Intelligent factory life cycle prediction management system and method based on digital twinning

Abstract

This invention discloses a smart factory lifecycle prediction management system and method based on digital twins, belonging to the field of factory lifecycle prediction management. The system includes: a data acquisition and preprocessing module, configured to collect multi-source status data in real time from sensor groups on key equipment in each production stage of the smart factory, generating a standardized time-series data stream; an independent digital twin construction module, configured to construct an independent digital twin for each key equipment; a degradation trajectory prediction module, integrated into each independent digital twin, configured to receive the standardized time-series data stream; and a lifecycle prediction output module, configured to calculate the remaining service life distribution function of each key equipment based on the multiple predicted degradation trajectories and their predicted distributions. This invention uses independent twin modeling for each device, integrating geometric, physical, and behavioral multi-dimensional characteristics to specifically characterize the degradation patterns of individual equipment, resulting in strong predictive adaptability.

Description

Technical Field

[0001] This invention relates to the field of factory lifecycle prediction management, specifically to an intelligent factory lifecycle prediction management system and method based on digital twins. Background Technology

[0002] As the core carrier of industrial digital and intelligent transformation, the stable operation of key production equipment in smart factories directly determines production efficiency, product quality, and production costs. Equipment lifecycle management (including health monitoring, degradation prediction, remaining life assessment, and maintenance scheduling) is a core component of the smart factory operation and maintenance system. Accurate lifecycle prediction can effectively avoid sudden equipment failures, reduce downtime losses, and optimize the allocation of maintenance resources.

[0003] Digital twin technology, through the construction of real-time mapping, data interaction, and virtual-real fusion between physical entities and virtual models, can accurately replicate the geometric shape, physical characteristics, and operating behavior of equipment, providing a technical path for the visualization and precise management of the entire equipment lifecycle. Therefore, intelligent factory lifecycle prediction management based on digital twins has gradually become the mainstream research direction in the industry. Existing technologies have realized functions such as basic data collection, overall factory twin modeling, and simple time-series degradation prediction, which have improved the digitalization level of lifecycle management to a certain extent.

[0004] However, existing digital twins mostly involve unified modeling of the entire factory, resulting in homogeneous digital twin models. They fail to construct independent twins for individual key equipment, thus failing to accurately depict the geometric, physical, and behavioral characteristics of individual equipment and leading to poor predictive specificity. Therefore, those skilled in the art have provided a smart factory lifecycle prediction management system and method based on digital twins to address the problems mentioned in the background. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies and provide a smart factory lifecycle prediction management system and method based on digital twins. It solves the problems of low data processing accuracy, homogenization of twin modeling, and weak capture of spatiotemporal features in existing technologies, and realizes accurate prediction, visual management, and intelligent maintenance scheduling of the lifecycle of key equipment in smart factories.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A digital twin-based intelligent factory lifecycle prediction and management system includes:

[0008] The data acquisition and preprocessing module is configured to acquire multi-source status data in real time from sensor groups on key equipment in each production link of the smart factory, and perform time synchronization, outlier removal and normalization processing on the multi-source status data to generate a standardized time-series data stream.

[0009] An independent digital twin construction module is configured to construct an independent digital twin for each of the key devices. Each digital twin includes a geometric model that has a geometric mapping to the physical entity of the key device, a physical model that describes its physical field distribution, a behavioral model that characterizes its current working condition and behavior, and a core prediction engine. The core prediction engine is based on a dynamic spatiotemporal attention-residual memory network architecture.

[0010] The degradation trajectory prediction module, integrated within each independent digital twin, is configured to receive the standardized time-series data stream and sequentially execute the following: a feature embedding layer maps to a high-dimensional feature space; a dynamic spatiotemporal attention layer calculates and weights attention weights along the time and sensor space dimensions; a residual memory layer processes the weighted features through a gated recurrent unit network regulated by a drift-diffusion joint state update mechanism; and a degradation trajectory generation layer outputs multiple predicted degradation trajectories and their predicted distributions through a fully connected neural network mapping and Monte Carlo dropout method.

[0011] The lifecycle prediction output module is configured to calculate the remaining service life distribution function of each key device based on the multiple predicted degradation trajectories and their predicted distributions, and to present the predicted degradation trajectory, current health indicators and remaining service life confidence interval of each device in a twin 3D scene in a visual form.

[0012] As a further aspect of the present invention: the data acquisition and preprocessing module further includes a multi-source data alignment sub-module; the multi-source data alignment sub-module performs elastic alignment based on dynamic time warping on two sensor data sequences with different sampling frequencies, and after calculating the point-to-point distance cost matrix of the two sensor data sequences, introduces a bending penalty term based on the local slope of the sequence. The bending penalty term is obtained by multiplying the absolute value of the difference in the local rate of change of the two sensor data sequences at the corresponding positions by an adaptive penalty coefficient that is negatively correlated with the amplitude of sequence fluctuation.

[0013] As a further aspect of the present invention: the physical model adopts a data-driven enhanced reduced-order physical model; the construction process of the reduced-order physical model is as follows: a full-order finite element model is established based on the physical mechanism of the key equipment; the main energy modes of the full-order model are extracted using the intrinsic orthogonal decomposition method using historical full-life cycle data, and the first few orders of modes are retained to form the reduced-order basis; a shallow neural network is trained as a residual corrector, the input of the residual corrector is the prediction error of the reduced-order model at the key measurement points and its corresponding operating parameters, and the output is the real-time correction amount of the output of the reduced-order model.

[0014] As a further aspect of the present invention: the process of calculating attention weights along the time dimension in the dynamic spatiotemporal attention layer is as follows: the original attention score is calculated by using a bilinear transformation matrix to transform the feature vector at the current moment and the feature vector at the historical moment; a decay factor based on a time interval is introduced, which is obtained by multiplying the time difference between the current moment and the historical moment by a learnable parameter characterizing the memory strength of the degradation process after exponential decay calculation; the original attention score is multiplied element-wise by the decay factor and then normalized by Softmax to obtain the temporal attention weight, the mathematical expression of which is:

[0015] ;

[0016] Where t is the index of the current time, s is the index of the historical time, and j is the summation index. The time attention weight of the current time t to the historical time s; and These are the feature vectors at the current time and the historical time, respectively, with the superscript T indicating vector transpose; It is a learnable bilinear transformation matrix; This is a learnable memory strength parameter, with a value ranging from 0 to 1.

