperature, and flow data at a frequency of 100 times per second through sensors. After Kalman filtering for noise reduction and outlier removal, the data is uploaded to the data monitoring center server with a unified timestamp. The physical-data dual-drive twin modeling module on the server side receives the measured data, inputs the data into the pre-trained physical information neural network, and outputs real-time high-fidelity prediction results of the displacement field, pressure field, and temperature field of the four hydraulic cylinders. This synchronously drives the three-dimensional virtual model to complete the attitude mapping, so that the extension and retraction state of the virtual hydraulic cylinders and the lifting posture of the virtual roof are kept in real time synchronized with the physical entity.

[0058] like Figure 2 As shown, the state deduction module driven by the twin residual of this invention constructs a complete closed-loop feedback evolution mechanism, and its workflow is as follows: The physical hydraulic cylinder system outputs measured observation values ​​through sensors, and the PINN digital twin model outputs corresponding predicted values ​​based on the current network parameters. The difference is calculated to obtain the real-time twin residual. The system continuously monitors and accumulates the residual. When the accumulated residual of a certain cylinder exceeds a preset threshold, the module determines that the cylinder has experienced operating condition drift or physical parameter mismatch, and immediately triggers the recursive least squares parameter identification program to identify key physical parameters such as the damping coefficient, equivalent stiffness, and leakage coefficient of the cylinder in real time. The identification results are fed back into the equation coefficients of the physical information neural network, enabling the twin model to adaptively track the current real state. For example, in the middle and later stages of operation in this embodiment, the hydraulic system oil temperature rises, causing the oil viscosity to decrease. The twin residual module successfully captures the slight drift of the cylinder leakage coefficient. Through parameter identification and reinjection, the twin model's prediction accuracy of cylinder pressure is always maintained above 95%, effectively avoiding the mapping distortion problem of traditional static twin models under long-term operation.

[0059] like Figure 3As shown, the hierarchical safety assessment module of this invention adopts a top-down three-layer structure to complete the safety status determination. Based on the multi-cylinder state vector output by the twin model, the hierarchical safety assessment module calculates the displacement synchronization error, pressure equalization error, and twin residual error of the four cylinders in real time. The comprehensive safety index is obtained by weight fusion using the analytic hierarchy process (AHP) and the safety level is determined accordingly. When the comprehensive safety index is below the safety threshold, the system is determined to be in normal operation, and the B / S interface displays a green safety indicator and continues to monitor. When the comprehensive safety index exceeds the safety threshold but does not reach the danger threshold, the system is determined to be in a state warning state, and the abnormal cylinder is highlighted on the interface, prompting maintenance personnel to check the corresponding hydraulic circuit or sensor. When the comprehensive safety index exceeds the danger threshold, the system is determined to be in a high-risk shutdown state, immediately triggering the audible and visual alarm of the monitoring center, and issuing an emergency shutdown command to the lifting control system via industrial Ethernet to prevent structural overloading or attitude instability accidents.

[0060] During the visualization and construction archiving phase, the B / S visualization monitoring module displays in real-time the displacement synchronization curves, pressure equilibrium bar charts, oil temperature trend curves, twin residual heat maps, and a 3D virtual model of the multiple cylinders and steel structure roof, allowing maintenance personnel to intuitively grasp the status of the entire construction process. All collected data, twin prediction results, error indicators, safety assessment records, and early warning events during construction are automatically stored in a time-series database. After construction is completed, the system automatically archives and generates a construction monitoring report, supporting subsequent construction deviation analysis, scheme optimization, and accident tracing. Maintenance personnel can remotely adjust sensor sampling parameters, modify safety thresholds, and view equipment operation logs through the B / S interface, enabling refined system operation and maintenance management without needing to visit the construction site.

[0061] It should be noted that the embodiments of the present invention are not limited to the above-mentioned four-cylinder synchronous lifting scenario. The present invention is also applicable to other multi-cylinder collaborative construction scenarios such as hydraulic synchronous jacking, multi-cylinder linkage support, and synchronous lifting of offshore platforms. Without departing from the technical concept of the present invention, those skilled in the art can make adaptive adjustments to the system module configuration, network structure parameters, safety threshold values, etc., and all such modifications fall within the protection scope of the present invention.

