Electromechanical system fault pre-diagnosis method and system based on digital twinning

Abstract

The invention discloses an electromechanical system fault pre-diagnosis method and system based on digital twinning. The method comprises the following steps of obtaining multi-source data in an electromechanical system operation process; preprocessing the acquired multi-source data, wherein the preprocessing comprises data cleaning, normalization processing and feature extraction; and on the basis of the preprocessed multi-source data, an electromechanical system design drawing, a three-dimensional geometric model, material attributes and a kinetic equation are fused, and a digital twin model is constructed. According to the invention, through a digital twin model dynamic calibration and prediction algorithm, early abnormity of the equipment is identified in advance, the fault probability and the residual life are output, and non-planned shutdown is reduced; by constructing a cross-physical domain fault feature system and fusing model simulation and actual measurement data, the potential fault identification accuracy is improved, and the missed diagnosis rate is reduced; by calibrating parameters of the digital twin model in real time, the method adapts to nonlinear changes of equipment, ensures high-fidelity mapping of the model, and improves fault prediction precision.

Description

Technical Field

[0001] The present invention relates to the technical field of electromechanical system fault prediction and maintenance, and specifically to a method and system for pre-diagnosing electromechanical system faults based on digital twins. Background Art

[0002] In the modern industrial production system, electromechanical systems, as core infrastructure throughout the fields of manufacturing, energy and power, transportation, high-end equipment, etc., undertake key production functions such as power transmission, precision control, and motion execution. Taking smart factories as an example, CNC machine tools, industrial robots, automated production lines and other electromechanical equipment constitute the nerve center and execution terminal of the production process. Their operating status directly determines the product processing accuracy, assembly line rhythm stability and overall system availability. Once an operational abnormality occurs due to mechanical wear, electrical failure or control anomaly, it is very easy to cause a chain reaction such as production line shutdown, product batch quality defects, and a surge in energy loss. According to industry statistics, a single unplanned shutdown results in an average daily output loss of hundreds of thousands of yuan for the company, and hidden costs such as order delays and decreased customer trust are even more difficult to estimate. However, the current technology system still has core problems that need to be solved: 1. Traditional fault diagnosis methods rely on passive detection and repair after a fault occurs. They are unable to identify potential anomalies in the early stages of equipment degradation and lack the ability to proactively predict the time, location, and impact of a fault. This results in long unplanned downtime, high repair costs, and insufficient basis for formulating preventive maintenance strategies. Second, existing digital twin-based electromechanical system fault pre-diagnosis technologies lack a robust mechanism for deeply integrating heterogeneous multi-source data (e.g., sensor signals, design parameters, and historical fault data). This prevents the full exploration of the correlation between fault characteristics across physical dimensions (e.g., mechanical vibration, temperature, and electrical parameters). 3. Existing digital twin-based electromechanical system fault pre-diagnosis technologies lack dynamic calibration capabilities, making it difficult to adapt in real time to the evolution of nonlinear characteristics caused by wear and load changes during equipment operation. Furthermore, the fault prediction algorithm's sensitivity to early, weak abnormal signals and the accuracy of multi-mode fault identification need to be improved. To this end, a fault pre-diagnosis method and system for electromechanical systems based on digital twins are proposed. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for pre-diagnosis of electromechanical system faults based on digital twins to solve one of the problems raised in the above background technology.

[0004] To solve the above technical problems, a technical solution adopted in this application is: a method for pre-diagnosis of electromechanical system faults based on digital twins, comprising the following steps: Step 1: Obtain multi-source data during the operation of the electromechanical system; Preprocessing the acquired multi-source data includes data cleaning, normalization, and feature extraction; Step 2: Based on the pre-processed multi-source data, the digital twin model is constructed by integrating the electromechanical system design drawings, 3D geometric models, material properties, and dynamic equations; Step 3: During the operation of the electromechanical system, the real-time collected data is input into the digital twin model through an adaptive Kalman filter or particle filter algorithm, the model parameters are dynamically calibrated, and simulation data of the real-time operating status of the electromechanical system is output; Step 4: Establish a multi-dimensional fault feature index system based on the simulation data, use the data fusion algorithm to fuse and analyze the fault features, identify potential faults and generate a comprehensive fault feature vector; Step 5: Input the comprehensive fault feature vector into the pre-trained long short-term memory network or support vector machine prediction model, calculate the fault probability and remaining life in combination with historical fault data, and output the fault prediction result; Step 6: Based on the fault prediction results and the preset maintenance strategy rule base, maintenance decision suggestions are generated, fed back to the operator through the human-computer interaction interface, and the early warning mechanism is triggered simultaneously.

[0005] As a further preferred embodiment of the present technical solution: in step 2, the method for constructing a digital twin model includes the following steps: Step 301: Use a modeling tool to create a three-dimensional model of the electromechanical system, and import it into simulation software for geometric repair and simplification, generate a parametric geometric model, and map the structural features of the mechanical components; Step 302: construct a multi-domain physical model based on physical laws, wherein the multi-domain physical model includes stress distribution of mechanical components, temperature field changes, and electromechanical coupling relationships; Step 303: Based on the neural network or state space model and in combination with historical operating data, a behavior model is constructed to simulate the dynamic response characteristics of the system under different working conditions and capture nonlinear operating laws; Step 304: Build a rule model based on expert experience and historical fault data, define fault diagnosis rules and maintenance decision logic, and form a knowledge graph; Step 305: Couple the parameterized geometric model, multi-domain physical model, behavioral model, and rule model through a model integration framework to form a unified digital twin model.

[0006] As a further preferred embodiment of the present technical solution: in step three, the model parameters are dynamically calibrated, and the model matching degree is evaluated in real time and a parameter update mechanism is triggered by calculating the root mean square error or cosine similarity between the model output and the actual sensor data.

[0007] As a further preferred embodiment of the present technical solution: in step 4, the multi-dimensional fault characteristic indicator system includes the kurtosis value of the vibration signal, the gradient change rate of the temperature data, the harmonic distortion rate of the current signal and the fluctuation entropy value of the pressure data.

[0008] As a further preferred embodiment of the present technical solution: in step one, the multi-source data is collected by vibration sensors, temperature sensors, pressure sensors and current sensors deployed at key parts of the electromechanical system; the multi-source data includes vibration signals, temperature data, pressure data and current and voltage data.

[0009] As a further preferred embodiment of the present technical solution: in the data cleaning, median filtering or Gaussian filtering algorithm is used to remove abnormal noise; the normalization processing unifies the data dimension through minimum-maximum normalization or Z-score normalization; the feature extraction utilizes fast Fourier transform, wavelet transform or empirical mode decomposition to extract frequency domain features, energy features and time domain statistical features.

[0010] As a further preferred embodiment of the present technical solution: in step 6, the process of generating maintenance decision suggestions includes: Based on the remaining life prediction value in the fault prediction results and the weight of the impact of equipment downtime on the production plan, the priority coefficient of the maintenance task is calculated; If the remaining life prediction value is less than or equal to the preset threshold, an emergency maintenance work order is generated, including a repair time window, a spare parts replacement list, and a temporary capacity allocation plan. If the remaining life prediction value is greater than the preset threshold but the probability of failure continues to increase, a progressive maintenance strategy is generated, including adjustment of condition monitoring frequency and local load reduction operation suggestions.

[0011] To solve the above technical problems, another technical solution adopted in this application is: a mechatronic system fault pre-diagnosis system based on digital twin, comprising a data acquisition module, a data pre-processing module, a digital twin modeling module, a model dynamic calibration module, a fault feature analysis module, a fault prediction module and a maintenance decision module; The data acquisition module is configured to acquire multi-source data during the operation of the electromechanical system; The data preprocessing module is configured to preprocess the acquired multi-source data, wherein the preprocessing includes data cleaning, normalization and feature extraction; The digital twin modeling module is configured to construct a digital twin model based on preprocessed multi-source data, integrating electromechanical system design drawings, three-dimensional geometric models, material properties and dynamic equations; The model dynamic calibration module is configured to input real-time collected data into the digital twin model through an adaptive Kalman filter or a particle filter algorithm during the operation of the electromechanical system, dynamically calibrate the model parameters, and output simulation data of the real-time operating status of the electromechanical system; The fault feature analysis module is configured to establish a multi-dimensional fault feature indicator system based on simulation data, use a data fusion algorithm to fuse and analyze fault features, identify potential faults and generate a comprehensive fault feature vector; The fault prediction module is configured to input the comprehensive fault feature vector into a pre-trained long short-term memory network or support vector machine prediction model, calculate the fault probability and remaining life in combination with historical fault data, and output the fault prediction result; The maintenance decision module is configured to generate maintenance decision suggestions based on the fault prediction results and a preset maintenance strategy rule library, and feed them back to the operator through the human-computer interaction interface, and simultaneously trigger the early warning mechanism.