[0017] As a further aspect of the present invention: the drift-diffusion joint state update mechanism is implemented by defining the latent health state variable of the device at time t. Its update process follows discretized stochastic differential equations; the drift term is obtained by processing the input features of the residual memory layer through an adaptive linear layer, and then adding a correction amount to adjust the matching degree between the previous state value and the current observation information; the diffusion term is obtained by extracting the time-varying noise intensity coefficient by performing statistical characteristic analysis on the high-frequency components in the multi-source state data at the current time, and then multiplying it by a standard Gaussian random variable; latent variables The summation of the drift term and the diffusion term, plus the latent variable from the previous time step, is then performed. The mathematical expression for obtaining is:

[0018] ;

[0019] Where t is the discrete time step index. Let be the latent variable of health status at time t. For the latent variable of health status at the previous moment, For the multi-source observation characteristics at time t, Let θ be the drift term function, and θ be the learnable model parameter vector within the drift term function. Let be the diffusion term function, which passes through the vibration signal at time t. Envelope spectrum entropy and acoustic emission signal The output noise intensity is calculated by combining the energy release rate. These are random numbers from a standard normal distribution. is the discrete time step.

[0020] As a further aspect of the present invention, it also includes a cross-digital twin coupling field analysis module; the cross-digital twin coupling field analysis module calculates and maintains a coupling field matrix describing the implicit interactions between all independent digital twins; each element in the coupling field matrix represents the degradation influence intensity between two key equipment digital twins, the degradation influence intensity is obtained by performing dynamic time warping on the operating condition sequences of the two equipment to calculate similarity, and then multiplying by an attenuation coefficient characterizing the topological distance between the two in the production process; the cross-digital twin coupling field analysis module inputs the coupling field matrix as an additional feature to the degradation trajectory prediction module of each independent digital twin in each prediction cycle, specifically by concatenating it with the feature tensor before the dynamic spatiotemporal attention layer.

[0021] As a further aspect of the present invention: the lifecycle prediction output module further includes a competitive failure mode discriminator; the competitive failure mode discriminator detects the time point at which each of the multiple predicted degradation trajectories first crosses a preset failure threshold within a preset prediction time window; clusters all trajectories according to different physical failure modes; calculates the kernel density estimate of the remaining lifetime for each group to obtain the remaining lifetime probability distribution under that failure mode; determines the failure mode that occurs first and whose probability density peak exceeds a preset confidence threshold as the dominant failure mode, and outputs the degradation characteristics and highlighted data of the predicted trajectory corresponding to the dominant failure mode.

[0022] As a further aspect of the present invention, it also includes a predictive maintenance strategy self-generation module; the predictive maintenance strategy self-generation module receives the remaining useful life distribution function of each key device and preset maintenance resource constraints; the predictive maintenance strategy self-generation module embeds a scheduling agent based on deep reinforcement learning, the state space of the scheduling agent includes the current health indicators, the mean and variance of the remaining useful life prediction, and the occupancy of maintenance resources for each device; the action space of the scheduling agent is a combination of actions of different maintenance levels assigned to each device; the reward function of the scheduling agent is a weighted sum of the total downtime cost, preventive maintenance cost, and production loss cost caused by sudden failure of all devices, multiplied by a risk penalty term inversely proportional to the uncertainty of the remaining useful life prediction; the scheduling agent is trained through historical simulation interaction with the environment and outputs a collaborative maintenance task list.

[0023] As a further aspect of the present invention, it also includes a twin online evolution and incremental learning module; when the prediction error between the health index output by the degradation trajectory prediction module and the subsequent actual collected observation data feedback continuously exceeds a preset dynamic threshold, the twin online evolution and incremental learning module triggers an incremental update of the core prediction engine inside the independent digital twin; the incremental update adopts an elastic weight consolidation algorithm, which applies a secondary penalty term when calculating the loss gradient of the new batch of data; the secondary penalty term is obtained by multiplying the square of the change in each historical weight in the core prediction engine by a regularization coefficient determined according to the diagonal elements of the Fisher information matrix, and this coefficient is positively correlated with the importance of the weight to the historical task.

[0024] This application also discloses a smart factory lifecycle prediction and management method based on digital twins, which adopts a smart factory lifecycle prediction and management system based on digital twins, including the following steps:

[0025] Step S1: Collect multi-source status data in real time from the sensor groups on key equipment in each production link of the smart factory, and perform time synchronization, outlier removal and normalization processing on the multi-source status data to generate a standardized time-series data stream.

[0026] Step S2: Construct an independent digital twin for each key device. Each digital twin includes a geometric model that has a geometric mapping to the physical entity of the key device, a physical model that describes its physical field distribution, a behavioral model that characterizes its current working condition and behavior, and a core prediction engine. The core prediction engine is based on a dynamic spatiotemporal attention-residual memory network architecture.

[0027] Step S3: Receive the standardized time-series data stream and execute the following steps in sequence: the feature embedding layer maps to a high-dimensional feature space; the dynamic spatiotemporal attention layer calculates and weights attention weights along the time and sensor space dimensions; the residual memory layer processes the weighted features through a gated recurrent unit network regulated by a drift-diffusion joint state update mechanism; and the degradation trajectory generation layer outputs multiple predicted degradation trajectories and their predicted distributions through a fully connected neural network mapping and Monte Carlo dropout method.

[0028] Step S4: Based on multiple predicted degradation trajectories and their predicted distributions, calculate the remaining useful life distribution function for each key device, and present the predicted degradation trajectory, current health indicators, and remaining useful life confidence interval of each device in a twin 3D scene in a visual format.

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

[0030] High data processing accuracy: Dynamic time warping + local slope curvature penalty term accurately aligns multi-frequency time series data, eliminates time series deviations, and provides high-quality data support for prediction.

[0031] Accurate twin modeling: Independent twin modeling of single device, integrating geometric, physical and behavioral characteristics, to specifically characterize the degradation pattern of individual devices, with strong predictive adaptability.

[0032] Strong spatiotemporal feature capture capability: The dynamic spatiotemporal attention network takes into account both the temporal degradation correlation and the sensor spatial coupling, and the residual memory layer accurately models the random degradation process, which greatly improves the prediction accuracy.

