Intelligent equipment full-life reliability evaluation method and system based on AI dynamic update

Abstract

This invention relates to the field of intelligent equipment condition monitoring and health management, specifically to an AI-based dynamically updated method and system for full-lifecycle reliability assessment of intelligent equipment. The method first constructs a physical information neural network model embedding specific physical equations for key components of the equipment. Low-dimensional health features are extracted from high-dimensional sensor data using an autoencoder, and Gaussian mixture model clustering is used for condition perception. When a new condition is identified, the model is proactively updated. The update process employs an incremental learning strategy that combines fixed elastic weights with experience replay, effectively avoiding catastrophic forgetting. The updated surrogate model is used to predict the system response, construct a performance function, and calculate the failure probability and reliability in real time using the Monte Carlo method, achieving risk warning. This invention enables real-time, dynamic, and high-precision full-lifecycle reliability assessment of intelligent equipment under complex operating conditions without disassembling the equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent equipment condition monitoring and health management. Specifically, it relates to a method and system for dynamically updating a physical information neural network model using artificial intelligence technology for complex intelligent autonomous operating equipment such as tunnel boring machines, and for conducting real-time reliability assessments throughout the entire life cycle based on this model. Background Technology

[0002] Full-face tunnel boring machines (MTBF) are intelligent, autonomous operating equipment integrating mechanics, hydraulics, electrical systems, sensing, and control. Their key components, such as main bearings, propulsion hydraulic cylinders, articulated bearings, and cutting teeth, endure high loads, strong impacts, and harsh environments during tunneling. Wear and malfunctions directly affect construction safety and efficiency. However, the health status of these components is difficult to measure directly in real-time underground: internal damage to main bearings requires disassembly for inspection; internal leakage in hydraulic cylinders lacks effective online methods; and wear on articulated joints and cutting teeth is difficult to perform and infrequently due to structural enclosure, rock debris coverage, and safety space limitations, making traditional inspection methods challenging.

[0003] Existing reliability assessment methods are mostly based on probability statistics and historical failure data, which have significant limitations: First, they rely on a large amount of failure data, making them unsuitable for new equipment or small sample scenarios; second, they are mostly based on static assumptions, making it difficult to reflect reliability degradation under dynamic and time-varying operating conditions; third, they ignore the coupling effects and physical mechanisms between components, resulting in significant deviations between the assessment results and actual conditions; and fourth, they cannot utilize massive amounts of real-time operational data for adaptive model updates. Therefore, there is an urgent need for a new reliability assessment method that can integrate physical mechanisms, real-time data, and artificial intelligence, and can dynamically evolve throughout the entire life cycle. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for assessing the full life-cycle reliability of intelligent equipment based on AI dynamic updates, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for assessing the full life-cycle reliability of intelligent equipment based on AI-dynamic updates includes the following steps:

[0007] S1. Construction of Component-Specific PINN Models: For multiple core subsystems of intelligent autonomous operating equipment, physical information neural network models embedded with corresponding physical control equations are constructed respectively; the PINN model takes spatial coordinates, time coordinates and operating condition random variables as inputs, and outputs the predicted value of the key health status of the corresponding component.

[0008] S2. Online data feature extraction: Real-time acquisition of multi-source sensor data during equipment operation, and dimensionality reduction processing of the data using a pre-trained autoencoder to extract low-dimensional feature vectors characterizing the health status of components.

[0009] S3. Working Condition Perception and Update Trigger: Input the low-dimensional feature vector into the Gaussian mixture model clusterer and calculate its Mahalanobis distance with the existing working condition cluster centers; if the Mahalanobis distance exceeds a preset threshold, it is determined that a new working condition has been entered, and the update process of the corresponding subsystem PINN model is triggered.

[0010] S4. Anti-forgetting incremental learning update: When an update is triggered, the PINN model is updated using an incremental learning strategy; the loss function of the incremental learning adds a new data prediction error term and an elastic weight fixed regularization term to the original physical constraint loss, and introduces historical key samples for experience replay in order to absorb new knowledge and retain historical knowledge.

[0011] S5. Online Reliability Assessment and Early Warning: The updated PINN model is used to predict the system response over a future period, and a performance function is constructed to measure the distance between the system state and the failure boundary. The Monte Carlo method is used to sample the input random variables, and the failure probability and reliability of the system within a specified lifespan are calculated based on the performance function. When the instantaneous failure rate calculated based on the reliability exceeds a preset threshold, the corresponding component is determined to be in a high-risk state and an early warning is triggered.

