Equipment fault prediction method and system based on digital twinning

Abstract

The invention relates to the technical field of artificial intelligence and digital twinning crossing, and discloses an equipment fault prediction method and system based on digital twinning. Comprising the steps of collecting multi-dimensional physical data and full life cycle data of equipment, constructing a digital twin model and synchronizing physical states in real time, quantifying implicit coupling interference through three-level linkage nonlinear modeling, dynamically correcting features, calculating fault evolution probability, and optimizing parameters through reinforcement learning. The system correspondingly comprises five functional modules. The method solves the problem of inaccurate prediction caused by neglect of multi-physics field coupling interference and feature processing lag in the prior art, realizes accurate pre-judgment and early warning of faults, and is suitable for a full-life-cycle monitoring scene of industrial equipment.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and digital twin technology, and in particular to a method and system for predicting equipment failures based on digital twins. Background Technology

[0002] During long-term operation, industrial equipment's core components are subjected to a combination of physical factors, including electromagnetic radiation, temperature variations, and load fluctuations. These factors are not isolated but form complex, implicitly coupled interference relationships. Most existing equipment fault prediction technologies only monitor and analyze single physical quantities, failing to fully consider the implicit coupling effects between multiple physical fields, resulting in incomplete identification of fault causes. Furthermore, existing technologies are insufficiently adaptable to dynamic changes in equipment operating conditions, employing fixed and singular feature processing methods that cannot be dynamically adjusted based on real-time interference. This leads to low fidelity of extracted fault features and the inclusion of significant redundant noise. These two major problems directly result in existing fault prediction technologies exhibiting predictive lag and low accuracy, making it difficult to identify early-stage equipment faults. Warnings are often only issued after a fault has already occurred or is about to occur, leading to substantial economic losses and safety risks in industrial production.

[0003] Based on the above problems, there is an urgent need for a device fault prediction technology that can accurately quantify the implicit coupling interference of multi-physics fields and dynamically optimize the feature processing process. Summary of the Invention

[0004] This application proposes a device fault prediction method based on digital twins, which includes collecting multi-dimensional physical data and full lifecycle data during device operation, constructing a digital twin model of the device and synchronizing the device's physical state in real time, performing fault-related feature processing and predictive analysis based on the digital twin model, and outputting the probability of fault occurrence and evolution trend. The feature processing and predictive analysis process adopts a three-level linkage nonlinear modeling approach. First, the implicit coupling interference of multi-physics fields during device operation is quantified. Then, the original features are dynamically corrected based on the interference quantification results. Finally, the corrected features are fused with the device's full lifecycle correlation data to calculate the fault evolution probability. The entire process uses artificial intelligence reinforcement learning algorithms to dynamically optimize modeling parameters, achieving accurate fault prediction.

[0005] Preferably, the multi-dimensional physical data includes the electromagnetic radiation intensity, temperature gradient, continuous operating time, real-time load fluctuation rate, and sensor attenuation coefficient of key components of the equipment. The full life cycle data includes the cumulative operating time of the equipment, the number of maintenance operations, the material fatigue coefficient, and historical fault characteristic data. All data are collected through distributed sensors and data acquisition interfaces, and then transmitted to the processing unit after digital conversion.

[0006] Preferably, the digital twin model achieves real-time synchronization with the physical state of the device through a two-way interactive channel. The synchronized content includes device operating parameters, physical quantity change data, aging status data, and feature correlation data. The synchronization frequency is consistent with the data acquisition frequency to ensure the consistency between the model and the physical device's state.

[0007] Preferably, the artificial intelligence reinforcement learning algorithm uses the fault prediction error as the reward function, and the reward function value is negatively correlated with the prediction error. The algorithm dynamically adjusts key parameters such as coupling coefficient, correction weight coefficient, and probability amplification coefficient in the modeling process through iterative training, so that the prediction result continuously approaches the actual fault state.

[0008] Preferably, the multiphysics implicit coupling interference quantization is achieved through the following mathematical formula:

[0009] ;

[0010] in The implicit coupling interference coefficient is dimensionless. The mean electromagnetic radiation intensity of the key components of the equipment is expressed in V / m. The temperature gradient change rate between the core component and the environment is expressed in °C / s. The continuous operating time of the equipment is measured in hours (h). is the electromagnetic-temperature coupling coefficient, with dimensions (m / V)². This is the duration decay coefficient, with dimensions 1 / h; and This is achieved through reinforcement learning training.

[0011] Preferably, the dynamic feature correction is achieved through the following mathematical formula:

[0012] ;

[0013] in To achieve high-fidelity features after correction, The original features acquired by the sensor For real-time load volatility, The sensor attenuation coefficient, To correct the weighting coefficients, and It is updated in real time through digital twin models.

[0014] Preferably, the fault evolution probability calculation is achieved through the following mathematical formula:

[0015] ;

[0016] in This represents the probability of a failure occurring, with a value ranging from 0 to 1. This refers to the equipment aging index. The trace is the cosine similarity matrix between the current features and historical fault features. This is the probability amplification factor. It is calculated by combining cumulative running time, number of maintenance operations, and material fatigue coefficient. Quantify multi-dimensional correlations through matrix operations.

[0017] Preferably, the fault evolution probability calculation further includes generating short-term and long-term fault evolution curves. The short-term fault evolution curve corresponds to the fault development trend from 1 to 24 hours, and the long-term fault evolution curve corresponds to the fault development trend from 1 to 30 days. The curve generation process combines the changing patterns of equipment operating conditions and historical fault evolution data.