[0062] This embodiment also provides a digital twin safety monitoring method for multi-cylinder synchronous equipment based on the above system, including the following steps: S1: Real-time acquisition of data on the extension displacement, working pressure, oil temperature and flow rate of each cylinder in the multi-cylinder synchronous construction site through the multi-source sensor acquisition module, and completion of spatiotemporal alignment; S2: Based on historical operating data and the physical equations of the hydraulic system, offline pre-training of the physical information neural network twin model is completed through joint optimization training of data fitting loss, physical residual loss, boundary condition loss and multi-cylinder coupling constraint loss. S3: Input the real-time collected data into the twin model and output the real-time high-fidelity prediction of the displacement field, pressure field and temperature field of multiple cylinders, and complete the real-time synchronous mapping between the physical entity and the virtual model. S4: Calculate the twin residuals in real time and monitor their cumulative value. When the value exceeds the preset threshold, trigger online parameter identification of recursive least squares to achieve adaptive evolution of the twin model. S5: Based on the twin model, the displacement synchronization error, pressure balance error and twin residual error are calculated. The comprehensive safety index is obtained by weighted fusion through the analytic hierarchy process, and the three-level safety level is determined and graded early warning and linkage protection are executed. S6: Displays twin mapping results, security levels, and early warning information through a B / S visual interactive module, supporting historical data tracing and remote parameter adjustment.

[0063] In step S2, the pre-training of the physical information neural network twin model adopts a two-stage training strategy. In the first stage, an adaptive moment estimation optimizer is used for coarse tuning with a high learning rate. In the second stage, a quasi-Newton optimizer is switched to fine tuning until convergence. During the training process, a course learning strategy is used to gradually increase the weight of physical loss to avoid getting trapped in local minima of physical constraints in the early stage of training. The monitoring method is applicable to multi-cylinder collaborative construction scenarios such as synchronous lifting of large steel structures, synchronous jacking of ultra-large bridges, and multi-cylinder linkage support of offshore platforms.

[0064] It should be noted that the specific content and beneficial effects of the method in this application can be found in the above system embodiments, and will not be repeated here.

[0065] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A physical information neural network-based multi-oil cylinder synchronous equipment digital twin safety monitoring system, characterized in that, It includes a multi-source sensing acquisition module, a physical-data dual-driven twin modeling module, a twin residual-driven state deduction module, and a graded security assessment module; The multi-source sensor acquisition module is deployed at the end of each cylinder in the multi-cylinder synchronous construction site to collect data on the extension displacement, working pressure, oil temperature and flow rate of each cylinder in real time. The physical-data dual-drive twin modeling module constructs a multi-cylinder digital twin model based on a physical information neural network. This multi-cylinder digital twin model embeds the physical equations of the hydraulic synchronization system as soft constraints into the loss function of the physical information neural network. The soft constraints include physical residual loss and multi-cylinder coupling constraint loss. The physical residual loss is obtained by back-solving the dynamic equation residuals of the hydraulic cylinder rigid body and the continuity equation residuals of the hydraulic cavity fluid based on the output of the physical information neural network, and combining them. The multi-cylinder coupling constraint loss includes a constraint term to force the displacement of each cylinder to remain synchronous and consistent, and a constraint term to force the sum of the bearing capacities of each cylinder to equal the total weight of the component. The twin residual-driven state deduction module is used to calculate the twin residual between the predicted value of the multi-cylinder digital twin model and the measured value of the corresponding physical quantity, so as to trigger parameter identification and update the scalar coefficients of the physical equation in the physical information neural network. The graded safety assessment module is used to calculate a comprehensive safety index based on the multi-cylinder state vector output by the multi-cylinder digital twin model, so as to assess the safety status of the multi-cylinder synchronous construction site.