[0012] As a further aspect of the present invention: in step S1, the core subsystem includes at least a main bearing, a propulsion hydraulic cylinder, and a cutting gear; the physical control equation embedded in the PINN model constructed for the main bearing is a vibration response equation: The PINN model constructed for the propulsion hydraulic cylinder is embedded with the correlation between oil pressure pulsation, seal wear and internal leakage; the PINN model constructed for the cutting teeth is embedded with the mapping relationship between rock hardness, impact stress and chipping / wear.

[0013] As a further aspect of the present invention: in step S1, the training of the PINN model employs a composite loss function: ;in, For physical residual loss, For boundary condition loss, For the initial conditional loss, λ f , λ BC , λ IC These are the weighting coefficients determined through cross-validation.

[0014] As a further aspect of the present invention: in step S4, the loss function for incremental learning is specifically:

[0015] ;

[0016] in, For the new data prediction error term, For its weight, The weights of the regularization term are fixed for the elastic weights. These are the model parameters for the current task. It is the optimal solution of parameters obtained after training the model. This represents the total number of parameters.

[0017] As a further aspect of the present invention: in step S5, the performance function is: ,in This represents the predicted system response by the PINN model. The failure threshold is a preset value; the failure probability is... The calculation formula is: ,in, Given the number of Monte Carlo samplings, the reliability R(T) = 1 − Pf(0, T), and the failure rate function. The formula for calculating (T) is: .

[0018] As a further aspect of the present invention, it also includes step S6: intelligent management of the model lifecycle; constructing a hierarchical management system, including:

[0019] The data perception layer continuously collects real-time operating condition data and metadata of each PINN model;

[0020] The analysis and decision-making layer calculates the health of the model based on the data, analyzes the evolution trend of the working condition, monitors computing resources, and outputs decision instructions.

[0021] The execution control layer executes dynamic management strategies, ranging from routine incremental maintenance to active network structure optimization and cross-model collaborative training, based on the decision instructions.

[0022] An AI-based dynamically updated intelligent equipment lifecycle reliability assessment system is provided to implement the AI-based dynamically updated intelligent equipment lifecycle reliability assessment method. The system includes:

[0023] The model building and deployment module is used to build, initially verify, and deploy the PINN models of each core subsystem to the equipment edge computing terminal.

[0024] The data acquisition and feature extraction module includes a sensor array, a data acquisition circuit, and an automatic encoder deployed on various components, used to acquire and extract health feature vectors in real time.

[0025] The working condition perception and update control module integrates a GMM clusterer to determine the working condition and trigger model update instructions;

[0026] The incremental learning update engine is used to execute an incremental learning algorithm that combines EWC and experience replay to update the specified PINN model after receiving an update instruction.

[0027] The reliability assessment and early warning module is used to call the updated PINN model to perform response prediction, failure probability calculation and risk warning.

[0028] The model lifecycle management module is used to monitor the status of each model, coordinate resources, and execute dynamic management strategies.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] 1. Deep integration of physical mechanisms and data: By constructing the PINN model with embedded physical equations, the limitations of pure data-driven models in terms of mechanism interpretability, extrapolation and small sample size are fundamentally overcome, laying a solid physical foundation for reliability assessment.

[0031] 2. Intelligent perception and proactive update: The working condition perception mechanism of "auto encoder + GMM clustering" is adopted to realize intelligent triggering of model update, so that the model update is synchronized with the change of working condition and avoids lag or over-update.

[0032] 3. Stable and efficient continuous learning: The incremental learning strategy that integrates EWC and experience replay ensures that the model retains the core historical patterns while absorbing new knowledge, achieving continuous improvement and stability of evaluation accuracy throughout the entire life cycle.

[0033] 4. Closed-loop management and autonomous optimization: The introduction of a model lifecycle intelligent management system enables closed-loop management of multi-model collaboration and adaptive optimization, improving the robustness and intelligence of the entire evaluation system.

[0034] 5. High engineering applicability: The entire method is based on online monitoring data and can achieve real-time and dynamic reliability assessment and early warning without stopping the machine for disassembly and inspection. It effectively ensures the safe, efficient and continuous operation of equipment such as TBMs and has great engineering application value. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the sensor installation arrangement deployed on a key component of the TBM in an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of the overall technical process of the evaluation method described in this invention.