[0018] Preferably, the output fault occurrence probability and evolution trend include triggering graded early warning signals. The early warning signals are divided into level 1, level 2 and level 3 early warnings, which correspond to different fault occurrence probability ranges and emergency handling priorities, respectively. The early warning signals are displayed through a visual interface and synchronized to the device management terminal.

[0019] Preferably, a device fault prediction system based on digital twins includes an implicit coupling interference perception module, a dynamic feature correction module, a fault evolution prediction module, a digital twin mapping interaction module, and a reinforcement learning optimization module. The implicit coupling interference perception module is used to collect multi-dimensional physical data and quantify the implicit coupling interference coefficients. The dynamic feature correction module is used to correct the original features based on the interference coefficients. The fault evolution prediction module is used to calculate the probability of fault occurrence and the evolution trend. The digital twin mapping interaction module is used to achieve bidirectional synchronization between the physical device and the digital twin model. The reinforcement learning optimization module is used to dynamically adjust the key modeling parameters. Each module realizes data interaction and collaborative work through a data bus.

[0020] The technical advantages of this invention lie in the deep integration of three-level linkage nonlinear modeling and artificial intelligence reinforcement learning. By quantifying the implicit coupling interference of multi-physics fields, it solves the problem of incomplete fault cause identification caused by neglecting coupling effects in existing technologies; based on dynamic correction features of interference results, it addresses the shortcomings of fixed feature processing methods, such as poor adaptability and low fidelity; and by integrating full lifecycle data to calculate the fault evolution probability, it achieves accurate prediction of fault trends. The synergistic effect of these three technologies effectively solves the core problems of lagging and low accuracy in equipment fault prediction in the background technologies, providing a reliable technical means for early fault identification of industrial equipment and ensuring production continuity and safety.

[0021] The beneficial effects of this invention are as follows:

[0022] This application provides a digital twin-based equipment fault prediction method and system. It addresses the problem of incomplete fault cause identification caused by neglecting coupling effects in existing technologies by quantifying implicit coupling interference from multi-physics fields. It overcomes the shortcomings of fixed feature processing methods, such as poor adaptability and low fidelity, by dynamically correcting features based on interference results. Furthermore, it integrates full lifecycle data to calculate fault evolution probability, achieving accurate fault trend prediction. These three key technologies work synergistically to effectively solve the core problems of lagging and low accuracy in equipment fault prediction in the background technologies, providing a reliable technical means for early fault identification of industrial equipment and ensuring production continuity and safety. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort:

[0024] Figure 1 This is a flowchart of the equipment fault prediction method based on digital twins in this application;

[0025] Figure 2 This is a connection block diagram of the equipment fault prediction system based on digital twins in this application. Detailed Implementation

[0026] 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.

[0027] Traditional fault prediction technology has the following technical problems: it ignores the implicit coupling interference between multiple physical fields during equipment operation, only monitors and analyzes a single physical quantity, and the feature processing method is fixed and cannot adapt to dynamic changes in operating conditions, resulting in delayed fault prediction and low accuracy.

[0028] Based on this, please refer to Figure 1 This embodiment provides a device fault prediction method based on digital twins, including:

[0029] S1: Collect multi-dimensional physical data and full lifecycle data during equipment operation;

[0030] S2: Build a digital twin model of the equipment and synchronize the physical status of the equipment in real time;

[0031] S3: Fault-related feature processing and predictive analysis based on digital twin models;

[0032] S4: Output failure probability and evolution trend.

[0033] The feature processing and predictive analysis process adopts a three-level linkage nonlinear modeling approach. First, the implicit coupling interference of multiple physical fields during equipment operation is quantified. Then, the original features are dynamically corrected based on the interference quantification results. Finally, the corrected features are fused with the equipment's life-cycle correlation data to calculate the probability of fault evolution. The entire process uses artificial intelligence reinforcement learning algorithms to dynamically optimize modeling parameters and achieve accurate fault prediction.

[0034] The core of this technical solution lies in the collaborative design of three-level linkage nonlinear modeling and reinforcement learning. Multi-dimensional physical data is acquired through distributed sensor deployment, encompassing electromagnetic radiation sensors, temperature sensors, runtime statistics modules, load monitoring units, and sensor performance monitoring modules. Full lifecycle data is retrieved through the equipment management system, including cumulative runtime records, maintenance archives, material fatigue test data, and a historical fault feature database. The digital twin model is constructed based on the three-dimensional structure and operating principles of the physical equipment. It acquires various data through a real-time data transmission channel, dynamically updating the equipment state parameters in the model to ensure real-time synchronization between the model and the physical equipment. In the three-level linkage modeling process, the first level quantifies the coupling interference between electromagnetic radiation and temperature gradients using a specific mathematical model. The second level dynamically adjusts the original acquired features based on the interference quantification results, combined with load fluctuations and sensor attenuation. The third level fuses the corrected high-fidelity features with data such as equipment aging status and historical fault correlation to calculate the probability of fault occurrence. The artificial intelligence reinforcement learning algorithm uses prediction error as a reward signal. Through continuous iterative training, it dynamically adjusts various key parameters in the modeling process, continuously optimizing the prediction model and improving prediction accuracy.