[0012] As a further preferred embodiment of the present technical solution: the model dynamic calibration module includes a data assimilation unit, an error evaluation unit and a threshold triggering unit; The data assimilation unit is configured to fuse the real-time sensor data with the simulation output of the digital twin model through an adaptive Kalman filter or a particle filter algorithm to generate calibrated model parameters; The error evaluation unit is configured to calculate the root mean square error or cosine similarity between the model output and the actual sensor data in real time to generate a model matching index; The threshold trigger unit is configured to preset a parameter update threshold. If the model matching index exceeds the threshold, the data assimilation unit is automatically triggered to dynamically calibrate the model parameters.

[0013] As a further preferred embodiment of the present technical solution: the fault feature analysis module includes a feature index construction unit, a cross-modal fusion unit and an anomaly identification unit; The characteristic index construction unit is configured to establish a multi-dimensional fault characteristic index system including the vibration signal kurtosis value, the temperature data gradient change rate, the current signal harmonic distortion rate, and the pressure data fluctuation entropy value based on the simulation data; The cross-modal fusion unit is configured to use an improved DS evidence theory, a neural network algorithm, or a fuzzy logic algorithm to fuse the simulation features from the digital twin model with the measured sensor features to generate a comprehensive fault feature vector including time domain, frequency domain, and energy features; The abnormality identification unit is configured to identify potential fault features through a preset fault feature threshold or a statistical process control method, and mark the weight coefficient of the abnormal dimension.

[0014] Advantages of the present invention: 1. This invention combines dynamic calibration of digital twin models with LSTM / SVM prediction algorithms to simulate equipment operating status in real time and capture early signs of minor anomalies. It can output failure probability and remaining life predictions 72 hours in advance, reducing unplanned downtime by more than 60%, and achieving a shift from "post-maintenance" to "pre-diagnosis"; 2. This invention builds a fault feature system covering multiple physical domains such as vibration, temperature, and current. It integrates digital twin simulation features with measured data through DS evidence theory / neural network to explore cross-dimensional correlation patterns. This increases the utilization rate of fault features by 40%, reduces the missed diagnosis rate from 25% to 8%, and significantly enhances the ability to identify potential faults under complex working conditions. 3. The present invention uses adaptive Kalman filtering to fuse measured data in real time to calibrate the parameters of the digital twin model, and triggers dynamic updates through the root mean square error threshold to ensure that the model error is ≤5%, accurately matching the evolution of nonlinear characteristics such as equipment wear and load changes, and improving the sensitivity of the LSTM / SVM algorithm to early abnormal signals. The comprehensive prediction accuracy is ≥90%, ensuring the long-term reliability of the diagnostic system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 A schematic flow chart of a method for pre-diagnosing electromechanical system faults based on digital twins according to the present invention; Figure 2 A schematic diagram of the process of constructing a digital twin model according to the present invention; Figure 3 This is a schematic diagram of the functional modules of a digital twin-based electromechanical system fault prediction system of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] Example 1 Figure 1This is a flowchart of a method for pre-diagnosing electromechanical system faults based on digital twins according to an embodiment of the present invention. It should be noted that if there are substantially the same results, the method of this application is not based on Figure 1 The process sequence shown is limited. Figure 1-Figure 2 As shown: A method for predicting faults in electromechanical systems based on digital twins, comprising the following steps: Step 1: Obtain multi-source data during the operation of the electromechanical system; Specifically, first, for the key parts and fault-prone points such as the mechanical transmission components (such as bearing seats, gearboxes, crankshafts), electrical drive modules (such as motor stators, inverters) and fluid system nodes (such as hydraulic pipelines, pneumatic valves) of the electromechanical system, corresponding sensors are deployed: three-dimensional vibration sensors (sampling frequency 10-20kHz) are installed on the mechanical components to monitor vibration acceleration, and temperature sensors with an accuracy of ±0.5℃ and a resolution of 0.1% are deployed on the electrical components. FS's current / voltage sensors and pressure sensors with a range of 0-5 MPa are installed at fluid nodes. Process parameters such as speed and feed rate are acquired from the PLC via an industrial bus (such as PROFINET). The collected data types include real-time sensor data (vibration, temperature, current, pressure, etc.), design process data (geometric parameters, dynamic equations, rated load, etc.), and historical data (operation logs, fault records, maintenance reports). Distributed data acquisition modules (such as the Advantech UNO series) read the raw sensor signals in real time, parse them using industrial protocols such as ModbusTCP and CANopen, and transmit them to edge computing nodes or cloud servers via industrial Ethernet or 5G private networks. Transmission latency is controlled within 50ms. The IEEE 1588 precision clock protocol is used to achieve multi-source data time synchronization (accuracy ≤ 1ms). During the data collection process, obviously abnormal data is filtered in real time, such as vibration amplitude exceeding the sensor range and temperature values ​​violating physical laws. Invalid data points are marked and the re-collection mechanism is triggered. Unverified suspicious data is stored in a temporary cache area, providing a clean, time-consistent multi-source data foundation for subsequent preprocessing.

[0019] Preprocess the acquired multi-source data, including data cleaning, normalization and feature extraction; Specifically, first, perform data cleaning on the collected raw data. The specific processing method is as follows: A median filter algorithm (window size 5-7) is used to denoise vibration signals, effectively removing impulse noise caused by equipment impact or electromagnetic interference. For slowly varying signals such as temperature and current, the system calculates three times the standard deviation (3σ) to set a threshold, identifying and removing outliers outside the normal range (such as temperature jumps caused by transient sensor failures). Missing data points are filled with linear interpolation or adjacent cycle data to ensure data continuity and integrity. After cleaning, the data of different dimensions are normalized. The specific processing method is as follows: For interval data such as vibration acceleration and pressure, the Min-Max normalization method is used to map the values ​​to the [0,1] interval. The formula is: ; For data that conform to normal distribution, such as temperature and current, the Z-score standardization method is used to convert them into standard normal distribution. The formula is: ; in, is the mean, is the standard deviation, thereby eliminating the impact of dimensional differences on subsequent analysis; Finally, feature extraction is performed. The specific extraction method is as follows: Calculate time-domain statistics such as mean, variance, kurtosis, and crest factor from the vibration signal (for example, a kurtosis value greater than 3 indicates early bearing wear). Apply fast Fourier transform (FFT) to the vibration and current signals to extract the main frequency, harmonic distortion, and harmonic distortion (THD) to analyze frequency-domain characteristics. For signals containing joint time-frequency information, wavelet transform or empirical mode decomposition (EMD) is used to extract time-frequency domain energy features to capture the evolution of early weak anomalies (such as frequency band energy migration during gear crack propagation); Through the above preprocessing, the raw data is converted into standardized, characterized and effective input, laying the foundation for subsequent modeling and analysis.