[0037] Figure 3This is a schematic diagram of the basic structure of the physical information neural network used in this invention.

[0038] Figure 4 This is a schematic diagram of the model update triggering decision process based on GMM clustering in this invention.

[0039] Figure 5 This is a flowchart illustrating the construction and workflow of the intelligent management system for the model lifecycle in this invention. Detailed Implementation

[0040] The technical solution of this application will be further described in detail below with reference to specific embodiments.

[0041] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0042] Please see Figure 2 In one embodiment of the present invention, an AI-based dynamically updated intelligent equipment lifecycle reliability assessment method continuously injects new data and dynamically calibrates the physical constraint network during the equipment operation phase, thereby achieving rapid reassessment of system response and failure probability throughout the entire lifecycle. This provides real-time and evolvable reliability assurance for intelligent autonomous operating equipment under complex working conditions. Considering the complex working conditions of TBM during tunneling, such as sudden changes in strata, component coupling excitation, and a surge in sensor data, the present invention first constructs and embeds dedicated physical equations into PINN models for three important components: bearings, hydraulic cylinders, and cutting teeth. It uses an automatic encoder to extract health indicators highly correlated with cutting tooth wear from high-dimensional data such as vibration, temperature, and pressure. The extracted indicators are updated using GMM clustering to proactively ensure that the model does not evolve excessively or lag behind. Elastic weights are applied to the network parameters to consolidate incremental learning, which both absorbs new working condition knowledge and prevents catastrophic forgetting. Finally, the updated PINN proxy model is used for reliability analysis, and the system is judged to determine whether it has entered a high-risk state, thus realizing real-time reliability judgment of TBM under variable geological conditions.

[0043] The following are the specific steps for constructing the PINN model:

[0044] Step 1: Based on the TBM tunneling requirements, define specific physical constraints. For the three core components of the TBM—main bearing, propulsion hydraulic cylinder, and cutting gear—extract physical constraints directly related to their failure modes, rather than general physical equations:

[0045] Main bearing: Taking the coupling characteristics of rotating support and impact load during tunneling as the core, the vibration response relationship between cutterhead speed and tunneling thrust is selected as the physical constraint, the constraint boundary is defined as the rated tunneling parameter range of TBM, and the initial condition is set as the standard vibration baseline of component factory inspection.

[0046] For hydraulic cylinder propulsion: physical constraints are constructed based on the correlation between oil pressure pulsation, seal wear, and internal leakage. Boundary conditions are limited to the working pressure range of the hydraulic cylinder and the ambient temperature. The initial conditions are based on the stable oil pressure value after the installation of new seals.

[0047] Cutting teeth: Physical constraints are established based on the mapping relationship between rock hardness, impact stress, and chipping / wear. Boundary conditions are set according to the rock compressive strength range of the tunneling strata, and the initial condition is the stress wave reference value after the new cutting teeth are installed.

[0048] Step 2: Adapting the neural network structure design for TBM field deployment. Based on the computing power limitations and real-time requirements of the TBM on-board edge computing terminal, a lightweight neural network structure is designed with the following specific parameters:

[0049] (1) Input layer: The dimension is fixed at 6 dimensions, which directly corresponds to the core parameters collected by the TBM sensor, including spatial coordinates (radial / axial position of the cutterhead), time coordinates (tunneling time), and random variables (load fluctuation value, formation hardness deviation value), without the need for additional data conversion.

[0050] (2) Hidden layer: A lightweight 3-layer structure is adopted, with 64, 32 and 16 neurons in each layer respectively. The activation function is adapted to the edge computing for fast operation, avoiding the computational power consumption caused by complex activation functions.

[0051] (3) Output layer: Based on the component differentiation design, the peak value of vibration displacement output by the main bearing (unit: mm), the amplitude of oil pressure pulsation output by the hydraulic cylinder (unit: MPa), and the peak value of stress wave output by the cutting tooth (unit: MPa) directly correspond to the on-site fault judgment indicators.

[0052] Step 3: Taking into account the strong noise and large fluctuations in operating conditions of TBM tunneling data, construct a targeted composite loss function, as follows:

[0053] (1) Collect three sets of basic data (hard rock, soft rock, and normal wear conditions) during the TBM trial tunneling stage, extract the physical constraint residual benchmark values ​​of each component, and ensure that the loss function is adapted to the characteristics of the field data.