[0035] It is worth mentioning that the various types of data collected above support the twin model in the following ways: 3D model calibration: The collected data on the dimensions and materials of key equipment components are used to build the physical basis of the twin model to ensure structural consistency; Real-time status updates: Operating parameters and physical quantity change data are transmitted through a two-way channel to dynamically update the equipment operating status in the model; Aging status synchronization: Data such as cumulative running time and number of maintenance are used to calculate the equipment aging index and synchronize it to the aging status module of the model.

[0036] The data support for the feature processing in step S3 in step S2 is provided in the following ways: Directly provide dynamic parameters: synchronize real-time load fluctuation rate (L) and sensor attenuation coefficient (S) as the core input for dynamic feature correction; output coupling interference related data: based on synchronized electromagnetic radiation intensity (E) and temperature gradient (T), the implicit coupling interference coefficient (I) is calculated through the interference sensing module associated in S2; associate full life cycle data: synchronize equipment aging index (A) and historical fault characteristic data to provide basic parameters for fault evolution probability calculation.

[0037] The technical effects achieved by the above-mentioned technical solutions include: accurate quantification of implicit coupling interference of multi-physics fields, solving the limitations of monitoring a single physical quantity; dynamic feature correction mechanism improves the fidelity of fault features and eliminates redundant noise; reinforcement learning-driven parameter optimization enables the model to adapt to the dynamic changes in equipment operating conditions, significantly improving the accuracy and timeliness of fault prediction, and providing reliable support for early fault warning of equipment.

[0038] Traditional data acquisition techniques suffer from the following technical problems: incomplete data acquisition dimensions, lack of systematic acquisition of key physical quantities for equipment fault prediction, and distortion during data transmission and conversion, which affects the accuracy of subsequent analysis.

[0039] Based on this, the multi-dimensional physical data includes the electromagnetic radiation intensity, temperature gradient, continuous operating time, real-time load fluctuation rate, and sensor attenuation coefficient of key components of the equipment. The full life cycle data includes the cumulative operating time of the equipment, the number of maintenance operations, the material fatigue coefficient, and historical fault characteristic data. All data are collected through distributed sensors and data acquisition interfaces, and after being digitally converted, they are transmitted to the processing unit.

[0040] In this technical solution, multi-dimensional physical data acquisition is achieved through a specially designed distributed sensor network. Electromagnetic radiation intensity is collected by electromagnetic sensors deployed around key components of the equipment. These sensors employ a wideband design, capable of capturing electromagnetic radiation signals across different frequency ranges. The acquired analog signals are converted into digital signals by an analog-to-digital converter. Temperature gradients are collected by a group of temperature sensors installed on the surface of the core components and in the surrounding environment, calculating the rate of change of the temperature difference between the core components and the environment over time. Continuous operating time is recorded by a timing module in the equipment control system, showing the continuous operating time of the equipment from startup to the current moment, in hours. Real-time load fluctuation rate is calculated by collecting real-time operating load data from the load monitoring unit, calculating the load fluctuation amplitude per unit time. Sensor attenuation coefficients are obtained by periodically testing the output accuracy of each sensor through a sensor performance monitoring module, comparing it with the initial calibration accuracy. The cumulative operating time in the full lifecycle data is extracted from the equipment operation log; the number of repairs and repair details are retrieved from the equipment maintenance records; the material fatigue coefficient is calculated based on the material properties of the core components and the cumulative operating time; and historical fault characteristic data is extracted from feature parameters in the equipment's historical fault records. All collected physical data and lifecycle data are aggregated through a unified data acquisition interface, and after digital conversion and format standardization, they are transmitted to the central processing unit to provide high-quality data support for subsequent modeling and analysis.

[0041] The technical effects achieved by the above-mentioned technical solution include: comprehensively covering the key data dimensions required for equipment fault prediction, solving the problem of single dimensions in traditional acquisition schemes; the distributed acquisition method ensures the comprehensiveness and accuracy of data acquisition, and digital conversion and standardization processing reduce data distortion; and it provides a high-quality data foundation for subsequent coupling interference quantification, feature correction and fault prediction, thereby improving the reliability and accuracy of the entire prediction system.

[0042] Traditional digital twin models have the following technical problems: insufficient synchronization between the model and the physical device, and incomplete synchronization content, which makes it impossible for the model to truly reflect the real-time operating status of the device and affects the model-based fault prediction effect.

[0043] Based on this, the digital twin model achieves real-time synchronization with the physical state of the equipment through a two-way interactive channel. The synchronized content includes equipment operating parameters, physical quantity change data, aging status data, and feature correlation data. The synchronization frequency is consistent with the data acquisition frequency to ensure the consistency between the model and the physical equipment.

[0044] In this technical solution, the bidirectional interactive channel of the digital twin model is constructed using a high-speed data transmission protocol, including a data upload channel and a command issuance channel. The data upload channel transmits collected equipment operating parameters such as speed, voltage, and current; physical quantity change data such as changes in electromagnetic radiation intensity and temperature gradient; aging status data such as equipment aging index and component wear degree; and feature correlation data such as the correlation between current features and historical fault features to the digital twin model in real time. The synchronization frequency is set to the same as the data acquisition frequency to ensure that the model can obtain the latest status data of the equipment in a timely manner. After receiving the data, the digital twin model adjusts the corresponding parameters in the model in real time using a built-in state update algorithm, updating the virtual operating status of the equipment. The command issuance channel transmits adjustment commands obtained from model analysis, such as sensor acquisition frequency adjustment commands and equipment operating parameter optimization suggestions, to the equipment's control system, enabling the model to reverse control and optimize the physical equipment. To ensure the reliability of synchronization, the bidirectional interactive channel adopts a redundant design; when the main channel fails, it automatically switches to the backup channel to avoid data transmission interruption. At the same time, a data verification mechanism is set up to verify the integrity and accuracy of the transmitted data, and to trigger a retransmission mechanism in a timely manner when data anomalies are detected.

