A fault diagnosis method based on digital twinning

Abstract

This invention discloses a fault diagnosis method based on digital twins, comprising the following steps: 1) constructing a corresponding virtual model based on the physical characteristics of a physical model; 2) running the physical model and the virtual model respectively to obtain measured and virtual values ​​of several features; 3) selecting parameters to be optimized based on a distance metric method, and then updating the digital twin virtual model using a particle swarm optimization algorithm; 4) running the updated virtual model, constructing a training set based on the generated data, and training a convolutional neural network; 5) constructing a test set based on the data obtained from the physical model; analyzing the test set using the trained convolutional neural network, and outputting the diagnostic effect on equipment faults. This invention improves the virtual model through correlation analysis between virtual and real data, solving the problem that existing data-driven fault diagnosis methods do not consider the insufficient amount of data required by the algorithm.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing, and in particular to a fault diagnosis method based on digital twins. Background Technology

[0002] Traditional machine learning or deep learning algorithms for preventative maintenance of industrial equipment rely on historical data. Obtaining this data requires extensive equipment, which is costly and results in insufficient data for model training. Furthermore, training machine learning models using historical data cannot guarantee the reliability of fault diagnosis for equipment whose lifespan could have been extended.

[0003] To address the aforementioned issues, the field of intelligent manufacturing has introduced the digital twin theory and technology system.

[0004] Digital twins fully utilize data from physical models, sensor updates, and operational history to integrate virtual processes across multiple disciplines, physical quantities, scales, and probabilities. This results in a mapping within a virtual space, reflecting the entire lifecycle of the corresponding physical equipment. Digital twins are a concept that transcends reality; they can be viewed as digital mapping systems of one or more important, interdependent equipment systems. By constructing virtual models that reflect the physical systems using digital twin principles, and generating sufficiently diverse data in real time through these virtual models, digital twins can overcome the limitations of existing fault diagnosis methods that rely solely on historical data to train machine learning models. Summary of the Invention

[0005] The purpose of this invention is to provide a fault diagnosis method based on digital twins. By analyzing the correlation between virtual and real data, the virtual model is continuously improved, thereby simulating sufficiently diverse data for machine learning algorithms. This addresses the shortcomings of existing fault diagnosis methods that can only train machine learning models with historical data, and better serves the preventive maintenance of industrial equipment.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0007] A fault diagnosis method based on digital twins includes the following steps:

[0008] 1) Based on the physical characteristics of the physical model, construct its corresponding virtual model;

[0009] 2) Select several features, run the physical model and virtual model respectively, and obtain the measured values ​​and virtual values ​​of the features;

[0010] 3) Based on the distance metric method, select the parameters that need to be optimized in the virtual model from the measured values ​​and the virtual values, and then update the virtual model using the particle swarm optimization algorithm;

[0011] 4) Run the updated virtual model, construct a training set based on the data generated by the virtual model, and train a convolutional neural network using the training set;

[0012] 5) Construct a test set based on the data obtained from the physical model; analyze the test set using a trained convolutional neural network and output the diagnostic effect on equipment faults.

[0013] Furthermore, the physical model is a three-phase asynchronous motor fault test platform, which includes a controllable load motor, a rotor connected to the drive end of the controllable load motor, and two ball bearings mounted on the rotor; there is a fault point on the outer ring and inner ring of each ball bearing, and sensors for acquiring data are respectively provided in the horizontal, vertical, and axial directions at the drive end of the controllable load motor.

[0014] Furthermore, the process of constructing the virtual digital twin model is as follows:

[0015] 1) Determine the equations of motion of the physical entity

[0016] The rotor is considered as a typical multi-node Timoshenko beam, and two points are selected to represent the dynamic characteristics of the physical system; each point is restricted to three degrees of freedom.

[0017] in, and Represent each direction and The curvature, represent The direction of the torsion is reflected by considering the shear effect to reflect the cross-sectional change of the shaft; the governing equations of the physical system considering inertial force, restoring force and damping force as well as constant vertical force acting on the inner ring of the bearing are shown in equations (1) and (2);

[0018] + + (1)

[0019] (2)

[0020] in, This indicates the mass of the rotor supported by the bearing and the mass of the inner ring. Represents the equivalent viscous damping coefficient. It refers to the number of balls. Represents the Hertzian contact elastic deformation constant. Indicates internal radial clearance. Indicates the first The amplitude of the surface ripples at each ball position Indicates a radially controllable load. This indicates the force generated by rotor imbalance. It is the first The angular position of each rolling element It is the main displacement of the inner circle center. This indicates the angular velocity of the cage-type guide bar / outer ring / inner ring. Indicates time;

[0021] The subscript + sign in equations (1) and (2) indicates that...