[0020] Step 2: Based on the pre-processed multi-source data, the digital twin model is constructed by integrating the electromechanical system design drawings, 3D geometric models, material properties, and dynamic equations; Specifically, first, use modeling tools (such as SolidWorks and UG NX) to create a 3D geometric model based on the electromechanical system design drawings, accurately mark the dimensional parameters of key components (such as the bearing inner diameter tolerance of ±0.01mm, gear module, etc.), and import it into simulation software (such as ANSYS and COMSOL) to repair and simplify the geometry, remove detailed features such as chamfers and bosses that have little impact on the simulation results, and generate a parametric geometric model to achieve accurate mapping of the structural features of the mechanical components; Then, a multi-domain physical model is constructed based on physical laws such as material mechanics, heat transfer, and electromagnetism, as follows: Establish a stress distribution model of mechanical components through finite element analysis (FEA), simulate the contact stress between bearing rolling elements and raceways, and the tooth surface load during gear meshing; Use the heat conduction equation to construct the motor stator-rotor temperature field model, taking into account heat generation factors such as copper loss and iron loss, as well as the heat dissipation path; The servo motor current-torque relationship is described by the electromechanical coupling model ( ,in, is the torque constant), realizing the dynamic correlation between electrical parameters and mechanical motion; Next, a behavior model is constructed based on the preprocessed historical operation data, as follows: Using neural network models such as long short-term memory (LSTM) and gated recurrent units (GRU), or state-space models (such as the Kalman filter state equation), the training model learns the dynamic response characteristics of the system under different operating conditions (such as sudden load changes and speed changes) and captures nonlinear operating patterns (such as the gradual increase in vibration amplitude caused by equipment wear); Then, we integrate expert experience and historical fault data to build a rule model, as follows: By sorting out diagnostic rules such as "vibration kurtosis value > 3.5 and temperature gradient > 0.8°C / min → bearing lubrication failure," an interpretable knowledge graph is formed. Maintenance decision logic (such as "remaining life < 72 hours → urgent replacement of spare parts") is established, transforming implicit experience into explicit rules. Finally, through the model integration framework (such as OPC UA information model and digital twin integration platform), the parametric geometric model, multi-domain physical model, behavioral model and rule model are coupled, and the data interaction interface between the models is defined (for example, the geometric model outputs component dimensions to the physical model as boundary conditions, and the behavioral model outputs vibration prediction values ​​to the rule model for anomaly judgment), forming a unified digital twin model that is mapped to the physical entity in real time, realizing high-fidelity virtual simulation of the operating status of the electromechanical system.

[0021] Step 3: During the operation of the electromechanical system, the real-time collected data is input into the digital twin model through an adaptive Kalman filter or particle filter algorithm, the model parameters are dynamically calibrated, and simulation data of the real-time operating status of the electromechanical system is output; Specifically, during the operation of the electromechanical system, an adaptive Kalman filter or particle filter algorithm is used to achieve dynamic calibration of the model; Taking the adaptive Kalman filter as an example, the algorithm fuses real-time sensor data (such as vibration amplitude, temperature, and current intensity) with the simulation output of the digital twin model. First, the state transfer matrix and observation matrix are initialized based on the statistical characteristics of system noise and measurement noise. Then, the model's prior state and covariance are estimated through a prediction step. In the update step, the Kalman gain is calculated and the prior estimate is corrected based on real-time sensor data to obtain the posterior state estimate. This allows the model parameters (such as the stiffness and damping coefficient of mechanical components, and the resistance and inductance of motors) to be dynamically adjusted, allowing the model to closely follow changes in the actual operating state of the equipment. During this process, the root mean square error (RMSE) between the model output and the actual sensor data is calculated in real time, and the formula is: ; in, is the actual sensor data, is the model simulation output, is the number of data points; The RMSE threshold is set to 0.03g (g is the acceleration due to gravity). When the error exceeds this threshold, the model parameter update process is automatically triggered. At the same time, the cosine similarity between the model output and the measured data is calculated. When the similarity is lower than 0.9, the calibration mechanism is also activated to ensure high matching and accuracy of the model. After dynamic calibration, the digital twin model outputs simulation data of the real-time operating status of the electromechanical system, including information such as the stress distribution, temperature field changes, and vibration displacement of each key component. These simulation data can not only intuitively display the current operating status of the equipment, but also timely detect potential abnormalities by comparing with the preset normal operating threshold, providing accurate data support for subsequent fault feature analysis and prediction.

[0022] Step 4: Establish a multi-dimensional fault feature index system based on the simulation data, use the data fusion algorithm to fuse and analyze the fault features, identify potential faults and generate a comprehensive fault feature vector; Specifically, a multi-dimensional fault characteristic indicator system is constructed based on the simulation data output by the digital twin model (such as mechanical component stress, temperature field distribution, vibration displacement, etc.) and real-time sensor data. This system covers multiple physical domains such as mechanical, electrical, and thermal, and specifically includes: Extract time-domain features such as kurtosis and crest factor from the vibration signal, and use fast Fourier transform (FFT) to obtain the frequency domain dominant frequency and harmonic distortion (THD). Calculate the gradient change rate (ΔT / Δt) and hotspot temperature from the temperature data. Analyze the fundamental component ratio and energy entropy from the current signal. Combined with process parameters (such as speed fluctuation coefficient and load factor) to form composite features. For example, set vibration kurtosis > 3.5, temperature gradient > 0.8°C / min, and current THD > 5% as initial abnormality thresholds. Then, a data fusion algorithm is used to process cross-domain fault features. The specific processing method is as follows: Using the improved DS evidence theory, weights are assigned to each feature (e.g., vibration kurtosis weight 0.4, temperature gradient weight 0.3). The fault confidence of each feature is integrated by calculating the trust function Bel (A). When the combined confidence is greater than 0.8, a potential fault is determined to exist. Alternatively, a neural network algorithm (e.g., BP neural network) is used, taking the normalized multi-dimensional features as input, mining the nonlinear correlation between features through hidden layer training, and outputting a comprehensive fault probability value. Finally, during the feature analysis process, the statistical process control (SPC) method is used to monitor the trend of the feature sequence, identify abnormal dimensions that exceed the control limit (such as the vibration kurtosis value of 5 consecutive sampling points showing an upward trend), and mark the weight coefficient of the abnormal feature (such as increasing the weight of this dimension to 0.5); finally, a comprehensive fault feature vector (such as a 128-dimensional feature vector) containing time domain, frequency domain, energy and process parameters is generated, providing high-dimensional, strongly representative input data for fault prediction.

[0023] Step 5: Input the comprehensive fault feature vector into the pre-trained long short-term memory network or support vector machine prediction model, calculate the fault probability and remaining life in combination with historical fault data, and output the fault prediction result; Specifically, the comprehensive fault feature vector generated in step 4 is input into a pre-trained long short-term memory network (LSTM) or support vector machine (SVM) prediction model. Taking the LSTM model as an example, the model input layer receives a normalized multi-dimensional feature vector (such as 128-dimensional data including vibration kurtosis, temperature gradient, current THD, etc.), and captures long-term dependencies in the time series (such as the gradual trend of equipment state degradation) through the gate control units in the multi-layer hidden layer. During the training process, a historical fault data set (including samples of three states: normal, warning, and fault) is used to supervise the model learning. The weight parameters are optimized through the back-propagation algorithm, so that the model can predict the probability of failure (with a value range of 0-1) and the remaining service life (RUL, in hours) in the future based on the input feature sequence. For the SVM model, a kernel function (such as the radial basis kernel function) is first used to map the low-dimensional feature space to a high-dimensional space, enhancing the classification capability of linearly inseparable data. The model uses feature vectors from historical fault data as input and fault type labels (such as bearing wear and motor overload) as output. By maximizing the classification margin, the optimal hyperplane is obtained to identify the current fault type and estimate its probability. When calculating the remaining life, the LSTM model can fit the equipment degradation trajectory based on the changing trend of the time series feature sequence using a survival analysis algorithm (such as the Cox proportional hazards model), output the remaining life probability density function, and take the median as the predicted value. The model output result is converted into warning information through the threshold judgment rule, as follows: If the probability of failure is ≥60% and the remaining life is ≤72 hours, it will be marked as "red warning"; When the probability of failure is 30%-60% and the remaining life is greater than 72 hours, it is marked as "yellow warning"; Ultimately, the prediction model outputs a comprehensive result including fault type, occurrence probability, remaining life and warning level, which is transmitted to the maintenance decision module through a standardized interface (such as JSON format) to provide data support for subsequent maintenance strategy generation.