[0033] More reliable global predictions: The implicit degradation effects between devices are quantified by cross-twin coupled field matrix and incorporated into the prediction process to avoid the one-sidedness of predictions from a single device.

[0034] Model Adaptive Evolution: The Elastic Weight Consolidation Algorithm enables online incremental updates of the prediction engine, balances the performance of new and old tasks, adapts to long-term dynamic degradation of equipment, and reduces maintenance costs.

[0035] Intelligent maintenance scheduling: Deep reinforcement learning scheduling agent comprehensively considers multi-dimensional costs and constraints to generate the optimal maintenance plan, avoiding over- or under-maintenance and improving operation and maintenance efficiency.

[0036] Intuitive and visual management: The 3D twin scene visualization presents the degradation trajectory, health indicators, and remaining lifespan, and supports highlighting of dominant failure modes, facilitating quick decision-making by maintenance personnel. Attached Figure Description

[0037] Figure 1 This is a structural block diagram of a smart factory lifecycle prediction and management system based on digital twins.

[0038] Figure 2 This is a flowchart of a smart factory lifecycle prediction and management method based on digital twins. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] As mentioned in the background section of this application, research has found that existing digital twins are mostly unified models of the entire factory, resulting in homogeneous digital twin models. They do not construct independent twins for individual key equipment, and therefore cannot accurately depict the geometric, physical, and behavioral characteristics of individual equipment, leading to poor predictive targeting and certain defects.

[0041] To address the aforementioned shortcomings, this application discloses a smart factory lifecycle prediction management system and method based on digital twins. It features independent twin modeling for each device, integrating geometric, physical, and behavioral characteristics to specifically characterize the degradation patterns of individual devices, resulting in strong predictive adaptability.

[0042] The following will describe in detail, with reference to the accompanying drawings, how the solution of this application solves the above-mentioned technical problems.

[0043] Please see Figure 1In this embodiment of the invention, the intelligent factory lifecycle prediction management system based on digital twins includes: a data acquisition and preprocessing module, configured to acquire multi-source status data in real time from sensor groups on key equipment in each production stage of the intelligent factory, and perform time synchronization, outlier removal, and normalization processing on the multi-source status data to generate a standardized time-series data stream; an independent digital twin construction module, configured to construct an independent digital twin for each key equipment, each digital twin including a geometric model with geometric mapping to the physical entity of the key equipment, a physical model describing its physical field distribution, a behavioral model characterizing its current working condition and behavior, and a core prediction engine; the core prediction engine is based on a dynamic spatiotemporal attention-residual memory network architecture; and a degradation trajectory prediction module. The module, integrated within each independent digital twin, is configured to receive a standardized time-series data stream and execute the following sequentially: a feature embedding layer maps to a high-dimensional feature space; a dynamic spatiotemporal attention layer calculates and weights attention weights along the time and sensor space dimensions; a residual memory layer processes weighted features through a gated recurrent unit network regulated by a drift-diffusion joint state update mechanism; a degradation trajectory generation layer outputs multiple predicted degradation trajectories and their predicted distributions through fully connected neural network mapping and Monte Carlo dropout method; and a lifespan prediction output module, configured to calculate the remaining lifespan distribution function of each key device based on multiple predicted degradation trajectories and their predicted distributions, and to present the predicted degradation trajectory, current health indicators, and remaining lifespan confidence interval of each device in a visualized form within the twin's 3D scene. The data acquisition and preprocessing module collects multi-source status data of key equipment in real time and completes time synchronization, outlier removal, and normalization processing to ensure the consistency of input data time sequence and reliable quality, laying a solid data foundation for subsequent accurate predictions. An independent digital twin construction module creates a dedicated digital twin for each key piece of equipment, integrating geometric, physical, and behavioral models with a dynamic spatiotemporal attention-residual memory network core prediction engine to accurately replicate individual equipment characteristics and improve prediction targeting and adaptability. The degradation trajectory prediction module performs feature mapping, spatiotemporal attention weighting, and random degradation modeling in a layered manner, outputting multiple degradation trajectories and probability distributions to accurately capture degradation patterns and quantify prediction uncertainties. The lifecycle prediction output module calculates the remaining service life distribution and presents the equipment degradation trajectory, health indicators, and lifespan confidence intervals in a 3D visualization format, supporting rapid operation and maintenance decisions. Overall, this achieves high-precision, visualized, and refined prediction management of the lifecycle of key equipment in the smart factory.

[0044] In this embodiment, the data acquisition and preprocessing module further includes a multi-source data alignment submodule. This submodule performs elastic alignment based on dynamic time warping on two sensor data sequences with different sampling frequencies. After calculating the point-to-point distance cost matrix of the two sensor data sequences, it introduces a bending penalty term based on the local slope of the sequence. This bending penalty term is obtained by multiplying the absolute value of the difference in the local rate of change of the two sensor data sequences at corresponding positions by an adaptive penalty coefficient negatively correlated with the sequence fluctuation amplitude. By adding the multi-source data alignment submodule, applying dynamic time warping elastic alignment to sensor data sequences with different sampling frequencies and introducing a bending penalty term based on the local slope of the sequence, the temporal misalignment and alignment distortion caused by differences in sampling frequencies are effectively eliminated. This improves the temporal consistency and matching accuracy of multi-source data, prevents data deviations from being transmitted to subsequent prediction stages, and ensures the accuracy and reliability of degradation trajectory prediction and lifecycle assessment.

[0045] Specific example: Taking the timing data alignment of a high-frequency vibration sensor (20kHz) and a low-frequency temperature sensor (1Hz) on a machine tool in a smart factory as an example, the specific implementation is as follows:

[0046] Data sequence setting; assuming a high-frequency vibration data sequence within 1 second is... The low-frequency temperature data sequence is The sampling frequencies of the two sequences differ significantly, and direct alignment will result in serious timing deviations.

[0047] Basic distance cost matrix; calculate the point-to-point distance cost matrix between two sequences. The formula for the elements is: ;

[0048] Bending penalty term calculation; introduction of a bending penalty term based on local slope. The formula is: ;in, The local slope of the vibration sequence. (Temperature remains unchanged for 1 second) It is an adaptive penalty coefficient (negatively correlated with the vibration fluctuation amplitude).