2. The physical information neural network-based multi-cylinder synchronous equipment digital twin safety monitoring system according to claim 1, wherein The physical equations include the hydraulic cylinder rigid body dynamics equation for calculating the residuals of the dynamic equations, and the hydraulic cavity fluid continuity equation for calculating the residuals of the continuity equations. The physical parameters covered by the rigid body dynamics equation of the hydraulic cylinder include: equivalent mass of the cylinder piston, damping coefficient, equivalent stiffness, effective working area of ​​the two chambers, and external load force. The physical parameters covered by the fluid continuity equation of the hydraulic cavity include: cavity volume, bulk modulus of the oil, instantaneous flow rate, and leakage coefficient. The physical residual loss is obtained by normalizing the residuals of the dynamic equation and the residuals of the continuity equation, and then performing a mean-square weighted summation. The residuals of the dynamic equations are normalized to the square of the rated load force, and the residuals of the continuity equations are normalized to the square of the rated pressure.

3. The physical information neural network-based multi-ram synchronous equipment digital twin safety monitoring system according to claim 1, wherein, The calculation expression for the multi-cylinder coupling constraint loss is as follows: In the formula, is the multi-cylinder coupling constraint loss, is the multi-oil-cylinder digital twin model parameter, is the total number of oil cylinders, and the superscript represents the th oil cylinder, and are the displacement and pressure prediction values of the th oil cylinder, respectively, is the displacement value of the th oil cylinder, is the effective action area of the oil cylinder rodless cavity, is the total gravity of the lifted component, is the number of time sampling points, is the relative weight of the load conservation term.

4. The digital twin safety monitoring system for multi-cylinder synchronous equipment based on physical information neural network according to claim 1, characterized in that, The loss function of the physical information neural network also includes data fitting loss and boundary condition loss; The data fitting loss is the mean square error of the predicted cylinder state output by the multi-cylinder digital twin model and the corresponding measured value output by the multi-source sensor acquisition module over all sampling points. The boundary condition loss includes an initial condition part and a boundary condition part. The initial condition part corresponds to the initial displacement, initial pressure and initial temperature before the cylinder is lifted. The boundary condition part corresponds to the oil inlet flow boundary and pressure boundary of the hydraulic pump station. The loss function of the physical information neural network is a weighted sum of data fitting loss, physical residual loss, boundary condition loss, and multi-cylinder coupling constraint loss.

5. The digital twin safety monitoring system for multi-cylinder synchronous equipment based on physical information neural network according to claim 1, characterized in that, The multi-source sensing acquisition module includes a laser displacement sensor, an oil pressure sensor, a temperature sensor, and a flow sensor. The laser displacement sensor is installed at the end of the piston rod of each oil cylinder or at the lifting point of the component to collect the extension displacement of the oil cylinder in real time, so as to reflect the lifting or pushing displacement state of the component. The hydraulic pressure sensors are installed at the oil circuit interfaces of the rodless and rod chambers of each hydraulic cylinder to collect the working pressure of the hydraulic cylinder in real time, so as to reflect the load distribution status. The temperature sensor is installed in the oil circuit of the cylinder or at the oil outlet of the hydraulic station to collect the oil temperature in real time, so as to help evaluate the working status of the hydraulic system. The flow sensor is installed in the main oil inlet of each cylinder to collect the oil flow rate in real time, so as to assist in solving the physical equation residuals. The data collected by the laser displacement sensor, oil pressure sensor, temperature sensor, and flow sensor are calibrated with a unified timestamp to achieve spatiotemporal alignment for subsequent processing.

6. The digital twin safety monitoring system for multi-cylinder synchronous equipment based on physical information neural network according to claim 1, characterized in that, The physical information neural network includes an input embedding layer, a deep fully connected backbone network, and a physical constraint calculation layer connected in series. The input embedding layer is used to encode the spatiotemporal coordinates of the input in a high dimension using Fourier feature embedding technology. The spatiotemporal coordinates include time coordinates and spatial coordinates, and the spatial coordinates correspond to the cylinder number and piston position. The deep fully connected backbone network is used to learn the spatiotemporal evolution of the displacement field, pressure field, and temperature field of multiple hydraulic cylinders. The physical constraint calculation layer is used to solve the dynamic equation residuals and continuity equation residuals of the hydraulic system's physical equations from the output of the neural network based on automatic differentiation technology.