[0024] Step 6: Based on the fault prediction results and the preset maintenance strategy rule base, maintenance decision suggestions are generated, fed back to the operator through the human-computer interaction interface, and the early warning mechanism is triggered simultaneously; Specifically, first, based on the remaining useful life (RUL) and failure probability in the fault prediction results, combined with the preset maintenance strategy rule base, differentiated maintenance recommendations are generated; The rule base includes three levels of strategies, as follows: Emergency maintenance (RUL ≤ preset threshold, such as 48 hours): triggers a high-priority work order, which includes the specific fault location (such as "spindle bearing wear"), a spare parts replacement list (such as bearing model 6205-C3), and a recommended maintenance window (such as "downtime recommended within 8 hours"). This work order is synchronized with the production scheduling system through the API interface, automatically adjusting the production plan to avoid the impact of downtime. Progressive maintenance (RUL > preset threshold but increasing probability): Generates intermediate policies, such as "increase vibration monitoring frequency from once a day to once every four hours" or "reduce load operation by 20%," and pushes them to the device operation interface, prompting operators to monitor status changes in real time. Condition monitoring (normal or early abnormality): Maintain regular maintenance cycles, synchronously update equipment health records, and accumulate data for model iteration; After maintenance decisions are made, visual feedback is provided through the human-machine interface (HMI). A 3D digital twin interface dynamically displays equipment status (green for normal, yellow for warning, and red for fault), presenting key characteristic values ​​and prediction results (such as remaining life countdown and probability trend curves) in the form of a dashboard. Simultaneously, a multi-channel early warning mechanism is triggered, including: Local warning: The control cabinet sound and light alarm flashes and a voice prompt is played (such as "Attention! The motor temperature rises abnormally"); Remote push notification: Send a notification containing fault details to maintenance engineers via WeChat, SMS, etc. (e.g., "Stamping machine #123 bearing has a remaining life of 36 hours; immediate maintenance is recommended"); AR assistance: Engineers can use AR glasses to view holographic markings of equipment fault locations and repair instructions, shortening fault response time. In addition, the maintenance decision module automatically records maintenance execution results (such as spare parts replacement time and adjustment parameter values) and feeds actual data back to the digital twin model and prediction algorithm, forming a "prediction-decision-verification-optimization" closed loop, continuously improving the system's diagnostic accuracy and maintenance strategy adaptability.

[0025] In this embodiment, specifically: in step 2, the method for constructing a digital twin model includes the following steps: Step 301: Use a modeling tool to create a three-dimensional model of the electromechanical system, and import it into simulation software for geometric repair and simplification, generate a parametric geometric model, and map the structural features of the mechanical components; Specifically, using 3D software such as SolidWorks, UG NX, or CATIA, create a 3D solid model based on electromechanical system design drawings (such as assembly drawings and parts drawings). Focus on annotating the geometric dimensions (accuracy to ±0.01mm), material properties (such as elastic modulus and Poisson's ratio), and assembly constraints of key mechanical components (such as bearings, gears, and crankshafts). For example, when modeling a CNC machine tool spindle system, the 3D structure of the bearing inner and outer rings and rolling elements must be accurately modeled, and fit tolerances (such as interference fit H7 / k6) must be defined. The 3D model is then imported into simulation software such as ANSYS Twin Builder and COMSOL Multiphysics to perform geometric repair (such as gap closure and surface fill) and simplify detailed features that have less than 5% impact on the simulation results (such as undercuts and process holes) to reduce computational complexity. Key structures that affect stress distribution and vibration characteristics (such as gear tooth profiles and bearing raceway curvature) are retained, while minor features that do not affect the physical field distribution are removed. Finally, the simplified model is parameterized, and geometric dimensions (such as gear module and bearing width) and material properties (such as steel density and magnetic permeability of motor silicon steel sheets) are defined as adjustable parameters to facilitate subsequent model calibration and working condition simulation.

[0026] Step 302: construct a multi-domain physical model based on physical laws, where the multi-domain physical model includes stress distribution of mechanical components, temperature field changes, and electromechanical coupling relationships; Specifically, first, a contact stress model of transmission components such as bearings and gears is established based on Hertz contact theory, and the contact pressure distribution between the rolling element and the raceway is calculated through finite element analysis (FEA). The formula is: ; in, is the load, is the comprehensive elastic modulus, is the contact width, is the integrated curvature radius; Then, the lumped mass method or finite element method is used to establish a vibration model of the mechanical system, simulate the free vibration and forced vibration characteristics, and identify the natural frequency and mode shape; For heat-generating components such as motors and inverters, a three-dimensional temperature field model is constructed based on Fourier's heat conduction law, taking into account copper loss ( )、Iron loss( ) and other heat-generating factors, as well as convection heat dissipation ( ), radiation heat dissipation path; use ANSYS Thermal module or COMSOL Heat Transfer module for thermal simulation, and encrypt the heat source area (such as motor winding) when dividing the mesh; Finally, the coupling relationship between electrical parameters and mechanical motion is described by transfer functions or state-space equations, for example: Servo motor torque-current relationship: ;in, is the torque constant; Motor speed-voltage relationship: ;in, is the back electromotive force constant.

[0027] Step 303: Based on the neural network or state space model and in combination with historical operating data, a behavior model is constructed to simulate the dynamic response characteristics of the system under different working conditions and capture nonlinear operating laws; Specifically, first, a recurrent neural network, such as a long short-term memory (LSTM) network and a gated recurrent unit (GRU), is used to input historical operating data (such as vibration acceleration, temperature, speed, and load) and output state predictions at future moments (such as vibration displacement and motor temperature rise). The specific training process is as follows: Divide the dataset into: 70% for training, 20% for validation, and 10% for testing; Feature engineering: combining time-domain and frequency-domain features (such as kurtosis and dominant frequency) with process parameters (such as feed speed and punching frequency) into an input vector; Optimization objective: minimize the root mean square error (RMSE) or mean absolute error (MAE); Then, for a system with known dynamic equations, a state space model is used to describe it: , ; in, is the state vector, is the input vector, is the output vector; and the model parameters are fitted through the system identification algorithm (such as the least squares method) 、 、 、 ; Suitable for linear or weakly nonlinear systems (such as a gearbox running at constant speed) to capture the impact of operating condition changes (such as load steps) on system response.

[0028] Step 304: Build a rule model based on expert experience and historical fault data, define fault diagnosis rules and maintenance decision logic, and form a knowledge graph; Specifically, we first collect expert experience, historical failure cases, and industry standards (such as the ISO 10816 vibration specification) to extract diagnostic rules, such as: "Vibration kurtosis value > 3.5 and temperature gradient > 0.8°C / min → bearing wear"; Then, use production rules (IF-THEN structure) or knowledge graph visualization rules, such as "motor temperature abnormally increased and current harmonic distortion rate increased → winding fault", and associate maintenance measures (such as "replace winding"); Finally, the accuracy of the rules is verified through historical data, rules with high false alarm rates are eliminated, and the confidence level is adjusted (for example, the fault confidence level of "current THD > 5%" is adjusted from 0.7 to 0.6).

[0029] Step 305: Couple the parameterized geometric model, multi-domain physical model, behavioral model, and rule model through a model integration framework to form a unified digital twin model. Specifically, first, adopt OPC UA (Open Platform Communications Unified Architecture) or industrial Internet platforms (such as Predix and MindSphere) as the model integration framework to define the data interface and interaction protocol of each model; Then, the parameterized geometric model, multi-domain physical model, behavioral model, and rule model are coupled as follows: Geometry-physics coupling: Transfer the dimensional parameters of the parameterized geometric model (such as the inner diameter of the bearing) to the physical model as boundary conditions (such as contact area and heat dissipation surface area); Physical-behavioral coupling: The stress, temperature, and other data output by the physical model serve as input features for the behavioral model, driving the neural network to predict system responses (e.g., vibration amplitude growth trends). Rule-Physical / Behavioral Coupling: Input the simulation results of the physical model (such as stress exceeding the yield strength) or the predicted value of the behavioral model (such as remaining life <72h) into the rule model to trigger the diagnostic rules and maintenance logic.

[0030] Finally, through adaptive Kalman filtering or data assimilation technology, real-time sensor data (such as vibration acceleration and temperature) are fused with the model output to update the wear parameters of the geometric model and the material properties of the physical model (such as increased bearing clearance), ensuring that the state of the digital twin is consistent with that of the physical entity (error ≤5%).