[0049] Alignment effect verification; the corrected cost matrix is ​​as follows: Based on this matrix, dynamic regularization path optimization is achieved, enabling flexible alignment of high-frequency and low-frequency data. The timing matching error is reduced from 150ms in the traditional method to less than 5ms, completely eliminating the timing misalignment problem.

[0050] In this embodiment, the physical model adopts a data-driven enhanced reduced-order physical model. The construction process of the reduced-order physical model is as follows: a full-order finite element model is established based on the physical mechanism of key equipment; the main energy modes of the full-order model are extracted using the intrinsic orthogonal decomposition method using historical life cycle data, and the first few orders of modes are retained to form the reduced-order basis; a shallow neural network is trained as a residual corrector. The input of the residual corrector is the prediction error of the reduced-order model at key measurement points and its corresponding operating parameters, and the output is the real-time correction amount of the reduced-order model output. By extracting the core energy modes of the full-order finite element model through intrinsic orthogonal decomposition, the model order is reduced, which significantly reduces the computational complexity and improves the real-time performance of modeling. Then, the shallow neural network is used as a residual corrector to compensate for the output error of the reduced-order model in real time, which takes into account both the computational efficiency and prediction accuracy of the physical model. This effectively solves the problems of high computational consumption of the full-order model and error accumulation of the traditional reduced-order model, and is suitable for the modeling requirements of real-time interaction of digital twins.

[0051] Specific example: Taking the physical model construction of a high-speed spindle (critical rotating equipment) in a smart factory as an example, the specific implementation is as follows:

[0052] A full-order finite element model is constructed. Based on the material mechanics and heat conduction mechanisms of the spindle, a full-order finite element model with 10,000 mesh elements is established. The temperature and stress field data of key measuring points such as the bearing and journal are output. The output of the full-order model is denoted as... Where N=10000 is the full-order output dimension.

[0053] Intrinsic Orthogonal Decomposition (POD) for mode reduction; 1000 sets of historical full-lifecycle operating condition data of the spindle were collected, and energy modes were extracted through POD decomposition. The energy proportion formula for the k-th mode is as follows: ;in, Let k be the k-th order eigenvalue. The top 5 modes by energy percentage are selected to construct a reduced-order basis. The output of the reduced-order model is The output dimension was reduced from 10,000 to 5, and the calculation time per operation was reduced from 200ms to 5ms.

[0054] Shallow Neural Network Residual Correction: A 3-layer fully connected shallow neural network is constructed as the residual corrector, with the input being the measurement point error of the reduced-order model. Operating parameters (speed n, load L), output correction amount The final output is The average error of the reduced-order model before correction was 8.2%, and the error after correction was reduced to 1.1%, balancing low computing power and high accuracy.

[0055] In this embodiment, the process of calculating attention weights along the time dimension in the dynamic spatiotemporal attention layer is as follows: The original attention score is calculated by transforming the feature vector at the current moment and the feature vector at the historical moment using a bilinear transformation matrix; a time-interval-based decay factor is introduced, which is obtained by multiplying the time difference between the current moment and the historical moment by a learnable parameter characterizing the memory strength during the degradation process; the original attention score is multiplied element-wise by the decay factor and then normalized using Softmax to obtain the temporal attention weight, the mathematical expression of which is:

[0056] ;

[0057] Where t is the index of the current time, s is the index of the historical time, and j is the summation index. The time attention weight of the current time t to the historical time s; and These are the feature vectors at the current time and the historical time, respectively, with the superscript T indicating vector transpose; It is a learnable bilinear transformation matrix; This is a learnable memory strength parameter, ranging from 0 to 1. By introducing a time-interval-based decay factor in the time dimension through a dynamic spatiotemporal attention layer, and combining it with a bilinear transformation matrix to calculate the original attention score and then normalizing it using Softmax to obtain the temporal attention weight, this approach can adaptively weaken the interference of long-standing historical temporal features and strengthen the influence of recent key degradation information. It accurately captures the nonlinear degradation correlation of equipment over time, aligning with the actual pattern that recent operating conditions have a greater impact on the current degradation state. This effectively improves the targeting of temporal feature extraction and the temporal accuracy of degradation trajectory prediction, avoiding interference from irrelevant historical information in the prediction results.

[0058] Specific example: Taking the time-series feature data of a high-speed spindle in a smart factory as an example, we verify the effect of time attention weight calculation.

[0059] Parameter settings; extract the feature vectors of five consecutive moments along the principal axis. (Current time t=5, historical times s=1,2,3,4), eigenvector dimension is 64; bilinear transformation matrix Memory strength parameters (Values ​​range from 0 to 1).

[0060] Substitute into the formula to calculate the weights; calculate the time attention weights using the formula above:

[0061] ;

[0062] Time interval decay term: |5-1|=4, |5-2|=3, |5-3|=2, |5-4|=1 .

[0063] Weighting and Normalization: Scores at distant times s=1 are significantly decayed, while scores at recent times s=4 are retained more, resulting in a final weight of... .

[0064] The weights exhibit a distribution that is high in the near term and low in the long term, which aligns with the reality that recent vibration and temperature data of the main shaft have a greater impact on degradation. Compared with traditional attention without a decay factor, invalid historical interference is reduced, and the accuracy of temporal feature extraction is improved.

[0065] In this embodiment, the drift-diffusion joint state update mechanism is implemented as follows: defining the latent health state variable of the device at time t. Its update process follows discretized stochastic differential equations; the drift term is obtained by processing the input features of the residual memory layer through an adaptive linear layer, and then adding a correction amount to adjust the matching degree between the previous state value and the current observation information; the diffusion term is obtained by extracting the time-varying noise intensity coefficient by performing statistical characteristic analysis on the high-frequency components in the multi-source state data at the current time, and then multiplying it by a standard Gaussian random variable; latent variables The drift and diffusion terms are summed, and then the latent variable from the previous time step is added. The mathematical expression for obtaining is:

[0066] ;

[0067] Where t is the discrete time step index. Let be the latent variable of health status at time t. For the latent variable of health status at the previous moment, For the multi-source observation characteristics at time t, Let θ be the drift term function, and θ be the learnable model parameter vector within the drift term function. Let be the diffusion term function, which passes through the vibration signal at time t. Envelope spectrum entropy and acoustic emission signal The output noise intensity is calculated by combining the energy release rate. These are random numbers from a standard normal distribution. The time step is discrete. A drift-diffusion joint state update mechanism is used to model the latent variables of equipment health state using discretized stochastic differential equations. The drift term captures the deterministic trend of degradation and introduces an adaptive correction to fit the changes in operating conditions. The diffusion term extracts time-varying noise based on the statistical characteristics of high-frequency vibration and acoustic emission signals to characterize the randomness of degradation. This accurately integrates deterministic degradation patterns and random fluctuation characteristics, solving the problems of large prediction bias and poor robustness caused by traditional models that only fit trends and ignore random interference. It significantly improves the adaptability and reliability of degradation trajectory prediction to random degradation processes under complex operating conditions.