[0045] It should be noted that when data transmission delays on one side cause inconsistencies in the state, the following solutions can be used to resolve the issue:

[0046] Delay detection and graded processing: Delays are identified by comparing timestamps. Real-time retransmission is triggered when the delay is ≤1s. When the delay is >1s, a cache queue is started to temporarily store data. After recovery, the data is synchronized in batches according to time order.

[0047] Automatic switching of redundant channels: When the delay of the main transmission channel exceeds the threshold, the system automatically switches to the backup channel, while retaining the fault log of the main channel for easy troubleshooting.

[0048] Predictive data completion: Based on historical synchronization data and equipment operation patterns, complete data for delayed periods is generated through a short-term prediction model, and error correction is performed after the actual data is transmitted.

[0049] The technical effects achieved by the above-mentioned technical solution include: the two-way interactive channel realizes real-time data interaction between the model and the physical device, ensuring that the model can truly reflect the real-time operating status of the device; comprehensive synchronization content provides complete status data support for fault prediction; the synchronization frequency is consistent with the acquisition frequency, avoiding prediction errors caused by data delay; redundant design and data verification mechanism improve the reliability and stability of the synchronization process, ensuring the continuous operation of fault prediction work.

[0050] Traditional parameter optimization techniques suffer from the following technical problems: fixed modeling parameters or rigid adjustment methods make it impossible to dynamically optimize based on the actual error of fault prediction, resulting in poor model adaptability and difficulty in continuously improving prediction accuracy.

[0051] Based on this, the artificial intelligence reinforcement learning algorithm uses the fault prediction error as the reward function. The reward function value is negatively correlated with the prediction error. The algorithm dynamically adjusts key parameters such as coupling coefficient, correction weight coefficient, and probability amplification coefficient in the modeling process through iterative training, so that the prediction result continuously approaches the actual fault state.

[0052] In this technical solution, the artificial intelligence reinforcement learning algorithm is constructed using a deep deterministic policy gradient algorithm framework. The reward function is designed around the fault prediction error, which is calculated as the deviation between the fault occurrence probability output by the model and the actual fault state of the equipment. The smaller the prediction error, the larger the reward function value, and vice versa. The algorithm's agent continuously explores the optimal parameter adjustment strategy through interaction with the fault prediction environment. Key parameters in the modeling process, such as the electromagnetic-temperature coupling coefficient α, the duration decay coefficient β, the correction weight coefficient γ, and the probability amplification coefficient δ, serve as the agent's action space. The agent executes parameter adjustment actions, observes changes in the prediction error (i.e., environmental feedback), and adjusts subsequent parameter adjustment strategies based on the reward function value. During iterative training, the agent stores past interaction experiences through an experience replay mechanism, including parameter adjustment values, prediction error changes, and reward function values. Random sampling of this experience data is used for model training, improving the algorithm's stability and convergence speed. Meanwhile, an exploration rate decay mechanism is set up. A higher exploration rate is used in the early stage of training to encourage the agent to try different parameter combinations. In the later stage of training, the exploration rate is gradually reduced so that the agent can focus on the optimal parameter region for fine-tuning.

[0053] The technical effects achieved by the above-mentioned technical solutions include: the reinforcement learning algorithm realizes dynamic optimization of key modeling parameters, solving the problem of poor adaptability of fixed parameters or rigid adjustment methods; the reward function is directly linked to the prediction error, ensuring that parameter adjustments are always made in the direction of improving prediction accuracy; the experience replay mechanism and the exploration rate decay mechanism improve the training efficiency and stability of the algorithm, enabling the model to continuously approach the optimal prediction state, and significantly improving the accuracy and adaptability of fault prediction.

[0054] Traditional interference quantification techniques have the following technical problems: they cannot accurately quantify the implicit coupling interference between multiple physical fields, and can only consider the influence of a single physical field, resulting in inaccurate interference assessment, which in turn affects the subsequent feature processing and fault prediction effects.

[0055] Based on this, the quantization of the multiphysics implicit coupling interference is achieved through the following mathematical formula:

[0056] ;

[0057] in The implicit coupling interference coefficient is dimensionless. The mean electromagnetic radiation intensity of the key components of the equipment is expressed in V / m. The temperature gradient change rate between the core component and the environment is expressed in °C / s. The continuous operating time of the equipment is measured in hours (h). is the electromagnetic-temperature coupling coefficient, with dimensions (m / V)². This is the duration decay coefficient, with dimensions 1 / h; and The values ​​were obtained through reinforcement learning training, where α was initially 0.002 (m / V)²; β was initially 0.011 / h, corresponding to the interference attenuation pattern during 100 hours of continuous operation of the equipment; γ was initially 0.05, balancing the combined effects of coupling interference and load and sensor attenuation; and δ was initially 0.3, ensuring that the initial calculation of the failure probability was within a reasonable range (0-0.5). Subsequently, through reinforcement learning optimization, T∈(-1℃ / s,+∞), that is, the temperature gradient change rate is not lower than -1℃ / s.