[0054] (2) Each component of the loss function is clearly adapted to the TBM scenario:

[0055] Physical residual loss Lf: weight =0.4, focusing on core failure-related parameters of key components, such as main bearing vibration displacement residual and hydraulic cylinder oil pressure residual).

[0056] Boundary condition loss L BC Weight =0.3, using the TBM's rated operating parameters as the boundary to avoid the model exceeding the actual operating range;

[0057] Initial condition loss L IC Weight =0.3, using the measured data after the new component is installed as the initial benchmark to ensure that the model fits the initial state of the equipment.

[0058] Loss function calibration: After 20 hours of trial tunneling with the TBM, 1,000 sets of valid data were collected. The weight parameters were adjusted using cross-validation to ensure that the convergence speed of the model under actual working conditions meets the real-time monitoring requirements.

[0059] Step 4: TBM Field Model Deployment and Initialization Verification

[0060] The completed PINN model was integrated into the TBM main control PLC system. The sensor data interface was connected via a shielded channel of a high-voltage slip ring to ensure data transmission latency ≤50ms. Afterwards, the initial verification process was initiated, and the TBM was run unloaded for 5 minutes to collect sensor baseline data and complete the initial parameter calibration of the model. A 1-hour trial tunneling was conducted, comparing the model's predicted values ​​with the sensor's measured values, requiring an error ≤5%. If this was not met, the loss function weights were adjusted immediately. After successful verification, the model parameters were stored on the edge computing terminal, generating a component-model mapping table for easy subsequent updates and maintenance.

[0061] The following is combined Figure 3 Taking bearings as an example, a PINN model of bearing vibration is established. For bearings, this invention uses the control equation of their vibration response as a physical constraint. This equation expresses the relationship between the bearing displacement response and external excitation and its own structural parameters. The specific form of the control equation is as follows:

[0062] ;

[0063] in: This refers to the vibration displacement response of a bearing at position x and time t, where m is the equivalent mass, c is the equivalent damping coefficient, and k is the equivalent stiffness coefficient. It is the external time-varying excitation force acting on the main bearing. It is a set of random variables that characterize the uncertainty of the excitation, and it can include distribution parameters such as load amplitude and frequency.

[0064] Once the governing equations are determined, the structure of the physical neural network model is constructed. The input layer of this network receives an input vector X consisting of spatial coordinates, temporal coordinates, and random variables.

[0065] ;

[0066] The network's output layer provides predicted values ​​for the main bearing's vibration displacement response. :

[0067] ;

[0068] In the expression, N is the nonlinear mapping function implemented by the neural network, and W and b are the weight matrix and bias vector of the entire network, respectively.

[0069] After completing the above construction, in order to make the predicted value To satisfy the initially constructed physical governing equations, calculations are required. Take the partial derivative with respect to t, and substitute it into the governing equation to construct the physical residual term f.

[0070] ;

[0071] Based on the above formula, a composite loss function θ for training network parameters is constructed. :

[0072] ;

[0073] In the formula For physical residual loss, For boundary condition loss, Loss due to initial conditions

[0074] λ f , λ BC , λ IC It is a parameter used to balance the contributions of various losses, and its specific value is generally determined through empirical setting or cross-validation.

[0075] After the PINN model is established, the pre-PINN model is integrated into the TBM's main control PLC system or edge computing terminal to achieve real-time data acquisition, real-time model inference, and finally real-time instruction output.