[0068] Specific example: Taking the health status degradation model of a high-speed spindle in a smart factory as an example, the implementation effect of the drift-diffusion joint state update mechanism is verified.

[0069] Parameter and variable settings; determining the discrete time step Latent variables of health status at the previous moment (Normalized health value, 0 for failure, 1 for health); Current observed feature vector Model parameter vector θ; vibration signal Acoustic emission signals Standard normal random numbers =0.2.

[0070] Drift term calculation; drift term After correction for the adaptive linear layer and observation matching degree, the following is calculated: ;

[0071] Calculation of diffusion term; based on Envelope spectral entropy and The energy release rate is used to calculate the noise intensity factor. The diffusion term is: ;

[0072] Update latent variables of health status; substitute into the above formula to calculate the latent variables at the current time: ;

[0073] Substitute the values: ;

[0074] This mechanism simultaneously captures the deterministic degradation trend of spindle bearing wear and the random fluctuations caused by vibration and shock. Compared with traditional models that only fit the trend, the fitting error of random fluctuations is reduced, the robustness of degradation trajectory prediction is improved, and it fits the actual complex random degradation characteristics of the equipment.

[0075] In this embodiment, a cross-digital twin coupling field analysis module is also included. The cross-digital twin coupling field analysis module calculates and maintains a coupling field matrix that describes the implicit interactions between all independent digital twins. Each element in the coupling field matrix represents the degradation influence intensity between two key equipment digital twins. The degradation influence intensity is obtained by performing dynamic time warping on the operating condition sequences of the two equipment to calculate the similarity, and then multiplying it by a decay coefficient that characterizes the topological distance between the two in the production process. The cross-digital twin coupling field analysis module inputs the coupling field matrix as an additional feature to the degradation trajectory prediction module of each independent digital twin in each prediction cycle. Specifically, it concatenates the matrix with the feature tensor before the dynamic spatiotemporal attention layer. By combining the dynamic time warping similarity of the operating condition sequence with the process flow topology distance attenuation coefficient, the implicit degradation influence intensity between independent digital twins is quantified. The coupling field matrix is ​​used as an additional feature input to the degradation trajectory prediction module, which effectively solves the problem of isolated prediction bias caused by neglecting the process flow linkage between equipment in the existing technology. It accurately portrays the degradation correlation of mutual interference between equipment, realizes collaborative prediction from a global perspective, and significantly improves the degradation prediction accuracy and overall reliability of multi-equipment clusters.

[0076] Specific example: Taking the three core machine tools (named Machine Tool 1, Machine Tool 2, and Machine Tool 3) of the intelligent factory with interconnected process flow as an example, we construct a cross-twin coupling field matrix and verify the effect.

[0077] Parameter definition: Operating condition sequence similarity, calculated by dynamic time warping of the operating condition sequences of two machine tools. The value ranges from [0,1], with larger values ​​indicating stronger correlation between operating conditions; the topological distance attenuation coefficient is determined by the distance between adjacent machine tools. Take from one machine tool at intervals The machine tool itself takes Coupled field matrix elements, , representing the intensity of the degradation effect of machine tool i on machine tool j.

[0078] Coupled field matrix construction: The specific similarity of the measured working conditions is as follows: Self-similarity Substitute the values ​​to calculate the 3×3 coupling field matrix. :

[0079] ;

[0080] The coupling field matrix is ​​flattened (3×3 → 9 dimensions) and spliced ​​to the feature tensor in front of the dynamic spatiotemporal attention layer of each machine tool degradation trajectory prediction module (originally 64 dimensions → 73 dimensions). When predicting in isolation, the average degradation prediction error of the three machine tools is relatively large; after introducing the coupling field matrix, the average error is significantly reduced, the prediction deviation caused by mutual interference between equipment is reduced, and it conforms to the law of equipment linkage degradation in actual production.

[0081] In this embodiment, the lifecycle prediction output module also includes a competitive failure mode discriminator. The competitive failure mode discriminator detects the time point at which each of the multiple predicted degradation trajectories first crosses a preset failure threshold within a preset prediction time window; it clusters all trajectories according to different physical failure modes; it calculates the kernel density estimate of the remaining lifetime for each group to obtain the probability distribution of the remaining lifetime under that failure mode; it identifies the failure mode that occurs first and whose probability density peak exceeds a preset confidence threshold as the dominant failure mode, and outputs the degradation characteristics and highlighted data of the predicted trajectory corresponding to this dominant failure mode. Through the competitive failure mode discriminator, the system detects the time of first crossing of the failure threshold for multiple predicted degradation trajectories, clusters different physical failure modes, calculates the kernel density estimate probability distribution of the remaining lifetime for each mode, accurately identifies the dominant failure mode that occurs first and has the highest confidence, and highlights the relevant data. This effectively solves the problems of difficulty in distinguishing competing failures, difficulty in locating dominant modes, and lack of targeted maintenance decisions when multiple failure modes coexist. It avoids wasting maintenance resources and missing faults caused by ignoring secondary failures or misjudging dominant failures, significantly improving the accuracy of failure prediction and the scientific nature of maintenance strategy formulation.

[0082] Specific example: Taking the high-speed spindle in a smart factory as an example, there are two competing failure modes: mode A (bearing wear failure) and mode B (spindle runout accuracy failure). The working effect of the competing failure mode discriminator is verified.

[0083] Failure threshold setting; Normalized degradation index threshold: Mode A failure threshold (Wearing amount exceeds limit), Mode B failure threshold (Actuation volume exceeds limit); Preset confidence threshold .

[0084] Degradation trajectory crossing detection; generate 10 predicted degradation trajectories, and detect the time point of the first crossing threshold within a preset time window for each trajectory:

[0085] Crossing 6 tracks first This corresponds to mode A;

[0086] The four tracks cross first This corresponds to mode B.