[0058] In this technical solution, the mathematical formula for quantifying implicit coupling interference is designed based on the coupling mechanism of electromagnetic radiation and temperature gradient, while also considering the attenuation effect of continuous equipment operation time on interference intensity. The average electromagnetic radiation intensity E is obtained by collecting the electromagnetic radiation intensity from multiple monitoring points around key components of the equipment using electromagnetic sensors, and then taking the arithmetic mean, reflecting the overall intensity of the electromagnetic environment around the key components. The temperature gradient change rate T is obtained by calculating the difference between the surface temperature of the core component and the ambient temperature, and then taking the time derivative of this difference, reflecting the impact of the rate of temperature change on coupling interference. The continuous equipment operation time t is statistically analyzed in real time by a timing module. As the operation time increases, the anti-interference capability of the equipment components gradually changes, which is reflected by an exponential function. To describe this attenuation effect, when t is 0, the term is 0; as t increases, it gradually approaches 1, consistent with the attenuation law of interference in actual operation. The electromagnetic-temperature coupling coefficient α is used to adjust the contribution weight of electromagnetic radiation intensity to coupled interference, and the duration attenuation coefficient β is used to adjust the influence of runtime on interference attenuation. Both are dynamically optimized based on prediction error using reinforcement learning algorithms to ensure the accuracy of interference quantification. In the actual calculation process, the dimensions of each physical quantity are first standardized, and then substituted into the formula to calculate the implicit coupled interference coefficient I. The value of I ranges from 0 to 10; a larger value indicates stronger coupled interference.

[0059] The technical effects achieved by the above-mentioned technical solution include: accurate quantification of implicit coupling interference of multi-physics fields, solving the deficiency of traditional technology in being unable to assess coupling effects; the formula takes into account the comprehensive effects of electromagnetic radiation, temperature gradient and running time, and the quantification results are more in line with the actual situation; the coupling coefficient and attenuation coefficient are dynamically optimized through reinforcement learning, which improves the adaptability and accuracy of interference quantification and provides a reliable basis for subsequent feature correction.

[0060] Traditional feature processing techniques have the following technical problems: the feature correction method is fixed and does not take into account the combined effects of implicit coupling interference, load fluctuation and sensor attenuation, resulting in low feature fidelity after correction and the presence of a large amount of redundant noise, which affects the accuracy of fault prediction.

[0061] Based on this, the dynamic feature correction is achieved through the following mathematical formula:

[0062] ;

[0063] in For high-fidelity features after correction, dimensionless; These are the raw, dimensionless features acquired by the sensor. This represents the real-time load volatility, which is dimensionless. The sensor attenuation coefficient is dimensionless. The weighting coefficients are dimensionless to correct for weighting. and It is updated in real time through digital twin models.

[0064] In this technical solution, the core of the dynamic feature correction mathematical formula lies in incorporating the implicit coupling interference coefficient I, real-time load fluctuation rate L, and sensor attenuation coefficient S into the correction process to achieve dynamic adjustment of the original features. The original sensor acquisition feature F0 represents unprocessed fault-related features acquired by various sensors, such as vibration signal features and current signal features, all of which have been normalized and range from 0 to 1. The real-time load fluctuation rate L is obtained by calculating the difference between the maximum and minimum load values ​​of the equipment per unit time, divided by the rated load, reflecting the degree of fluctuation in the equipment's operating load, and ranges from 0 to 0.5. The sensor attenuation coefficient S is obtained by comparing the current acquisition accuracy of the sensor with the initial calibration accuracy. The initial value is 1, and as the sensor's usage time increases, the accuracy decreases, and the value of S gradually decreases, ranging from 0.1 to 1. The correction weight coefficient γ is used to adjust the influence of comprehensive interference factors on feature correction, ranging from 0.01 to 0.1, and is dynamically optimized through a reinforcement learning algorithm. The arctangent function... This function is used to nonlinearly adjust the ratio of load fluctuation rate to sensor attenuation coefficient, making the correction process more consistent with the actual laws of characteristic changes. The value range of this function is from π / 4 to π / 2. During the correction process, the comprehensive interference factor is first calculated: Then add it to 1 and multiply it by the original feature F0 to obtain the corrected high-fidelity feature F. The value of F ranges from 0 to 2 to ensure the effectiveness and consistency of the feature.

[0065] It should be noted that when the sensor malfunctions, S approaches 0, which can be addressed in the following ways:

[0066] Failure threshold: The failure threshold for the sensor attenuation coefficient S is set to 0.05, that is, the sensor is considered to be faulty when S≤0.05.

[0067] Data replacement after failure:

[0068] Switching to backup sensors: If the device deploys redundant sensors, it will automatically call up the data from backup sensors of the same type to replace them.

[0069] Model prediction completion: When there is no backup sensor, the theoretical data of the sensor is predicted based on the device operating status in the digital twin model, and then corrected by combining historical data from the same period.

[0070] Failure warning trigger: After a sensor failure is detected, a level 2 warning is triggered simultaneously to remind maintenance personnel to replace the sensor in a timely manner.

[0071] The technical effects achieved by the above-mentioned technical solution include: the dynamic feature correction mechanism comprehensively considers the effects of implicit coupling interference, load fluctuation and sensor attenuation, and solves the problem of poor adaptability of traditional fixed correction methods; the accuracy of feature correction is improved by adjusting nonlinear functions and optimizing dynamic weight coefficients; the corrected high-fidelity features eliminate redundant noise, provide high-quality input for fault evolution probability calculation, and significantly improve the accuracy of fault prediction.