[0076] During tunneling operations, sensors deployed on various key components of the TBM begin collecting various signals generated during the tunneling process. Industrial sensors are installed in the main bearing, propulsion hydraulic cylinder, and cutting gear areas, such as... Figure 1As described above, a piezoelectric accelerometer is mounted radially and axially on the outer ring end face of the main bearing. A PT100 thermocouple with a metal sheath is embedded in the same ring to acquire data on cage impact vibration and overall temperature rise. A 4-20mA pressure transmitter with a standard threaded interface is used in the inlet and outlet oil chambers of the propulsion hydraulic cylinder. A small piezoelectric accelerometer is attached to the axial direction of the cylinder outer wall to simultaneously record oil pressure pulsation and piston chatter. For tooth cutting monitoring, a resonant acoustic emission probe is installed on the thread of the cutter head back plate, and a triaxial MEMS accelerometer is arranged next to the adjacent tooth seat to capture high-frequency stress waves and rock impact vibrations during tooth tip chipping. All signals from the sensors are transmitted centrally to the onboard edge computer through the shielded channel of the high-pressure slip ring. This forms raw data that can be directly used to extract features. This raw data contains… Monitoring information from various aspects, such as cutterhead vibration, propulsion system pressure, and bearing temperature, undergoes simple preprocessing before being fed into the core model. This preprocessing removes outliers and unifies data ranges. Then, the autoencoder intervenes, using a specific network structure and computation to quickly transform high-dimensional data into low-dimensional latent vectors. These latent vectors encapsulate the core features of the TBM's critical component operating status. The decoder takes these latent vectors as input and performs inverse computation using a network structure symmetrical to the encoder to restore the data dimensions and generate a reconstructed signal consistent with the original input signal dimensions. In actual tunneling operations, newly acquired sensor data is input into the encoder in real time and quickly transformed into low-dimensional latent vectors. These compressed feature vectors are then sent to the subsequent update triggering stage to determine whether the PINN model needs to be updated.

[0077] To ensure that the PINN proxy model of each subsystem maintains high accuracy throughout its entire lifecycle, this invention designs an incremental learning mechanism based on condition-aware updates and anti-forgetting. The update-triggered model used is Gaussian Mixture Model (GMM) clustering, such as... Figure 4 As shown, it is determined by the mean and standard deviation of a multidimensional Gaussian distribution. After the signal collected by the sensor is processed, GMM clustering is used to complete the condition perception and update trigger judgment, and then to determine whether AI incremental learning needs to be started to realize the dynamic calibration and accuracy maintenance of the PINN model. The specific process is as follows: The low-dimensional feature vector output by the encoder is used as the input data u of GMM clustering. This input data covers the main bearing vibration feature parameters, hydraulic cylinder oil pressure pulsation feature parameters, and cutting tooth stress wave feature parameters. For the input data u, according to the mean of the input parameters of the GMM model... Standard deviation Residuals of measured parameters Standard deviation The calculation formula is used to calculate the statistic of the i-th feature parameter at the N-th data point in real time. The residual between the j-th measured parameter and the corresponding PINN model prediction is calculated for the real-time acquired data u by comparing its value with the mean of the input parameters of the GMM model for each sub-region. of The formula for determining whether the current operating condition is covered by the existing GMM model is as follows:

[0078] ;

[0079] In the formula, k represents the number of GMM models established within the current sub-region. Threshold1 is the weighting factor for the noise sensitivity of the i-th input parameter, and refers to the minimum Mahalanobis distance under various working conditions such as hard rock tunneling, soft rock tunneling, normal tool wear, and tool chipping warning.

[0080] (1) If calculation <Threshold1, it is determined that the current operating condition has been covered by the existing GMM model, and the current PINN model will continue to be used for system response prediction and reliability assessment.

[0081] (2) If If the value is greater than or equal to Threshold1, then the current operating condition is determined to be a new operating condition, triggering the PINN model update process.

[0082] When the GMM model determines that the current operating condition is new, it triggers the PINN model update process. At this point, AI is used to incrementally update the PINN model. This is a learning strategy for continuously updated data, allowing the model to adjust in real time as new data arrives, avoiding retraining the entire network, thus saving computational resources and accelerating the training process. Incremental learning is crucial in complex systems such as physical information neural networks because these systems typically need to handle large amounts of real-time data streams and maintain high accuracy and efficiency with each data update. The specific update steps are as follows:

[0083] 1. Prepare incremental learning data and parameters

[0084] 1) Real-time acquisition of raw sensor data under new working conditions, including vibration, temperature, and pressure signals.

[0085] 2) Use the trained autoencoder to extract features from the new data and output a low-dimensional feature vector.

[0086] 3) Retrieve key samples similar to the current operating conditions from the historical sample library, especially those close to the failure boundary, and construct an experience playback dataset.

[0087] 4) Load the parameter file of the current PINN model and load the corresponding elastic weight fixed importance weight matrix, which records the importance score of each parameter in the historical tasks.

[0088] 2. Construct the incremental loss function

[0089] 1) Calculate the prediction error term for new data Where N is the number of samples, The model is for the input The predicted value, For input The true value.