[0087] Failure mode clustering; clustered into two groups based on failure physical mechanism:

[0088] Group A (Pattern A): {Trajectory 1, 2, 3, 4, 5, 6};

[0089] Group B (Pattern B): {Trajectory 7, 8, 9, 10}.

[0090] Remaining lifetime kernel density estimation: The probability density of remaining lifetime for each group is calculated using Gaussian kernel density estimation, with the following formula:

[0091] Where r: the remaining life variable of the probability density to be estimated (any life value, such as day 10, 12, or 15); : The remaining lifetime value of the sample corresponding to the i-th trajectory (Mode A has 6 trajectories, i.e. n: Number of trajectories in this failure mode; h: Bandwidth; : Gaussian kernel function.

[0092] Pattern A: Mean remaining lifespan 12 days, peak probability density ;

[0093] Mode B: Mean remaining lifespan 18 days, peak probability density .

[0094] Mode A satisfies the criteria of "first occurrence + peak value exceeding the confidence threshold" and is determined to be the dominant failure mode. The degradation trajectory and bearing wear characteristic data corresponding to Mode A are highlighted in the twin interface to facilitate operation and maintenance in locating the core failure source.

[0095] Traditional methods only make general predictions of spindle failure and cannot distinguish between competing modes; this solution accurately identifies the primary failure as bearing wear, improves targeted maintenance, and avoids the cost waste caused by blindly replacing the spindle.

[0096] In this embodiment, a predictive maintenance strategy self-generation module is also included. This module receives the remaining useful life distribution function of each key device and preset maintenance resource constraints. A deep reinforcement learning-based scheduling agent is embedded within the module. The scheduling agent's state space includes the current health indicators, the mean and variance of the remaining useful life prediction, and the occupancy of maintenance resources for each device. The scheduling agent's action space consists of combinations of actions at different maintenance levels assigned to each device. The scheduling agent's reward function is a weighted sum of the total downtime cost, preventative maintenance cost, and production loss cost caused by sudden failures for all devices, multiplied by a risk penalty term inversely proportional to the uncertainty of the remaining useful life prediction. The scheduling agent is trained through historical simulation interaction with the environment and outputs a collaborative maintenance task list. By using a predictive maintenance strategy self-generation module and embedding a deep reinforcement learning scheduling agent, equipment health indicators, remaining life prediction parameters, and maintenance resource constraints are incorporated into the state space. An action space is constructed using differentiated maintenance level combinations. A reward function that integrates downtime costs, maintenance costs, production losses, and risk penalties is designed. Through historical simulation interactive training, a collaborative maintenance task list is output. This effectively solves the problems of high costs, large downtime losses, and resource waste caused by traditional maintenance relying on manual experience, resource allocation imbalance, and inappropriate maintenance timing. It achieves optimal allocation of maintenance resources and precise planning of maintenance timing, significantly reducing overall operation and maintenance costs and the risk of sudden failures.

[0097] Specific example: Taking the maintenance scheduling of three linked machine tools (machine tool 1, machine tool 2, and machine tool 3) in a smart factory as an example, the implementation effect of deep reinforcement learning scheduling agent is verified.

[0098] Scene and parameter settings; equipment status, machine tool health. Mean remaining lifespan Heaven, variance Machine tool health status , sky, Machine tool health status 3 , sky, Resource constraints: 2 maintenance personnel, 1 set of spare bearings, resource utilization rate. .

[0099] State space definition; state vector Includes health status, remaining lifespan parameters, and resource status:

[0100] Example status: .

[0101] Action space definition; maintenance actions Corresponding machine tool maintenance levels: 0 (no maintenance), 1 (inspection), 2 (minor repair), 3 (major overhaul), action combinations .

[0102] Reward function formula: ;in, Total downtime cost; Maintenance costs; Sudden failure loss; : Mean of variance of remaining life expectancy; the larger the variance, the stronger the risk penalty.

[0103] Optimal action and reward calculation; filter possible actions a={2, 1, 0} (machine tool 1 minor repair, machine tool 2 spot check, machine tool 3 no maintenance), substitute parameters: =500 yuan, =800 yuan, =300 yuan ;

[0104] This action has the highest reward value and is the optimal collaborative maintenance solution.

[0105] In this embodiment, an online evolution and incremental learning module for the twin is also included. When the prediction error between the health index output by the degradation trajectory prediction module and the subsequent actual collected observation data feedback continuously exceeds a preset dynamic threshold, the online evolution and incremental learning module for the twin triggers an incremental update of the core prediction engine inside the independent digital twin. The incremental update adopts an elastic weight consolidation algorithm, which applies a secondary penalty term when calculating the loss gradient of the new batch of data. The secondary penalty term is obtained by multiplying the square of the change in each historical weight in the core prediction engine by a regularization coefficient determined based on the diagonal elements of the Fisher information matrix. This coefficient is positively correlated with the importance of the weight to the historical task. The prediction error is monitored in real time through the online evolution and incremental learning module of the digital twin. When the error continues to exceed the dynamic threshold, the core prediction engine is triggered to update incrementally. The elastic weight consolidation algorithm introduces a quadratic penalty term based on the diagonal elements of the Fisher information matrix into the loss function, and applies differentiated regularization constraints to historically important weights. This effectively balances the prediction performance under new and old operating conditions, avoids catastrophic forgetting in incremental learning, and can adapt to the characteristics changes caused by long-term equipment operating condition drift and component aging without offline retraining. This greatly improves the long-term stability, adaptability and generalization of the digital twin prediction model, and reduces the model maintenance cost and the risk of prediction bias accumulation.

[0106] Specific example: Taking the incremental update of the prediction engine for the joint motor of an industrial robot in a smart factory (which is prone to characteristic drift during long-term operation) as an example, the implementation effect of the elastic weight consolidation algorithm is verified.

[0107] Scene and error threshold settings: After long-term operation, the load characteristics of the robot's joint motors drift, and the dynamic threshold for prediction error is set to 6%. When the prediction error exceeds the threshold three times consecutively, incremental learning of the core prediction engine is triggered. The engine contains two key weights: (Important for predicting degradation of old operating conditions) (Prediction of old operating conditions is secondary).