[0072] Traditional fault probability calculation techniques have the following technical problems: they calculate fault probability based on only a single feature or partial state data, without integrating corrected features, equipment aging status and historical fault correlation data, resulting in inaccurate prediction of fault evolution trends and inability to predict fault development in advance.

[0073] Based on this, according to the equipment fault prediction method based on digital twins as described in claim 6, the fault evolution probability calculation is achieved through the following mathematical formula:

[0074] ;

[0075] in The probability of failure occurrence is dimensionless and ranges from 0 to 1. The equipment aging index is dimensionless and ranges from 0 to 1. The trace is the cosine similarity matrix between the current feature and the historical fault features. It is dimensionless and ranges from 0 to n, where n is the feature dimension. This is the probability amplification factor, dimensionless, with a value range from 0.1 to 1; It is calculated by combining cumulative running time, number of maintenance operations, and material fatigue coefficient. Quantify multi-dimensional correlations through matrix operations.

[0076] In this technical solution, the mathematical formula for calculating the failure evolution probability integrates the corrected high-fidelity feature F, the equipment aging index A, the correlation degree of historical failure features C, and the probability amplification coefficient δ, comprehensively considering the key factors affecting the occurrence of failures. The equipment aging index A is calculated using a weighted summation method, with cumulative runtime weighted at 0.4, maintenance frequency weighted at 0.3, and material fatigue coefficient weighted at 0.3. All data are normalized before summing; a larger A value indicates a more severe degree of equipment aging. The trace C of the cosine similarity matrix between the current feature and historical failure features is obtained by constructing the cosine similarity matrix between the current corrected feature vector and the historical failure feature vector library, and calculating the trace of this matrix. The value of the trace reflects the overall correlation between the current feature and historical failure features; a larger C value indicates a more similar current equipment state to historical failure states. The probability amplification coefficient δ is used to adjust the amplification effect of comprehensive influencing factors on the failure probability, and is dynamically optimized based on the prediction error using a reinforcement learning algorithm. In the formula, This is used to describe the exponential effect of characteristics and aging conditions on the probability of failure. The contribution of historical fault correlation to the fault probability is corrected by multiplying the two factors and then subtracting 1 to obtain the final fault occurrence probability P. When F, A, and C are all small, P is close to 0, indicating an extremely low fault occurrence probability; as F, A, and C increase, P gradually approaches 1, indicating an extremely high fault occurrence probability.

[0077] The technical effects achieved by the above-mentioned technical solution include: integrating multi-dimensional key data to calculate the probability of fault evolution, solving the problem of inaccurate calculation by traditional single factors; comprehensively considering feature fidelity, equipment aging status and historical fault correlation, improving the accuracy of fault probability calculation; effectively predicting fault evolution trends, providing a reliable basis for early warning, and significantly improving the practicality and effectiveness of fault prediction.

[0078] Traditional fault trend prediction technology has the following technical problems: it only outputs the probability of fault occurrence, does not provide fault evolution curves at different time scales, cannot intuitively display the fault development trend, and is not conducive to equipment maintenance personnel to formulate targeted maintenance strategies.

[0079] Based on this, the fault evolution probability calculation also includes generating short-term and long-term fault evolution curves. The short-term fault evolution curve corresponds to the fault development trend from 1 to 24 hours, and the long-term fault evolution curve corresponds to the fault development trend from 1 to 30 days. The curve generation process combines the changing patterns of equipment operating conditions and historical fault evolution data.

[0080] In this technical solution, the generation of fault evolution curves is based on time-series predictions of fault occurrence probabilities. The short-term fault evolution curve is generated using a sliding window prediction method. Starting from the current moment, based on fault probability data from the past hour and real-time collected equipment operating parameters, a short-term prediction model, such as a temporal convolutional network, predicts the fault occurrence probability for each moment in the next 24 hours hour by hour. The predicted probability values ​​are then connected in chronological order to form the short-term fault evolution curve. The long-term fault evolution curve is generated by combining changes in equipment operating conditions, such as production plans and load change cycles, with historical fault evolution data. A long-term prediction model, such as a long short-term memory network based on an attention mechanism, predicts the average fault occurrence probability for each day over the next 30 days. The daily average probability values ​​are then connected in chronological order to form the long-term fault evolution curve. During curve generation, the phased characteristics of fault development are incorporated, such as slow probability growth in the early fault stage, rapid probability growth in the mid-term fault stage, and a probability approaching 1 in the late fault stage, making the curve more consistent with actual fault evolution patterns. Simultaneously, the prediction results are smoothed to eliminate abnormal fluctuations and ensure the continuity and readability of the curve. The generated fault evolution curve is displayed through a visual interface, marking key time nodes and corresponding fault probability values, making it easy for maintenance personnel to intuitively grasp the fault development trend.

[0081] The technical effects achieved by the above-mentioned technical solutions include: short-term and long-term fault evolution curves intuitively show the fault development trend, solving the problem that traditional methods only output probability values ​​and lack trend references; curve generation combines operating condition patterns and historical data, improving the accuracy of trend prediction; and providing a reliable basis for maintenance personnel to formulate short-term emergency response plans and long-term maintenance plans, thereby improving the pertinence and effectiveness of equipment maintenance.

[0082] Traditional early warning technologies have the following technical problems: the early warning signal is singular and the early warning level is not divided according to the probability of the fault occurrence, which makes it impossible for maintenance personnel to judge the urgency of the fault, making it difficult to prioritize the handling of high-risk faults and affecting maintenance efficiency.