[0090] 2) Based on the original loss function, a data error term and an EWC regularization term are added to form a composite loss function suitable for incremental learning: , in the formula For physical residual loss, For boundary condition loss, For the initial conditional loss, λ f , λ BC , λ IC It is a parameter used to balance the contributions of various losses. is the weight coefficient of the EWC regularization term. These are the model parameters for the current task. It is the optimal solution of parameters obtained after training the model.

[0091] 3) Set the initial values ​​for the hyperparameters. =1.0, =500, which can be fine-tuned based on performance on the validation set.

[0092] To avoid the loss of core knowledge from historical operating conditions due to indiscriminate updates of network parameters during incremental learning in the PINN model, such as the vibration control law of the main bearing, the feature mapping of hydraulic cylinder oil pressure pulsation, and the stress wave response relationship of the cutting teeth, this invention is based on the Elastic Weight Fixed (EWC) technique. It introduces a regularization term specifically designed to limit the updates of important parameters. The core idea of ​​EWC is to quantify the importance of each parameter to the task and add a penalty term to the loss function, so that the model can avoid excessive changes when adjusting these parameters, thereby protecting the learned knowledge. During incremental learning, the EWC regularization term can be implemented using the following formula:

[0093] ;

[0094] in These are the parameters for the current task. It is the optimal solution obtained after training on the previous task. It is a hyperparameter that controls the influence of the regularization term. M is the total number of model parameters. Through this regularization term, the model will try to maintain the stability of historical tasks during the update process and avoid excessive modification of parameters that are important to old tasks.

[0095] After completing the construction and incremental learning update of the surrogate model based on the physical information neural network, this invention utilizes the updated PINN model to predict the system response of key TBM components, including the main bearing, propulsion hydraulic cylinder, and cutting teeth, and further conducts reliability analysis. Based on the updated PINN surrogate model, this invention achieves reliability analysis through the following process: based on the system responses predicted by the PINN model, such as vibration displacement, hydraulic cylinder pressure pulsation, and cutting tooth stress waves, a performance function is constructed. It is used to measure the distance between the system state and the failure boundary, and its expression is:

[0096] ;

[0097] in This represents the predicted system response by the PINN model. The preset failure threshold,

[0098] when When <0, the system is in a failed state.

[0099] when When ≥0, the system is in normal working condition.

[0100] After completing the above calculations, the Monte Carlo method (MCS) is used to estimate the system failure probability, based on the joint distribution N of the input variables. MCS Within the generated group of random samples, the trained PINN model is used to quickly predict the failure probability of each sample over the entire time interval [0, T], and based on this prediction, it is determined whether the sample exceeds the failure boundary. The failure probability is calculated by the following formula:

[0101] ;

[0102] Lifespan Reliability within the range R(T) = 1− According to the failure rate function Formula for calculating (T)

[0103] ;

[0104] Set failure rate threshold max ,generally max The specific value is selected based on historical fault data. > max When a component is identified as being in a high-risk state, the system will trigger an early warning or maintenance decision.

[0105] Please see Figure 5 To implement the concept of full life cycle management and ensure that the PINN model maintains optimal performance throughout the entire equipment service life, this invention constructs a model life cycle intelligent management system. As the core management unit of the entire evaluation system, this system achieves collaborative management and autonomous optimization of multiple subsystem PINN models through multi-level monitoring, evaluation, and decision-making mechanisms. The system adopts a hierarchical decision-making architecture, including a data perception layer, an analysis and decision-making layer, and an execution control layer.

[0106] (1) Data perception layer: continuously collects data streams from two dimensions. The first dimension is the operating condition data stream, which is the real-time monitoring data from the sensors and the feature vectors extracted by the autoencoder. The second dimension is the model metadata stream, which includes the prediction confidence intervals of each PINN model, the time-series statistics of historical prediction errors, the convergence curve of the loss function during the incremental learning process, and the trajectory of model parameter changes.

[0107] (2) Analysis and Decision Layer: It has three built-in parallel analysis engines: The model health assessment engine calculates the real-time performance health score of each PINN model based on the model metadata. This score is composed of multiple sub-indicators, such as the rolling correlation coefficient between the prediction results and subsequent short-term measured data, the degree of recall decay of the model on the historical working condition sample set, and the entropy value of model parameter changes. Based on this, the model status is divided into four levels: excellent, good, attention, and intervention required. The working condition evolution analysis engine combines the current working condition characteristics with the historical working condition map to analyze whether the working condition change trend is gradual drift, periodic fluctuation or sudden leap, and predicts the types of working conditions that may occur in the future. The resource status monitoring engine monitors the CPU, memory and storage resource utilization of the edge computing platform in real time and evaluates the carrying capacity of the model update task. The execution control layer generates specific model management instructions and schedules the corresponding computing resources to execute based on the output of the analysis and decision layer.