[0108] The elastic weighting consolidation loss function; the total incremental update loss consists of the task loss and a quadratic penalty term, as shown in the formula:

[0109] ;in, Losses in the new batch of data prediction task; The diagonal element of the Fisher information matrix corresponding to the k-th weight (the higher the importance of the weight to the old task, the more important it is). The larger); : The change in the kth weight.

[0110] Parameter substitution and loss calculation; Task loss: Fisher Information: ( important), ( Secondary); weight change: .

[0111] Substitute into the formula to calculate the penalty term and total loss:

[0112] .

[0113] This application also discloses a method for predictive management of the lifecycle of a smart factory based on digital twins, such as... Figure 2As shown, a smart factory lifecycle prediction and management system based on digital twins is adopted, including the following steps: Step S1: Real-time collection of multi-source status data from sensor groups on key equipment in each production link of the smart factory, and time synchronization, outlier removal and normalization processing of the multi-source status data to generate a standardized time-series data stream; Step S2: Construction of an independent digital twin for each key equipment. Each digital twin includes a geometric model with geometric mapping to the physical entity of the key equipment, a physical model describing its physical field distribution, a behavioral model representing its current working condition and behavior, and a core prediction engine; the core prediction engine is based on a dynamic spatiotemporal attention-residual memory network architecture; Step S3: ... Step S3: Receive the standardized time-series data stream and execute the following steps in sequence: the feature embedding layer maps to a high-dimensional feature space; the dynamic spatiotemporal attention layer calculates and weights attention weights along the time and sensor space dimensions; the residual memory layer processes the weighted features through a gated recurrent unit network regulated by a drift-diffusion joint state update mechanism; the degradation trajectory generation layer outputs multiple predicted degradation trajectories and their predicted distributions through fully connected neural network mapping and Monte Carlo dropout method; Step S4: Based on the multiple predicted degradation trajectories and their predicted distributions, calculate the remaining service life distribution function for each key device, and present the predicted degradation trajectory, current health indicators, and remaining service life confidence interval of each device in a visualized 3D twin scene.

[0114] This invention significantly improves data processing accuracy and the accuracy and reliability of equipment degradation prediction by constructing independent digital twins of equipment, optimizing the elastic alignment mechanism of multi-source temporal data, designing a dynamic spatiotemporal attention-residual memory network prediction architecture, and introducing a drift-diffusion joint state update mechanism. It achieves global and accurate prediction by adding a cross-twin coupling field analysis module to quantify the implicit degradation impact between equipment. It enables adaptive model updates based on the online evolution and incremental learning modules of the twins, adapting to the long-term dynamic degradation process of equipment. It generates intelligent predictive maintenance strategies using deep reinforcement learning, optimizing the allocation of operation and maintenance resources. Simultaneously, it presents the prediction results and health status through 3D twin scene visualization, intuitively supporting operation and maintenance decisions. This invention effectively solves the pain points of existing technologies, such as low data processing accuracy, homogeneous twin modeling, weak spatiotemporal feature capture, neglect of equipment coupling, difficulty in model adaptation, and unintelligent maintenance strategies. It provides a high-precision, adaptive, and globally collaborative full lifecycle predictive management solution for key equipment in smart factories.

[0115] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0116] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart factory lifecycle prediction and management system based on digital twins, characterized in that, include: The data acquisition and preprocessing module is configured to acquire multi-source status data in real time from sensor groups on key equipment in each production link of the smart factory, and perform time synchronization, outlier removal and normalization processing on the multi-source status data to generate a standardized time-series data stream. An independent digital twin construction module is configured to construct an independent digital twin for each of the key devices. Each digital twin includes a geometric model that has a geometric mapping to the physical entity of the key device, a physical model that describes its physical field distribution, a behavioral model that characterizes its current working condition and behavior, and a core prediction engine. The core prediction engine is based on a dynamic spatiotemporal attention-residual memory network architecture. The degradation trajectory prediction module, integrated within each independent digital twin, is configured to receive the standardized time-series data stream and sequentially execute the following: a feature embedding layer maps to a high-dimensional feature space; a dynamic spatiotemporal attention layer calculates and weights attention weights along the time and sensor space dimensions; a residual memory layer processes the weighted features through a gated recurrent unit network regulated by a drift-diffusion joint state update mechanism; and a degradation trajectory generation layer outputs multiple predicted degradation trajectories and their predicted distributions through a fully connected neural network mapping and Monte Carlo dropout method. The lifecycle prediction output module is configured to calculate the remaining service life distribution function of each key device based on the multiple predicted degradation trajectories and their predicted distributions, and to present the predicted degradation trajectory, current health indicators and remaining service life confidence interval of each device in a twin 3D scene in a visual form.

2. The intelligent factory lifecycle prediction and management system based on digital twins according to claim 1, characterized in that, The data acquisition and preprocessing module also includes a multi-source data alignment submodule. The multi-source data alignment submodule performs elastic alignment based on dynamic time warping on two sensor data sequences with different sampling frequencies. After calculating the point-to-point distance cost matrix of the two sensor data sequences, a bending penalty term based on the local slope of the sequence is introduced. The bending penalty term is obtained by multiplying the absolute value of the difference in the local rate of change of the two sensor data sequences at the corresponding positions by an adaptive penalty coefficient that is negatively correlated with the amplitude of the sequence fluctuation.

3. The intelligent factory lifecycle prediction and management system based on digital twins according to claim 2, characterized in that, The physical model adopts a data-driven enhanced reduced-order physical model. The construction process of the reduced-order physical model is as follows: a full-order finite element model is established based on the physical mechanism of the key equipment; the main energy modes of the full-order model are extracted by the intrinsic orthogonal decomposition method using historical full-life cycle data, and the first few order modes are retained to form the reduced-order basis; a shallow neural network is trained as a residual corrector. The input of the residual corrector is the prediction error of the reduced-order model at the key measurement points and its corresponding operating parameters, and the output is the real-time correction amount of the output of the reduced-order model.