[0022] when When, it indicates the loading angle position. The rolling elements generate restoring force;

[0023] when When i is not loaded, it means that the rolling body at angular position i has a restoring force of 0.

[0024] The equation can be solved using Newton's method and the implicit Newmark method to obtain the displacement, velocity, and acceleration of each point in the system.

[0025] 2) Bearing fault modeling

[0026] The bearing inner and outer ring faults are modeled as small segments with a sinusoidal half-wave shape and an angular width of . Depth is When each ball passes through the defect area, it is introduced into the virtual model of the motor fault test platform by increasing the radial clearance; the instantaneous restoring force when the bearing fails is calculated by formulas (3) and (4);

[0027] (3)

[0028] (4)

[0029] in, The extra gap is represented by formula (5);

[0030] (5)

[0031] in, Indicates the angular location of the defect. Indicates the angular width of the defect. The relative angle between ball i and the defect location is a function of the bearing cage angular velocity as expressed in formula (6);

[0032] .

[0033] Furthermore, the specific steps of step 3) are as follows:

[0034] The root mean square (RMS), kurtosis, peak value, crest factor, and skewness are directly extracted from the time domain. The average frequency, frequency center, RMS of the frequency distribution, standard deviation of the frequency distribution, and power envelope spectrum are extracted from the frequency domain. The effective value and average envelope spectrum of the frequency distribution are extracted from the time-scale domain. Then, based on the distance metric method, the similarity between the measured feature values ​​obtained from the physical system and the feature values ​​output by the dynamic model are analyzed. The top 6 parameters are selected, and the virtual model is updated using the particle swarm optimization algorithm.

[0035] In summary, the present invention has the following beneficial effects:

[0036] Based on the digital twin paradigm, a virtual model of the faulty equipment is constructed, parameters are selected using a distance metric-based method, and then the particle swarm optimization algorithm is used to update it. This ensures that the virtual model can fully reflect the operating state of the physical model. The sufficiently diverse data generated by the virtual model solves the problem of insufficient data during the training of machine learning models.

[0037] When the input dimension of the dataset generated by the virtual model is 1024, the present invention can achieve an average accuracy of 99.48% on the training set and an average accuracy of 99.29% on the test set. The use of digital twins to generate virtual data can improve the reliability of fault diagnosis. Attached Figure Description

[0038] Figure 1 It is a virtual model diagram corresponding to the physical entity described in this invention.

[0039] Figure 2 This is the parameter selection and optimization diagram of the virtual model based on distance metric described in this invention.

[0040] Figure 3 This is a classification framework diagram of the machine learning algorithm based on digital twins as described in this invention. Detailed Implementation

[0041] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to the figures and specific embodiments.

[0042] like Figure 1 , Figure 2 and Figure 3 As shown, the present invention proposes a fault diagnosis method based on digital twins, which includes the following steps:

[0043] 1) Based on the physical characteristics of the physical model, construct its corresponding virtual model;

[0044] 2) Select several features, run the physical model and virtual model respectively, and obtain the measured values ​​and virtual values ​​of the features;

[0045] 3) Based on the distance metric method, select the parameters that need to be optimized for the virtual model from the measured values ​​and virtual values, and then use the particle swarm optimization algorithm to update the digital twin virtual model;

[0046] 4) Run the updated digital twin simulacrum, construct a training set based on the data generated by the virtual model, and train a convolutional neural network using the training set;

[0047] 5) Construct a test set based on the data obtained from the physical model; analyze the test set using a trained convolutional neural network and output the diagnostic effect on equipment faults.

[0048] Example

[0049] I. Prepare a three-phase asynchronous motor fault testing platform

[0050] The main components of the platform are a rotor driven by a faulty asynchronous motor, two 6206-RZ type ball bearings of the motor itself, and a controllable load motor. The outer ring of the bearing at the motor drive end has a fault of 1.778 mm, and the inner ring has a fault of 0.356 mm. Sensors are installed at the motor drive end in three directions: horizontal, vertical, and axial. Vibration data throughout the entire operating state is collected using a CT-9208 vibration signal acquisition system at a sampling frequency of 12 kHz. Whether in a healthy state or with two fault conditions, the motor drive speed is 1420 rpm, and the applied radial load is 2250 W during data acquisition. For each bearing in a healthy state, a 1-second vibration signal is acquired and divided into 10 0.1-second samples, providing 500 samples for each fault.