[0108] (3) Execution control layer: Based on the output of the analysis and decision layer, it generates specific model management instructions and schedules corresponding computing resources to execute them.

[0109] After completing the above system architecture, the system implements dynamic and differentiated management strategies based on model health, operating condition evolution trends, and resource status. There are two main modes. The first is the routine maintenance mode. When the model health is excellent or good and the operating condition is stable, the system only maintains the basic incremental learning process, which is triggered by GMM clustering and combined with EWC fine-tuning. At the same time, the system periodically compresses and solidifies the key feature extraction capabilities of the online model into a lighter model for backup and rapid recovery in extreme cases. The second type is the proactive optimization mode. When the model's health drops to the level of concern, or when the working condition evolution analysis indicates that it is about to enter a completely new working condition region, the system will proactively intervene. When the model's learning saturation rate for the new working condition features is too fast, the system will initiate adaptive adjustments to the network structure. For example, it will dynamically add a hidden layer containing a specific activation function to the bearing PINN model to enhance its ability to capture new vibration frequency components. Subsequently, it will retrain using a mixture of historical key samples and new data. When the coupling effect between models increases, the system will temporarily construct a cross-subsystem joint training task, allowing the bearing PINN and the hydraulic cylinder PINN to share some intermediate layer features and perform collaborative fine-tuning to more accurately model the mutual influence between components.

[0110] This AI-based, dynamically updated intelligent equipment lifecycle reliability assessment method and system innovatively constructs a multi-subsystem PINN model with embedded proprietary physical equations. It integrates physical constraints such as vibration response control equations of key components like main bearings, propulsion hydraulic cylinders, and cutting teeth into a neural network. Combined with vibration, temperature, and pressure data collected by multi-dimensional sensors, it achieves accurate modeling of the operating status of core components under complex working conditions. This solves the problem that traditional methods struggle to capture the nonlinear response and failure mechanisms of components, providing a solid physical modeling foundation for reliability assessment.

[0111] This invention designs an autoencoder feature extraction + GMM clustering triggered working condition perception mechanism. The autoencoder extracts core health indicators from high-dimensional, noisy raw data, and then uses Mahalanobis distance of GMM clustering to determine the working condition coverage, realizing the active triggering of model updates. This avoids the waste of resources caused by excessive model evolution and prevents the evaluation bias caused by delayed updates, ensuring that the model adapts to the working condition changes throughout the TBM's life cycle.

[0112] This invention employs an incremental learning strategy that combines Elastic Weight Fixed (EWC) and Experience Replay. During model updates, the EWC regularization term limits significant changes in key parameters, while experience replay preserves failure boundary samples. This effectively avoids catastrophic forgetting, achieves efficient absorption of new operating condition knowledge and stable preservation of core patterns from historical operating conditions, and ensures the continuous stability of model evaluation accuracy throughout the entire lifecycle.

[0113] This invention constructs a complete technical system of multi-dimensional data acquisition, intelligent feature engineering, dynamic model updating, and real-time reliability assessment. Without disassembling and inspecting key TBM components, it can predict system response and failure probability in real time through online monitoring data and dynamically updated PINN proxy models. This effectively avoids the failure risks and project delays caused by disassembly and inspection, and provides real-time and evolvable reliability assurance for the efficient and stable operation of TBM equipment under complex geological conditions.