[0083] Based on this, the output failure probability and evolution trend include triggering graded early warning signals. The early warning signals are divided into level 1, level 2 and level 3 early warnings, which correspond to different failure probability ranges and emergency handling priorities, respectively. The early warning signals are displayed through a visual interface and synchronized to the equipment management terminal.

[0084] In this technical solution, the classification of early warning signals is determined based on a combination of the probability range of fault occurrence and the rate of fault development. Level 1 warnings correspond to a fault occurrence probability between 0.7 and 1.0, with a fault probability growth rate greater than 0.05 / hour. This is a high-urgency warning, indicating that the equipment is highly likely to experience a serious fault in the short term, and has the highest priority for emergency handling. Level 2 warnings correspond to a fault occurrence probability between 0.4 and 0.7, with a fault probability growth rate between 0.01 and 0.05 / hour. This is a medium-urgency warning, indicating that the equipment has potential for faults and requires inspection and maintenance within a specified time; its emergency handling priority is lower. Level 3 warnings correspond to a fault occurrence probability between 0.1 and 0.4, with a fault probability growth rate less than 0.01 / hour. This is a low-urgency warning, indicating that the equipment has potential fault risks and can be addressed within the regular maintenance cycle; its emergency handling priority is the lowest. The warning signals are displayed using a combination of color-coded visual indicators and sound prompts. Level 1 warnings are indicated by red and accompanied by a high-frequency alarm sound; Level 2 warnings by yellow and accompanied by a medium-frequency alarm sound; and Level 3 warnings by blue and accompanied by a low-frequency alarm sound. Simultaneously, the warning signals are synchronized to equipment management terminals such as mobile apps and computer clients via wireless communication modules, ensuring that maintenance personnel receive warning information in a timely manner. Each warning signal also includes a screenshot of the fault occurrence probability and evolution trend curve, as well as preliminary handling suggestions, providing decision support for maintenance personnel.

[0085] The technical effects achieved by the above-mentioned technical solution include: the graded early warning signal clarifies the urgency of the fault and the priority of handling, solving the problems of traditional early warning signals being single and low maintenance efficiency; the multi-channel early warning information push ensures that maintenance personnel receive it in a timely manner, avoiding delays in handling; the accompanying fault information and handling suggestions improve the pertinence and efficiency of maintenance work, effectively reducing losses caused by equipment failure.

[0086] Traditional fault prediction systems suffer from the following technical problems: unreasonable system module division, single function of each module and lack of collaborative working mechanism, poor data interaction, resulting in low overall prediction efficiency and poor accuracy, which cannot meet the actual needs of equipment fault prediction.

[0087] Based on this, please refer to Figure 2A digital twin-based equipment fault prediction system, applied to any of the aforementioned digital twin-based equipment fault prediction methods, includes an implicit coupling interference perception module, a dynamic feature correction module, a fault evolution prediction module, a digital twin mapping interaction module, and a reinforcement learning optimization module. The implicit coupling interference perception module is used to collect multi-dimensional physical data and quantify the implicit coupling interference coefficients. The dynamic feature correction module is used to correct the original features based on the interference coefficients. The fault evolution prediction module is used to calculate the probability of fault occurrence and the evolution trend. The digital twin mapping interaction module is used to achieve bidirectional synchronization between the physical equipment and the digital twin model. The reinforcement learning optimization module is used to dynamically adjust the key modeling parameters. Each module achieves data interaction and collaborative work through a data bus.

[0088] In this technical solution, the system's five major modules adopt a modular design, each possessing independent core functions, while achieving close collaborative operation through a high-speed data bus. The implicit coupling interference sensing module includes a distributed sensor array and an interference quantization unit. The sensor array collects multi-dimensional physical data and transmits it to the interference quantization unit, which calculates the implicit coupling interference coefficient using a preset mathematical formula, and then transmits the data and coefficients to other modules. The dynamic feature correction module receives the interference coefficient and original sensor characteristics transmitted from the implicit coupling interference sensing module, combines them with the load volatility and sensor attenuation coefficient synchronized by the digital twin mapping interaction module, executes a dynamic feature correction algorithm, and outputs corrected high-fidelity characteristics. The fault evolution prediction module receives the high-fidelity characteristics from the dynamic feature correction module, as well as the equipment aging index and historical fault characteristic data synchronized by the digital twin mapping interaction module, calculates the probability of fault occurrence, generates short-term and long-term evolution curves, and triggers graded early warning signals. The digital twin mapping interaction module constructs a digital twin model of the equipment, and synchronizes physical equipment status data and model analysis results in real time through a bidirectional data channel, providing data support for other modules. The reinforcement learning optimization module monitors the prediction error of the fault evolution prediction module in real time. Based on the error, it dynamically adjusts key parameters in the interference quantization, feature correction, and probability calculation processes. The adjusted parameters are then sent to the corresponding modules via a data bus, achieving closed-loop optimization of the entire system. Data interaction between modules uses standardized data formats and communication protocols to ensure fast, accurate, and reliable data transmission. The system also has self-diagnostic capabilities, monitoring the operating status of each module in real time. When a module failure is detected, it automatically triggers an alarm and switches to a backup module to ensure continuous system operation.

[0089] The technical effects achieved by the above-mentioned technical solutions include: modular design makes the system structure clear, the functions of each module are well-defined, and it is easy to maintain and upgrade; each module achieves efficient collaborative work through a data bus, solving the problem of poor data interaction in traditional systems; the reinforcement learning optimization module realizes dynamic closed-loop optimization of system parameters, improving the overall prediction accuracy and adaptability of the system; the self-diagnostic function and backup module design enhance the reliability and stability of the system, ensuring the continuous operation of fault prediction.