7. The digital twin safety monitoring system for multi-cylinder synchronous equipment based on physical information neural network according to claim 1, characterized in that, The twin residual-driven state deduction module includes a real-time residual calculation submodule, an accumulated threshold judgment submodule, an online parameter identification submodule, and a parameter back-injection and update submodule. The residual real-time calculation submodule is used to calculate the twin residual between the predicted value of the multi-cylinder digital twin model and the measured value of the corresponding physical quantity in real time; The cumulative threshold judgment submodule is used to calculate the root mean square value of the twin residual norm output by the real-time residual calculation submodule under each sliding window, and use it as the cumulative residual value. The online parameter identification submodule is used to trigger when the cumulative residual value exceeds a preset threshold, and to identify the damping coefficient, equivalent stiffness, leakage coefficient and oil volumetric elastic modulus online based on the recursive least squares method. The parameter back-injection update submodule is used to back-inject the parameter identification results output by the online parameter identification submodule into the scalar coefficients of the physical equations in the physical information neural network.

8. The digital twin safety monitoring system for multi-cylinder synchronous equipment based on physical information neural network according to claim 1, characterized in that, The process of calculating the comprehensive safety index by the graded safety assessment module includes: calculating the displacement synchronization error, pressure equalization error and twin residual error based on the multi-cylinder state vector output by the multi-cylinder digital twin model, and then performing weighted fusion to obtain the comprehensive safety index. The displacement synchronization error is the normalized standard deviation of the real-time displacement of each cylinder relative to the average displacement of multiple cylinders. The pressure equalization error is the ratio of the standard deviation of the multi-cylinder pressure to the pressure mean. The twin residual error is the normalized result of the cumulative residual value of the twin residual output by the state inference module driven by the twin residual, and the corresponding normalization benchmark is the residual trigger threshold. The assessment of the safety status of the multi-cylinder synchronous construction site includes: comparing the comprehensive safety index with the preset safety threshold and danger threshold; if the comprehensive safety index is not higher than the safety threshold, it is in normal operation status; if the comprehensive safety index is higher than the safety threshold but not higher than the danger threshold, it is in status warning status; if the comprehensive safety index is higher than the danger threshold, it is in high-risk shutdown status.

9. A digital twin safety monitoring system for multi-cylinder synchronous equipment based on physical information neural networks according to claim 1, characterized in that, The system also includes a B / S visualization interaction module, which is used to display multi-cylinder displacement synchronization curves, pressure balance bar charts, oil temperature trend curves, twin residual heat maps, real-time mapping of three-dimensional virtual models, safety level color and sound and light prompts, historical data storage, query and playback, and remote adjustment of sensor parameters and safety thresholds.

10. A digital twin safety monitoring method for multi-cylinder synchronous equipment based on physical information neural networks, characterized in that, Includes the following steps: Real-time acquisition of data on the extension displacement, working pressure, oil temperature, and flow rate of each hydraulic cylinder during synchronous construction of multiple hydraulic cylinders; The collected data is input into a multi-cylinder digital twin model constructed based on a physical information neural network, and the predicted values ​​of the multi-cylinder state vectors are output to complete the real-time synchronous mapping between the physical entity and the virtual model. During the training process of the multi-cylinder digital twin model, the physical equations of the hydraulic synchronization system are embedded as soft constraints into the loss function of the physical information neural network, and offline pre-training is performed based on historical operating data. Calculate the twin residual between the predicted value of the multi-cylinder digital twin model and the measured value of the corresponding physical quantity to trigger parameter identification and update the scalar coefficients of the physical equation in the physical information neural network; The comprehensive safety index is calculated based on the multi-cylinder state vector output by the multi-cylinder digital twin model to assess the safety status of the multi-cylinder synchronous construction site.

Images