[0114] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these should also be considered within the scope of protection of the present invention. These will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for assessing the full life-cycle reliability of intelligent equipment based on AI dynamic updates, characterized in that, Includes the following steps: S1. Construction of Component-Specific PINN Models: For multiple core subsystems of intelligent autonomous operating equipment, physical information neural network models embedded with corresponding physical control equations are constructed respectively; the PINN model takes spatial coordinates, time coordinates and operating condition random variables as inputs, and outputs the predicted value of the key health status of the corresponding component. S2. Online data feature extraction: Real-time acquisition of multi-source sensor data during equipment operation, and dimensionality reduction processing of the data using a pre-trained autoencoder to extract low-dimensional feature vectors characterizing the health status of components. S3. Working Condition Perception and Update Trigger: Input the low-dimensional feature vector into the Gaussian mixture model clusterer and calculate its Mahalanobis distance with the existing working condition cluster centers; if the Mahalanobis distance exceeds a preset threshold, it is determined that a new working condition has been entered, and the update process of the corresponding subsystem PINN model is triggered. S4. Anti-forgetting incremental learning update: When an update is triggered, the PINN model is updated using an incremental learning strategy; the loss function of the incremental learning adds a new data prediction error term and an elastic weight fixed regularization term to the original physical constraint loss, and introduces historical key samples for experience replay in order to absorb new knowledge and retain historical knowledge. S5. Online Reliability Assessment and Early Warning: The updated PINN model is used to predict the system response over a future period, and a performance function is constructed to measure the distance between the system state and the failure boundary. The Monte Carlo method is used to sample the input random variables, and the failure probability and reliability of the system within a specified lifespan are calculated based on the performance function. When the instantaneous failure rate calculated based on the reliability exceeds a preset threshold, the corresponding component is determined to be in a high-risk state and an early warning is triggered.

2. The AI-based dynamic update-based intelligent equipment full-lifecycle reliability assessment method according to claim 1, characterized in that, In step S1, the core subsystem includes at least a main bearing, a propulsion hydraulic cylinder, and a cutting gear; the physical control equation embedded in the PINN model constructed for the main bearing is the vibration response equation: The PINN model constructed for the propulsion hydraulic cylinder is embedded with the correlation between oil pressure pulsation, seal wear and internal leakage; the PINN model constructed for the cutting teeth is embedded with the mapping relationship between rock hardness, impact stress and chipping / wear.

3. The AI-based dynamic update-based intelligent equipment full-lifecycle reliability assessment method according to claim 1 or 2, characterized in that, In step S1, the training of the PINN model uses a composite loss function: ;in, For physical residual loss, For boundary condition loss, For the initial conditional loss, λ f , λ BC , λ IC These are the weighting coefficients determined through cross-validation.

4. The AI-based dynamic update-based intelligent equipment full-lifecycle reliability assessment method according to claim 1, characterized in that, In step S4, the loss function for incremental learning is specifically as follows: ； in, For the new data prediction error term, For its weight, The weights of the regularization term are fixed for the elastic weights. These are the model parameters for the current task. It is the optimal solution of parameters obtained after training the model. This represents the total number of parameters.

5. The AI-based dynamic update-based intelligent equipment full-lifecycle reliability assessment method according to claim 1, characterized in that, In step S5, the performance function is: ,in This represents the predicted system response by the PINN model. The failure threshold is a preset value; the failure probability is... The calculation formula is: ,in, Given the number of Monte Carlo samplings, the reliability R(T) = 1 − Pf(0, T), and the failure rate function. The formula for calculating (T) is: .

6. The AI-based dynamic update-based intelligent equipment full-lifecycle reliability assessment method according to claim 1, characterized in that, It also includes step S6: Intelligent management of the model lifecycle; building a hierarchical management system, including: The data perception layer continuously collects real-time operating condition data and metadata of each PINN model; The analysis and decision-making layer calculates the health of the model based on the data, analyzes the evolution trend of the working condition, monitors computing resources, and outputs decision instructions. The execution control layer executes dynamic management strategies, ranging from routine incremental maintenance to active network structure optimization and cross-model collaborative training, based on the decision instructions.

7. An AI-based dynamically updated intelligent equipment lifecycle reliability assessment system, used to implement the AI-based dynamically updated intelligent equipment lifecycle reliability assessment method as described in any one of claims 1-6, characterized in that, The system includes: The model building and deployment module is used to build, initially verify, and deploy the PINN models of each core subsystem to the equipment edge computing terminal. The data acquisition and feature extraction module includes a sensor array, a data acquisition circuit, and an automatic encoder deployed on various components, used to acquire and extract health feature vectors in real time. The working condition perception and update control module integrates a GMM clusterer to determine the working condition and trigger model update instructions; The incremental learning update engine is used to execute an incremental learning algorithm that combines EWC and experience replay to update the specified PINN model after receiving an update instruction. The reliability assessment and early warning module is used to call the updated PINN model to perform response prediction, failure probability calculation and risk warning. The model lifecycle management module is used to monitor the status of each model, coordinate resources, and execute dynamic management strategies.

Images