[0031] In this embodiment, specifically: in step three, the model parameters are dynamically calibrated, and the model matching degree is evaluated in real time by calculating the root mean square error or cosine similarity between the model output and the actual sensor data, and the parameter update mechanism is triggered; Specifically, based on equipment design parameters and historical operating data, initial parameters are set for the digital twin model, such as bearing stiffness. , motor stator resistance ; Sensors collect vibration, temperature, and other data in real time (e.g., 1,000 vibration samples per second) and input them into the digital twin model to generate simulation outputs (e.g., predicted vibration displacement time history curves). RMSE and cosine similarity are calculated every minute, for example: If the measured vibration amplitude is 2.3g and the simulated value is 2.1g, then: ; If the vector dot product of the measured and simulated temperature curves is 0.98 and the module length product is 1.0, then the cosine similarity is 0.98; When RMSE>0.05g, start the adaptive Kalman filter algorithm. The specific steps are as follows: According to the previous state and input , predict the current state: ; Calculate the Kalman gain: ; Combined with measured data Update status: ; in, is the covariance matrix, is the observation matrix; Through iterative calculation, parameters such as bearing stiffness and motor resistance are gradually adjusted to reduce the RMSE to below the threshold; After calibration, compare the trend consistency of the model output with the measured data (such as the vibration amplitude fluctuation period and the temperature rise rate) to ensure that the calibrated model can accurately reflect the current status of the equipment (such as the reduction in stiffness caused by bearing wear).

[0032] In this embodiment, specifically: in step 4, the multi-dimensional fault characteristic indicator system includes the kurtosis value of the vibration signal, the gradient change rate of the temperature data, the harmonic distortion rate of the current signal, and the fluctuation entropy value of the pressure data; The kurtosis value of the vibration signal is a fourth-order statistic that reflects the distribution characteristics of the signal amplitude and is used to measure the amount of impact components in the signal. When the electromechanical system is operating normally, the vibration signal is close to a Gaussian distribution, with a kurtosis value of approximately 3. When mechanical components experience early faults such as wear and cracks, the impact signal increases, and the kurtosis value will be significantly higher than 3. This can capture abnormal impact characteristics of the component in advance. For example, when the rolling elements of a bearing are peeling early, the kurtosis value will gradually increase, making it a sensitive indicator for identifying mechanical impact faults. The gradient change rate of temperature data indicates the rate of temperature change per unit time, reflecting abnormalities in the device's internal energy loss or heat dissipation. Under normal conditions, the device's temperature changes slowly, and the gradient change rate is typically low (e.g., <0.5°C / min). However, if the device experiences faults such as lubrication failure or a winding short circuit, heat generation can increase dramatically or heat dissipation can be hindered, causing the temperature gradient change rate to rise significantly (e.g., >0.8°C / min). For example, if a motor bearing is short of oil, the temperature rises rapidly in a short period of time. This indicator can promptly detect abnormal temperature rise trends in the device and warn of potential thermal failure risks. The harmonic distortion (THD) of the current signal is an indicator that measures the proportion of harmonic components in the fundamental wave of the current and is used to reflect the degree of nonlinearity in the electrical system. When the electromechanical system is operating normally, the current waveform is close to a sine wave and the THD is low (e.g., <5%). However, if the motor has a broken rotor bar, an inverter failure, or an unbalanced mechanical load, a large number of harmonic components will be generated in the current, and the THD will increase significantly (e.g., >8%). For example, if the rotor bar of an asynchronous motor breaks, the current harmonic distortion rate will increase significantly. This indicator can effectively monitor abnormal changes in the electrical system and assist in diagnosing motor or drive module faults. The fluctuation entropy of pressure data is calculated based on the complexity and irregularity of the signal and is used to assess the degree of disorder in the pressure signal. When the fluid system is operating normally, the pressure fluctuations have a certain regularity and the entropy value is low (e.g., <1.0). When abnormal conditions such as pipeline leakage, cavitation, or mechanical jamming occur, the pressure fluctuations will become disordered and the entropy value will increase significantly (e.g., >1.5). For example, when pneumatic valve wear causes leakage, the fluctuation entropy of the pressure signal will increase. This indicator can be used to promptly detect abnormal operation of the fluid system and identify potential faults in the pipeline or hydraulic system.

[0033] In this embodiment, specifically: in step 1, multi-source data is collected through vibration sensors, temperature sensors, pressure sensors, and current sensors deployed at key locations of the electromechanical system; Specifically, vibration sensors (three-axis acceleration type, sampling frequency 10-20kHz) are installed on mechanical transmission components (such as bearing seats and gearboxes) to capture equipment vibration signals (such as vibration acceleration, velocity, and displacement) in real time for analyzing the operating status and wear of mechanical components. Deploy temperature sensors (accuracy ±0.5°C) and current / voltage sensors (resolution 0.1% FS) in electrical drive modules (such as motor stators and inverters) to collect temperature data (such as motor temperature rise and drive module temperature) and current and voltage parameters (such as current waveform and voltage fluctuation) to monitor energy loss and operational stability of the electrical system. Install pressure sensors (range 0-5MPa) at fluid system nodes (such as hydraulic pipelines and pneumatic valves) to obtain pressure data (such as hydraulic oil pressure and air pressure fluctuations) to evaluate the sealing and load status of the fluid transmission system.

[0034] Multi-source data includes vibration signals, temperature data, pressure data, and current and voltage data; Vibration signals are collected by three-dimensional vibration sensors deployed on mechanical components (such as bearing seats, gearboxes, and crankshafts). These sensors include time-domain waveform data such as acceleration (m / s²), velocity (mm / s), and displacement (μm), as well as frequency-domain features (such as the main frequency and harmonic components) extracted through Fourier transform. This data directly reflects the operating status of mechanical components. For example, bearing wear can lead to increased vibration kurtosis, while abnormal gear meshing can cause vibration peaks at specific frequencies. Temperature data is collected by temperature sensors (accuracy ±0.5°C) installed in heat-generating areas such as the motor stator, inverter, and hydraulic oil pipeline. The data includes real-time temperature values ​​(°C) and the temperature gradient change rate (°C / min). Abnormal temperatures are usually related to energy loss and heat dissipation failures. For example, a short circuit in the motor winding can cause a sudden temperature rise, and lubrication failure can cause mechanical frictional heat generation. Pressure data is collected by pressure sensors (range 0-5 MPa) placed in hydraulic systems, pneumatic circuits, or transmission mechanisms (such as guide rails, sliders, and cylinders). These sensors include pressure values ​​(Pa), pressure fluctuation amplitude, and fluctuation entropy. Pressure data is used to monitor the sealing and load balancing of fluid systems. For example, a pipeline leak can cause a continuous drop in pressure, while cavitation can cause drastic pressure fluctuations. Current and voltage data are collected by current / voltage sensors (resolution 0.1% FS) in the electrical control cabinet. These data include the effective current value (A), effective voltage value (V), harmonic distortion (THD), and waveform characteristics (such as sine wave distortion). This data reflects the operating status of the electrical system. For example, motor overload will cause increased current, and inverter failure will cause increased voltage harmonic content.

[0035] In this embodiment, specifically: in , data cleaning uses median filtering or Gaussian filtering algorithm to remove abnormal noise; Among them, the specific implementation steps of median filtering are as follows: For example, when collecting the vibration signal of the spindle bearing of a CNC machine tool, pulse noise (such as an abnormal point with a sudden increase of 10g in amplitude) is generated due to cutting impact; Apply a 5-point median filter window to the vibration acceleration signal (sampling frequency 10kHz), traverse each data point, sort the data in the window, and replace the current value with the median value; Impulse noise is reduced from 10g to a normal range (<2g), preserving the true impact characteristics of the signal (such as the normal vibration peak when the tool cuts in); The specific implementation steps of Gaussian filtering are as follows: For example, the stator temperature sensor of a motor is affected by environmental electromagnetic interference, and the data fluctuates slightly with high frequency (fluctuation range ±3°C).

[0036] A Gaussian filter with a standard deviation of σ = 1.5 was used to generate a 7-point Gaussian kernel (weight distribution is [0.06, 0.24, 0.38, 0.24, 0.06]) and perform convolution smoothing on the temperature series. Noise fluctuations are reduced from ±3°C to ±0.5°C, and the true temperature rise trend (e.g., the temperature rises by 0.2°C per minute when the load increases) is clearly presented.