4. The intelligent factory lifecycle prediction and management system based on digital twins according to claim 3, characterized in that, The process of calculating attention weights along the time dimension in the dynamic spatiotemporal attention layer is as follows: The original attention score is calculated by transforming the feature vector at the current moment and the feature vector at the historical moment using a bilinear transformation matrix; a time-interval-based decay factor is introduced, which is obtained by multiplying the time difference between the current moment and the historical moment by a learnable parameter characterizing the memory strength during the degradation process; the original attention score is multiplied element-wise by the decay factor and then normalized using Softmax to obtain the temporal attention weight, which is mathematically expressed as: ； Where t is the index of the current time, s is the index of the historical time, and j is the summation index. The time attention weight of the current time t to the historical time s; and These are the feature vectors at the current time and the historical time, respectively, with the superscript T indicating vector transpose; It is a learnable bilinear transformation matrix; This is a learnable memory strength parameter, with a value ranging from 0 to 1.

5. The intelligent factory lifecycle prediction and management system based on digital twins according to claim 4, characterized in that, The drift-diffusion joint state update mechanism is implemented by defining the latent health state variable of the device at time t. Its update process follows discretized stochastic differential equations; the drift term is obtained by processing the input features of the residual memory layer through an adaptive linear layer, and then adding a correction amount to adjust the matching degree between the previous state value and the current observation information; the diffusion term is obtained by extracting the time-varying noise intensity coefficient by performing statistical characteristic analysis on the high-frequency components in the multi-source state data at the current time, and then multiplying it by a standard Gaussian random variable; latent variables The summation of the drift term and the diffusion term, plus the latent variable from the previous time step, is then performed. The mathematical expression for obtaining is: ； Where t is the discrete time step index. Let be the latent variable of health status at time t. For the latent variable of health status at the previous moment, For the multi-source observation characteristics at time t, Let θ be the drift term function, and θ be the learnable model parameter vector within the drift term function. Let be the diffusion term function, which passes through the vibration signal at time t. Envelope spectrum entropy and acoustic emission signal The output noise intensity is calculated by combining the energy release rate. These are random numbers from a standard normal distribution. is the discrete time step.

6. The intelligent factory lifecycle prediction and management system based on digital twins according to claim 5, characterized in that, It also includes a cross-digital twin coupling field analysis module; the cross-digital twin coupling field analysis module calculates and maintains a coupling field matrix describing the implicit interactions between all independent digital twins; each element in the coupling field matrix represents the degradation influence intensity between two key equipment digital twins, the degradation influence intensity is obtained by dynamically time warping the operating condition sequences of the two equipment to calculate similarity, and then multiplying it by a decay coefficient characterizing the topological distance between the two in the production process; the cross-digital twin coupling field analysis module inputs the coupling field matrix as an additional feature to the degradation trajectory prediction module of each independent digital twin in each prediction cycle, specifically by concatenating it with the feature tensor before the dynamic spatiotemporal attention layer.

7. The intelligent factory lifecycle prediction and management system based on digital twins according to claim 6, characterized in that, The lifecycle prediction output module also includes a competitive failure mode discriminator; the competitive failure mode discriminator detects the time point at which each of the multiple predicted degradation trajectories first crosses a preset failure threshold within a preset prediction time window; it clusters all trajectories according to different physical failure modes; it calculates the kernel density estimate of the remaining lifetime for each group to obtain the remaining lifetime probability distribution under that failure mode; it determines the failure mode that occurs first and whose probability density peak exceeds a preset confidence threshold as the dominant failure mode, and outputs the degradation characteristics and highlighted data of the predicted trajectory corresponding to the dominant failure mode.

8. The intelligent factory lifecycle prediction and management system based on digital twins according to claim 7, characterized in that, It also includes a predictive maintenance strategy self-generation module; this module receives the remaining useful life distribution function of each key device and preset maintenance resource constraints; the predictive maintenance strategy self-generation module embeds a scheduling agent based on deep reinforcement learning, the state space of which includes the current health indicators, the mean and variance of the remaining useful life prediction, and the occupancy of maintenance resources for each device; the action space of the scheduling agent consists of combinations of actions of different maintenance levels assigned to each device; the reward function of the scheduling agent is a weighted sum of the total downtime cost, preventive maintenance cost, and production loss cost caused by sudden failure of all devices, multiplied by a risk penalty term inversely proportional to the uncertainty of the remaining useful life prediction; the scheduling agent is trained through interaction with historical simulations of the environment and outputs a list of collaborative maintenance tasks.

9. The intelligent factory lifecycle prediction and management system based on digital twins according to claim 8, characterized in that, It also includes a twin online evolution and incremental learning module; when the prediction error between the health index output by the degradation trajectory prediction module and the subsequent actual collected observation data feedback continuously exceeds a preset dynamic threshold, the twin online evolution and incremental learning module triggers an incremental update of the core prediction engine inside the independent digital twin; the incremental update adopts an elastic weight consolidation algorithm, which applies a secondary penalty term when calculating the loss gradient of the new batch of data; the secondary penalty term is obtained by multiplying the square of the change in each historical weight in the core prediction engine by a regularization coefficient determined according to the diagonal elements of the Fisher information matrix.

10. A method for predictive management of the lifecycle of a smart factory based on digital twins, characterized in that: The intelligent factory lifecycle prediction management system based on digital twins as described in any one of claims 1-9 includes the following steps: Step S1: Collect multi-source status data in real time from the sensor groups on key equipment in each production link of the smart factory, and perform time synchronization, outlier removal and normalization processing on the multi-source status data to generate a standardized time-series data stream. Step S2: Construct an independent digital twin for each key device. Each digital twin includes a geometric model that has a geometric mapping to the physical entity of the key device, a physical model that describes its physical field distribution, a behavioral model that characterizes its current working condition and behavior, and a core prediction engine. The core prediction engine is based on a dynamic spatiotemporal attention-residual memory network architecture. Step S3: Receive the standardized time-series data stream and execute the following steps in sequence: the feature embedding layer maps to a high-dimensional feature space; the dynamic spatiotemporal attention layer calculates and weights attention weights along the time and sensor space dimensions; the residual memory layer processes the weighted features through a gated recurrent unit network regulated by a drift-diffusion joint state update mechanism; and the degradation trajectory generation layer outputs multiple predicted degradation trajectories and their predicted distributions through a fully connected neural network mapping and Monte Carlo dropout method. Step S4: Based on multiple predicted degradation trajectories and their predicted distributions, calculate the remaining useful life distribution function for each key device, and present the predicted degradation trajectory, current health indicators, and remaining useful life confidence interval of each device in a twin 3D scene in a visual format.

Images