[0090] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting equipment failure based on digital twins, comprising collecting multi-dimensional physical data and full lifecycle data during equipment operation, constructing a digital twin model of the equipment and synchronizing the physical state of the equipment in real time, performing fault-related feature processing and predictive analysis based on the digital twin model, and outputting the probability of failure occurrence and evolution trend; characterized in that, The feature processing and predictive analysis process adopts a three-level linkage nonlinear modeling approach. First, the implicit coupling interference of multiple physical fields during equipment operation is quantified. Then, the original features are dynamically corrected based on the interference quantification results. Finally, the corrected features are fused with the equipment's life-cycle correlation data to calculate the probability of fault evolution. The entire process uses artificial intelligence reinforcement learning algorithms to dynamically optimize modeling parameters and achieve accurate fault prediction.

2. The equipment fault prediction method based on digital twins according to claim 1, characterized in that, The multi-dimensional physical data includes the electromagnetic radiation intensity, temperature gradient, continuous operating time, real-time load fluctuation rate, and sensor attenuation coefficient of key components of the equipment. The full life cycle data includes the cumulative operating time of the equipment, the number of maintenance operations, the material fatigue coefficient, and historical fault characteristic data. All data are collected through distributed sensors and data acquisition interfaces, and after being digitally converted, they are transmitted to the processing unit.

3. The equipment fault prediction method based on digital twins according to claim 1, characterized in that, The digital twin model achieves real-time synchronization with the physical state of the equipment through a two-way interactive channel. The synchronized content includes equipment operating parameters, physical quantity change data, aging status data, and feature correlation data. The synchronization frequency is consistent with the data acquisition frequency to ensure the consistency between the model and the physical equipment.

4. The equipment fault prediction method based on digital twins according to claim 1, characterized in that, The artificial intelligence reinforcement learning algorithm uses the fault prediction error as the reward function. The reward function value is negatively correlated with the prediction error. The algorithm dynamically adjusts key parameters such as coupling coefficient, correction weight coefficient, and probability amplification coefficient in the modeling process through iterative training, so that the prediction results continuously approach the actual fault state.

5. The equipment fault prediction method based on digital twins according to claim 1, characterized in that, The quantization of the multiphysics implicit coupling interference is achieved through the following mathematical formula: ； in The implicit coupling interference coefficient is dimensionless. The mean electromagnetic radiation intensity of the key components of the equipment is expressed in V / m. The temperature gradient change rate between the core component and the environment is expressed in °C / s. The continuous operating time of the equipment is measured in hours (h). is the electromagnetic-temperature coupling coefficient, with dimensions (m / V)². This is the duration decay coefficient, with dimensions 1 / h; and This is achieved through reinforcement learning training.

6. The equipment fault prediction method based on digital twins according to claim 5, characterized in that, The dynamic feature correction is achieved through the following mathematical formula: ； in For high-fidelity features after correction, dimensionless; These are the raw, dimensionless features acquired by the sensor. This represents the real-time load volatility, which is dimensionless. The sensor attenuation coefficient is dimensionless. The weighting coefficients are dimensionless to correct for weighting. and It is updated in real time through digital twin models.

7. The equipment fault prediction method based on digital twins according to claim 6, characterized in that, The fault evolution probability is calculated using the following mathematical formula: ； in The probability of failure occurrence is dimensionless and ranges from 0 to 1. The equipment aging index is dimensionless and ranges from 0 to 1. The trace is the cosine similarity matrix between the current feature and the historical fault features. It is dimensionless and ranges from 0 to n, where n is the feature dimension. This is the probability amplification factor, dimensionless, with a value range from 0.1 to 1; It is calculated by combining cumulative running time, number of maintenance operations, and material fatigue coefficient. Quantify multi-dimensional correlations through matrix operations.

8. The equipment fault prediction method based on digital twins according to claim 1, characterized in that, The fault evolution probability calculation also includes generating short-term and long-term fault evolution curves. The short-term fault evolution curve corresponds to the fault development trend from 1 to 24 hours, and the long-term fault evolution curve corresponds to the fault development trend from 1 to 30 days. The curve generation process combines the changing patterns of equipment operating conditions and historical fault evolution data.

9. The equipment fault prediction method based on digital twins according to claim 1, characterized in that, The output fault occurrence probability and evolution trend include triggering graded early warning signals. The early warning signals are divided into level 1, level 2 and level 3 early warnings, which correspond to different fault occurrence probability ranges and emergency handling priorities, respectively. The early warning signals are displayed through a visual interface and synchronized to the equipment management terminal.

10. A digital twin-based equipment fault prediction system, applied to the digital twin-based equipment fault prediction method as described in any one of claims 1-9, characterized in that, include: The system comprises an implicit coupling interference sensing module, a dynamic feature correction module, a fault evolution prediction module, a digital twin mapping interaction module, and a reinforcement learning optimization module. The implicit coupling interference sensing module collects multi-dimensional physical data and quantifies the implicit coupling interference coefficients. The dynamic feature correction module corrects the original features based on the interference coefficients. The fault evolution prediction module calculates the probability of fault occurrence and its evolution trend. The digital twin mapping interaction module enables bidirectional synchronization between the physical device and the digital twin model. The reinforcement learning optimization module dynamically adjusts key modeling parameters. All modules interact and collaborate through a data bus.

Images