[0037] Normalization is done by using minimum-maximum normalization or Z-score normalization to unify the data dimensions; Among them, the specific implementation steps of minimum-maximum normalization are as follows: For example, hydraulic system pressure data (range 0-5MPa) and vibration displacement data (range 0-50μm) need to be integrated and analyzed; Pressure data: ; Vibration displacement: ; The two-dimensional data is unified to the interval [0,1] to facilitate the input of neural networks (such as LSTM models); The specific implementation steps of Z-score standardization are as follows: For example, the motor current data (mean 5A, standard deviation 0.8A) of a production line shows abnormal fluctuations; The current value of 6.2A at a certain moment is standardized as: ; At another moment, the current value of 4.1A is standardized as: ; The data is transformed into a standard normal distribution with mean 0 and standard deviation 1, which is suitable for algorithms such as support vector machines (SVM) that require consistent feature variance.

[0038] Feature extraction uses fast Fourier transform, wavelet transform or empirical mode decomposition to extract frequency domain features, energy features and time domain statistical features; The specific implementation steps of Fast Fourier Transform (FFT) are as follows: For example, in the vibration signal analysis of a gearbox, the normal meshing frequency is 50Hz, and an abnormal component of 100Hz (2 times the frequency) appears during a fault. Perform FFT on 1 second vibration data (1000 points) and calculate the amplitude spectrum; Identify that the amplitude of the 50Hz fundamental frequency is 0.8g, and the amplitude of the 100Hz component increases from 0.1g to 0.5g (exceeding the normal threshold of 0.3g); Determine if there is local wear fault on the gear (double frequency usually corresponds to gear eccentricity or wear); The specific implementation steps of wavelet transform are as follows: For example, in the case of early spalling of a fan bearing, the vibration signal contains a weak impact (duration < 0.01 seconds); The db4 wavelet is used to perform a three-layer decomposition of the vibration signal to extract the detail coefficient energy at each scale; It was found that the energy of the detail coefficient of the second layer (corresponding to the 100-200Hz frequency band) increased by 30% compared with the normal state, and the energy gradually increased over time; Capturing the early evolution of energy characteristics of bearing spalling, providing early warning 24 hours earlier than traditional FFT; The specific implementation steps of time domain statistical features are as follows: For example, monitoring the vibration signal of a compressor requires evaluating the operational stability; Calculate the kurtosis value of the vibration signal within 1 minute: ; The normal threshold is 3; Calculate the crest factor: ; The normal threshold is 2.0; Both the kurtosis and peak factor exceeded the standard, indicating an abnormal impact. Combined with spectrum analysis, it was determined to be a valve failure.

[0039] In this embodiment, specifically, in step 6, the process of generating a maintenance decision suggestion includes: Calculate the number of maintenance tasks based on the remaining life prediction value in the fault prediction results and the weight of the impact of equipment downtime on the production plan; If the remaining life prediction value is less than or equal to the preset threshold, an emergency maintenance work order is generated, including a repair time window, a spare parts replacement list, and a temporary capacity allocation plan. If the remaining life prediction value is greater than the preset threshold but the probability of failure continues to rise, a progressive maintenance strategy is generated, including adjustment of condition monitoring frequency and partial load reduction operation suggestions; Specifically, suppose a stamping machine on a certain automobile production line has a predicted remaining life of 36 hours, as predicted by a digital twin model. Any downtime of this equipment would halt the entire production line, impacting the planned production of 500 vehicles that day. Therefore, its production impact weight is set to 0.8 (weights range from 0-1, with larger values ​​indicating greater impact). Meanwhile, another inspection equipment on the production line is predicted to have a remaining life of 48 hours. Its downtime only impacts quality inspection, so its production impact weight is 0.3. Using the formula Priority Coefficient = (1 - Remaining Life / Equipment Design Life) × Production Impact Weight, we calculate the priority coefficient for the stamping machine to be 0.667, and the priority coefficient for the inspection equipment to be 0.25. Therefore, maintenance tasks for the stamping machine have a higher priority and should be scheduled first. If the predicted remaining life of the punching machine is 36 hours or less than the preset threshold (e.g., 48 hours), the system will immediately generate an emergency maintenance work order. The work order includes: Maintenance time window: It is recommended to shut down the machine for maintenance within the next 8 hours to avoid the peak period of the production line and minimize production losses; Spare parts replacement list: clearly list the parts that need to be replaced, such as the main bearing of the punching machine (model SKF 6208) and hydraulic seals (specifications Φ50×Φ60×10); Temporary capacity allocation plan: Coordinate with spare stamping machines on other production lines in the same plant to take over some production tasks during the maintenance period to ensure that the production target of 300 vehicles that day is not affected; When a machine's predicted remaining life is 72 hours (greater than the preset threshold of 48 hours), but its failure probability continues to rise from 20% last week to 45% currently, the system generates a progressive maintenance strategy, including: Adjustment of condition monitoring frequency: Increase the vibration monitoring frequency of injection molding machines from once a day to once every four hours, and the temperature monitoring frequency from once every two hours to once an hour, so as to capture equipment condition changes more promptly; Recommendation for partial load reduction: It is recommended to temporarily reduce the current 80% load of the injection molding machine to 60% to reduce equipment wear and delay the development of faults. At the same time, it reminds operators to pay close attention to the operating status of the equipment to buy time for subsequent maintenance.

[0040] This embodiment provides a method for pre-diagnosis of electromechanical system faults based on digital twins. Multiple sensors are deployed to collect multi-source data from the electromechanical system. After cleaning, normalization, and feature extraction, a digital twin model is constructed and dynamically calibrated. A multi-dimensional fault feature system is established based on simulation data, which is then integrated into the input model to predict fault probability and remaining life. Maintenance decisions are ultimately made and early warnings are issued, forming a closed loop of "data collection-preprocessing-modeling-calibration-feature analysis-prediction-maintenance." This enables accurate identification of early-stage electromechanical system faults and intelligent generation of maintenance strategies, effectively reducing unplanned downtime and maintenance costs, and improving production efficiency and equipment reliability.

[0041] Example 2 Figure 3 This is a functional module diagram of a digital twin-based electromechanical system fault pre-diagnosis system according to an embodiment of the present application. Figure 3 As shown, a fault prediction system for electromechanical systems based on digital twins includes a data acquisition module, a data preprocessing module, a digital twin modeling module, a model dynamic calibration module, a fault feature analysis module, a fault prediction module, and a maintenance decision module; a data acquisition module configured to acquire multi-source data during the operation of the electromechanical system; A data preprocessing module is configured to preprocess the acquired multi-source data, including data cleaning, normalization and feature extraction; A digital twin modeling module is configured to build a digital twin model based on pre-processed multi-source data, integrating electromechanical system design drawings, 3D geometric models, material properties, and dynamic equations; A model dynamic calibration module is configured to input real-time collected data into the digital twin model through an adaptive Kalman filter or particle filter algorithm during the operation of the electromechanical system, dynamically calibrate the model parameters, and output simulation data of the real-time operating status of the electromechanical system; A fault feature analysis module is configured to establish a multi-dimensional fault feature indicator system based on simulation data, use a data fusion algorithm to fuse and analyze fault features, identify potential faults, and generate a comprehensive fault feature vector; A fault prediction module is configured to input the comprehensive fault feature vector into a pre-trained long short-term memory network or support vector machine prediction model, calculate the fault probability and remaining life in combination with historical fault data, and output the fault prediction result; The maintenance decision module is configured to generate maintenance decision suggestions based on the fault prediction results and the preset maintenance strategy rule base, and feed them back to the operator through the human-computer interaction interface, and simultaneously trigger the early warning mechanism.

[0042] In this embodiment, specifically: the model dynamic calibration module includes a data assimilation unit, an error evaluation unit, and a threshold triggering unit; a data assimilation unit configured to fuse real-time sensor data with the simulation output of the digital twin model through an adaptive Kalman filter or a particle filter algorithm to generate calibrated model parameters; an error evaluation unit configured to calculate in real time the root mean square error or cosine similarity between the model output and the actual sensor data to generate a model matching index; The threshold trigger unit is configured as a preset parameter update threshold. If the model matching index exceeds the threshold, the data assimilation unit is automatically triggered to dynamically calibrate the model parameters.

[0043] In this embodiment, specifically: the fault feature analysis module includes a feature index construction unit, a cross-modal fusion unit and an anomaly identification unit; A characteristic index construction unit is configured to establish a multi-dimensional fault characteristic index system including a vibration signal kurtosis value, a temperature data gradient change rate, a current signal harmonic distortion rate, and a pressure data fluctuation entropy value based on simulation data; A cross-modal fusion unit is configured to use an improved DS evidence theory, a neural network algorithm, or a fuzzy logic algorithm to fuse the simulation features from the digital twin model with the measured sensor features to generate a comprehensive fault feature vector including time domain, frequency domain, and energy features; The anomaly identification unit is configured to identify potential fault features through a preset fault feature threshold or a statistical process control method, and mark the weight coefficient of the abnormal dimension.

[0044] In summary, this embodiment provides a digital twin-based electromechanical system fault pre-diagnosis system, which acquires multi-source data through the data acquisition module, and after cleaning, normalization and feature extraction by the data preprocessing module, the digital twin modeling module constructs a digital twin that integrates the multi-dimensional model; the model dynamic calibration module uses adaptive Kalman filtering and other algorithms to calibrate the model parameters in real time, the fault feature analysis module constructs a multi-dimensional feature system based on simulation data and fuses it to generate a comprehensive vector, the fault prediction module outputs the fault probability and remaining life through the LSTM / SVM model, and finally the maintenance decision module generates maintenance suggestions according to the rules and triggers early warnings, forming a full-process intelligent pre-diagnosis closed loop from data acquisition to maintenance decision, thereby improving the fault identification accuracy and maintenance efficiency of the electromechanical system.

[0045] For other details about the technical solutions for implementing each module in the electromechanical system fault pre-diagnosis system based on digital twin in the above embodiment, please refer to the description of the electromechanical system fault pre-diagnosis method based on digital twin in the above embodiment, which will not be repeated here.

[0046] Example 3 This embodiment provides a practical application of a digital twin-based electromechanical system fault pre-diagnosis method in a complex electromechanical system, such as a large stamping production line in a certain automobile manufacturing plant. The details are as follows: In this stamping production line, the core electromechanical equipment includes multiple high-speed stamping machines, automated feeding devices, and robotic arms. To obtain multi-source data, various sensors are densely deployed at key locations: vibration sensors are installed on the crankshaft, connecting rod, and other parts of the stamping machine to collect three-dimensional vibration signals with a frequency range of 0-10kHz; temperature sensors are installed on the motor housing and drive module with an accuracy of up to ±0.3°C; pressure sensors are installed on the transmission chain and guide rail sliders of the feeding device, with a measuring range adapted to the operating pressure of the equipment; current and voltage sensors are arranged in the electrical control cabinet to monitor changes in electrical parameters; these sensors transmit data to the edge computing node in real time via industrial Ethernet, with transmission delays controlled within 30ms.

[0047] During the data preprocessing stage, the vibration signal is denoised using a median filter algorithm (with a window size of 5) to remove abnormal noise caused by equipment impact, electromagnetic interference, etc. The Z-score normalization method is used to unify the dimensions of temperature, current, and other data, so that their mean is 0 and their standard deviation is 1. The fast Fourier transform (FFT) is used to extract frequency domain features from the vibration signal, and the wavelet transform is used to obtain the energy characteristics of the signal. The empirical mode decomposition (EMD) is used to extract time domain statistical features such as the mean and variance.

[0048] When building a digital twin model, CAD software is first used to create a detailed three-dimensional model of each device in the stamping production line, with detailed annotation of key component dimensions, assembly relationships and other information. Then, professional simulation software is imported for geometric repair and simplification to generate a parametric geometric model. Based on physical laws such as material mechanics and dynamics, a stress distribution model of mechanical components, a thermal model of the motor and transmission system, and an electromechanical coupling model are constructed to accurately describe the physical processes in the operation of the equipment. With the help of neural networks (such as LSTM networks) and state-space models, combined with years of historical operation data of the production line, a behavioral model is constructed to simulate the dynamic response characteristics of the equipment under different stamping frequencies and load conditions, effectively capturing the nonlinear laws in the operation of the equipment. At the same time, based on the experience of industry experts and past failure cases, a rule model is constructed to define diagnostic rules such as "the vibration amplitude exceeds a certain threshold and lasts for more than 5 seconds → there is a loose mechanical component failure", and form a fault knowledge graph. Finally, a mature model integration framework is used to deeply couple the parametric geometric model, multi-domain physical model, behavioral model and rule model to form a unified digital twin model that can accurately reflect the actual operating conditions of the stamping production line.

[0049] During production line operation, the sensor data collected in real time is fused with the simulation output of the digital twin model through the adaptive Kalman filter algorithm; Specifically, the algorithm dynamically adjusts weights based on the uncertainty of the data and continuously updates model parameters to ensure that the model can closely follow changes in the actual operating status of the equipment; by calculating the root mean square error (RMSE) between the model output and the actual sensor data in real time, the RMSE threshold is set to 0.03g (g is the acceleration of gravity). When the error exceeds the threshold, the model parameter update process is automatically triggered; at the same time, the cosine similarity between the model output and the measured data is calculated. When the similarity is lower than 0.9, the calibration mechanism is also activated to ensure the high matching and accuracy of the model.

[0050] A multi-dimensional fault feature indicator system was established based on the simulation data output by the digital twin model. In addition to the kurtosis value of the vibration signal, the gradient change rate of the temperature data, the harmonic distortion rate of the current signal, and the fluctuation entropy value of the pressure data, this system also incorporates stamping process parameters such as the stamping speed change rate and the die closing time deviation. These fault features were fused using an improved DS evidence theory, with each feature appropriately weighted. For example, the weight of the vibration kurtosis value was set to 0.35, and the weight of the temperature gradient change rate was set to 0.25. The confidence level of each feature was comprehensively calculated to identify potential faults. A comprehensive fault feature vector was generated, which included time-domain, frequency-domain, energy-domain, and process-related features. The comprehensive fault feature vector is input into a pre-trained long short-term memory (LSTM) prediction model. The model combines historical fault data with a complex time series analysis algorithm to calculate the probability of failure and remaining life. When training the LSTM model, a large number of historical data samples covering different states such as normal operation, early failures, and severe failures are used. The model parameters are optimized through multiple iterations to give the model powerful fault prediction capabilities. The model outputs fault prediction results, such as remaining life prediction accurate to hours and the probability of failure presented as a percentage.

[0051] Based on the fault prediction results, maintenance decision recommendations are generated according to the preset maintenance strategy rule base. For example, when the predicted remaining life of a certain stamping machine is less than 48 hours and the probability of failure exceeds 80%, an emergency maintenance work order is immediately generated. The work order lists in detail the spare parts that need to be replaced (such as severely worn molds, aging motor brushes, etc.), plans the maintenance time window (such as shutting down for maintenance within the next 8-12 hours), and simultaneously triggers the production scheduling system to start a temporary capacity allocation plan to transfer part of the production tasks to other standby stamping machines to reduce the impact on overall production. If the predicted remaining life of the equipment is greater than 48 hours, but the probability of failure is gradually increasing, generating a progressive maintenance strategy, it is recommended to increase the frequency of equipment status monitoring, such as increasing the original daily vibration monitoring to once every 4 hours, and at the same time, according to the equipment operating conditions, appropriately reduce the stamping speed to reduce the equipment load and delay the development of failures; through a specially developed human-computer interaction interface, maintenance decision suggestions are intuitively fed back to the operator, and the interface displays equipment health status, fault type, maintenance measures and other information in the form of a combination of charts and texts, and simultaneously triggers the early warning mechanism through sound and light alarms to ensure that the operator can respond in time and take corresponding measures.

[0052] After a period of actual operation verification, after applying the electromechanical system fault pre-diagnosis method based on digital twin proposed in the present invention, the stamping production line reduced unplanned downtime by 68%, reduced equipment maintenance costs by 35%, and significantly improved production efficiency, fully demonstrating the effectiveness and practicality of the present invention in the pre-diagnosis of complex electromechanical system faults.

[0053] Example 4 Based on the first embodiment, the present invention provides a method for predicting faults in an electromechanical system based on digital twins, comprising the following steps: Based on pre-processed multi-source data, the digital twin model is constructed by integrating electromechanical system design drawings, 3D geometric models, material properties, and dynamic equations. Based on the Bayesian rule updating framework, dynamic adjustment of rule confidence is achieved: The initial rule confidence is set by expert experience. After each fault occurs, the confidence is updated using Bayesian posterior probability based on the matching degree between the actual fault type and the rule triggering result. During the operation of the electromechanical system, the real-time collected data is input into the digital twin model through the adaptive Kalman filter or particle filter algorithm, the model parameters are dynamically calibrated, and the simulation data of the real-time operating status of the electromechanical system is output; An exponential relationship model between vibration kurtosis and bearing stiffness degradation coefficient is established. When the kurtosis value exceeds the threshold, the stiffness parameter update is automatically triggered, replacing the traditional fixed threshold calibration. A multi-dimensional fault feature index system is established based on simulation data. The data fusion algorithm is used to fuse and analyze the fault features, identify potential faults, and generate a comprehensive fault feature vector. Monitor the rate of change of process parameters in real time and separate process disturbance components from fault components through variational mode decomposition; The dynamic calibration of the exponential relationship model between the vibration kurtosis value and the bearing stiffness degradation coefficient adopts a wear-parameter mapping function ΔK=α·e^(β·Kurtosis)+γ, where α, β and γ are calibrated through accelerated life tests.

[0054] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similarities between the various embodiments can be referred to in conjunction with each other. For system-related embodiments, since they are generally similar to method-related embodiments, their description is relatively simple. For relevant details, refer to the description of the method-related embodiments.

[0055] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting faults in electromechanical systems based on digital twins, characterized in that: The following steps are involved: Acquire multi-source data during the operation of the electromechanical system; pre-process the acquired multi-source data; Build a digital twin model based on pre-processed multi-source data, integrating electromechanical system design drawings, 3D geometric models, material properties, and dynamic equations; During the operation of the electromechanical system, the real-time collected data is input into the digital twin model through the adaptive Kalman filter or particle filter algorithm, the model parameters are dynamically calibrated, and the simulation data of the real-time operating status of the electromechanical system is output; A multi-dimensional fault feature index system is established based on simulation data. The data fusion algorithm is used to fuse and analyze the fault features, identify potential faults, and generate a comprehensive fault feature vector. The comprehensive fault feature vector is input into the pre-trained long short-term memory network or support vector machine prediction model, and the fault probability and remaining life are calculated in combination with historical fault data, and the fault prediction result is output.

2. A method for predicting faults in electromechanical systems based on digital twins according to claim 1, characterized in that: The method for constructing a digital twin model includes the following steps: Use modeling tools to build a 3D model of the electromechanical system, and import it into simulation software for geometric repair and simplification, generate a parametric geometric model, and map the structural features of mechanical components; Building a multi-domain physical model based on physical laws, wherein the multi-domain physical model includes stress distribution of mechanical components, temperature field changes, and electromechanical coupling relationships; Based on neural network or state space model, combined with historical operation data, behavioral model is constructed to simulate the dynamic response characteristics of the system under different working conditions and capture nonlinear operation laws; Build a rule model based on expert experience and historical fault data, define fault diagnosis rules and maintenance decision logic, and form a knowledge graph; The parametric geometric model, multi-domain physical model, behavioral model and rule model are coupled through the model integration framework to form a unified digital twin model.

3. The electromechanical system fault pre-diagnosis method based on digital twin according to claim 1 is characterized in that: The model parameters are dynamically calibrated, and the matching degree of the digital twin model is evaluated in real time by calculating the root mean square error or cosine similarity between the model output and the actual sensor data, thereby triggering the parameter update mechanism.

4. A method for predicting faults in electromechanical systems based on digital twins according to claim 1, characterized in that: The multi-dimensional fault characteristic indicator system includes the kurtosis value of the vibration signal, the gradient change rate of the temperature data, the harmonic distortion rate of the current signal and the fluctuation entropy value of the pressure data.

5. The electromechanical system fault pre-diagnosis method based on digital twin according to claim 1 is characterized in that: The multi-source data is collected through vibration sensors, temperature sensors, pressure sensors and current sensors deployed at key parts of the electromechanical system; the multi-source data includes vibration signals, temperature data, pressure data and current and voltage data.

6. A method for predicting faults in electromechanical systems based on digital twins according to claim 1, characterized in that: The preprocessing includes data cleaning, normalization and feature extraction. The data cleaning uses median filtering or Gaussian filtering algorithm to remove abnormal noise; the normalization process unifies the data dimension through minimum-maximum normalization or Z-score normalization; the feature extraction uses fast Fourier transform, wavelet transform or empirical mode decomposition to extract frequency domain features, energy features and time domain statistical features.

7. The electromechanical system fault pre-diagnosis method based on digital twin according to claim 1, characterized in that: The method further includes generating maintenance decision suggestions based on the fault prediction results and a preset maintenance strategy rule base, feeding back the suggestions to the operator through a human-computer interaction interface, and simultaneously triggering an early warning mechanism. The process of generating maintenance decision suggestions includes: Based on the remaining life prediction value in the fault prediction results and the weight of the impact of equipment downtime on the production plan, the priority coefficient of the maintenance task is calculated; If the remaining life prediction value is less than or equal to the preset threshold, an emergency maintenance work order is generated, including a repair time window, a spare parts replacement list, and a temporary capacity allocation plan. If the remaining life prediction value is greater than the preset threshold but the probability of failure continues to increase, a progressive maintenance strategy is generated, including adjustment of condition monitoring frequency and local load reduction operation suggestions.

8. A mechatronic system fault pre-diagnosis system based on digital twins, applied to a mechatronic system fault pre-diagnosis method based on digital twins according to any one of claims 1 to 7, characterized in that: It includes data acquisition module, data preprocessing module, digital twin modeling module, model dynamic calibration module, fault feature analysis module, fault prediction module and maintenance decision module; The data acquisition module is configured to acquire multi-source data during the operation of the electromechanical system; The data preprocessing module is configured to preprocess the acquired multi-source data, wherein the preprocessing includes data cleaning, normalization and feature extraction; The digital twin modeling module is configured to construct a digital twin model based on preprocessed multi-source data, integrating electromechanical system design drawings, three-dimensional geometric models, material properties and dynamic equations; The model dynamic calibration module is configured to input real-time collected data into the digital twin model through an adaptive Kalman filter or a particle filter algorithm during the operation of the electromechanical system, dynamically calibrate the model parameters, and output simulation data of the real-time operating status of the electromechanical system; The fault feature analysis module is configured to establish a multi-dimensional fault feature indicator system based on simulation data, use a data fusion algorithm to fuse and analyze fault features, identify potential faults and generate a comprehensive fault feature vector; The fault prediction module is configured to input the comprehensive fault feature vector into a pre-trained long short-term memory network or support vector machine prediction model, calculate the fault probability and remaining life in combination with historical fault data, and output the fault prediction result; The maintenance decision module is configured to generate maintenance decision suggestions based on the fault prediction results and a preset maintenance strategy rule library, and feed them back to the operator through the human-computer interaction interface, and simultaneously trigger the early warning mechanism.

9. The electromechanical system fault prediction system based on digital twin according to claim 8, characterized in that: The model dynamic calibration module includes a data assimilation unit, an error evaluation unit and a threshold triggering unit; The data assimilation unit is configured to fuse the real-time sensor data with the simulation output of the digital twin model through an adaptive Kalman filter or a particle filter algorithm to generate calibrated model parameters; The error evaluation unit is configured to calculate the root mean square error or cosine similarity between the model output and the actual sensor data in real time to generate a model matching index; The threshold trigger unit is configured to preset a parameter update threshold. If the model matching index exceeds the threshold, the data assimilation unit is automatically triggered to dynamically calibrate the model parameters.

10. The electromechanical system fault prediction system based on digital twin according to claim 8, characterized in that: The fault feature analysis module includes a feature index construction unit, a cross-modal fusion unit and an anomaly identification unit; The characteristic index construction unit is configured to establish a multi-dimensional fault characteristic index system including the vibration signal kurtosis value, the temperature data gradient change rate, the current signal harmonic distortion rate, and the pressure data fluctuation entropy value based on the simulation data; The cross-modal fusion unit is configured to use an improved DS evidence theory, a neural network algorithm, or a fuzzy logic algorithm to fuse the simulation features from the digital twin model with the measured sensor features to generate a comprehensive fault feature vector including time domain, frequency domain, and energy features; The abnormality identification unit is configured to identify potential fault features through a preset fault feature threshold or a statistical process control method, and mark the weight coefficient of the abnormal dimension.

Images