[0051] II. Constructing a digital twin model of the motor fault testing platform

[0052] 1) Determine the equations of motion of the physical entity

[0053] Determining the system's equations of motion essentially involves realizing the dynamic characteristics of the physical system, the motor fault testing platform. To express the motion restoration equations using the Lagrange energy formula, the rotor is considered a typical multi-node Timoshenko beam, with two points selected to represent the dynamic characteristics of the physical system. Figure 1 Each point is limited to 3 degrees of freedom.

[0054] in, and Represent each direction and The curvature, represent The direction of the torsion is reflected by considering the shear effect to reflect the cross-sectional change of the shaft; the governing equations of the physical system considering inertial force, restoring force and damping force as well as constant vertical force acting on the inner ring of the bearing are shown in equations (1) and (2);

[0055] + + (1)

[0056] (2)

[0057] in, This indicates the mass of the rotor supported by the bearing and the mass of the inner ring. Represents the equivalent viscous damping coefficient. It refers to the number of balls. Represents the Hertzian contact elastic deformation constant. Indicates internal radial clearance. Indicates the first The amplitude of the surface ripples at each ball position Indicates a radially controllable load. This indicates the force generated by rotor imbalance. It is the first The angular position of each rolling element It is the main displacement of the inner circle center. This indicates the angular velocity of the cage-type guide bar / outer ring / inner ring. Indicates time;

[0058] The subscript + sign in equations (1) and (2) indicates that...

[0059] when When, it indicates the loading angle position. The rolling elements generate restoring force;

[0060] when When i is not loaded, it means that the rolling body at angular position i has a restoring force of 0.

[0061] The equation can be solved using Newton's method and the implicit Newmark method to obtain the displacement, velocity, and acceleration of each point in the system.

[0062] 2) Bearing fault modeling

[0063] The bearing inner and outer ring faults are modeled as small segments with a sinusoidal half-wave shape and an angular width of . Depth is When each ball passes through the defect area, it is introduced into the virtual model of the motor fault test platform by increasing the radial clearance; the instantaneous restoring force when the bearing fails is calculated by formulas (3) and (4);

[0064] (3)

[0065] (4)

[0066] in, The extra gap is represented by formula (5);

[0067] (5)

[0068] in, Indicates the angular location of the defect. Indicates the angular width of the defect. The relative angle between ball i and the defect location is a function of the bearing cage angular velocity as expressed in formula (6);

[0069] .

[0070] 3) Dynamic model updates

[0071] The purpose of employing a dynamic model update strategy is to ensure that the constructed virtual model accurately reflects the physical system. To implement this strategy, displacement, velocity, and acceleration data at each node during system operation need to be collected. Twelve features are extracted from the recorded signals. Specifically, root mean square (RMS), kurtosis, peak value, crest factor, and skewness are extracted directly from the time domain; average frequency, frequency center, RMS of the frequency distribution, standard deviation of the frequency distribution, and power envelope spectrum are extracted from the frequency domain; and the effective value and average envelope spectrum of the frequency distribution are extracted from the time-scale domain. Then, a distance metric-based method is used to analyze the similarity between the measured feature values ​​from the physical system and the feature values ​​output by the dynamic model. The top six parameters are selected, and the virtual model is updated using a particle swarm optimization algorithm.

[0072] III. Constructing a fault classification model based on digital twins

[0073] A training set is constructed based on data obtained from the virtual model. A traditional convolutional neural network is then trained using this training set. A test set is then constructed based on data obtained from the physical model. The trained model is then used to analyze the test set to observe its diagnostic effectiveness for equipment faults.

[0074] This invention is achieved through Figure 1 The physical characteristics of a three-phase asynchronous motor fault testing platform are modeled; then, based on... Figure 2 The selection of virtual models requires dynamically updated parameters; when the virtual model can simulate and reflect the physical system, based on... Figure 3Data from the physical system is collected, data generated by the virtual model is obtained, and a traditional convolutional neural network is trained using the data generated by the virtual model. A test set is constructed using the data collected from the physical system, and finally the trained convolutional neural network is used to analyze the fault classification effect in the test set.

[0075] In this document, the terms "upper," "lower," "front," "back," "left," "right," "top," "bottom," "inner," "outer," "vertical," and "horizontal," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only used for the clarity of expressing the technical solution and for the convenience of description, and therefore should not be construed as limiting the present invention.

[0076] In this document, the terms “including,” “comprising,” or any other variations thereof are intended to cover non-exclusive inclusion, which includes not only the elements listed but also other elements not expressly listed.

[0077] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A fault diagnosis method based on digital twins, characterized in that, Includes the following steps: 1) Based on the physical characteristics of the physical model, construct its corresponding virtual model; 2) Select several features, run the physical model and virtual model respectively, and obtain the measured values ​​and virtual values ​​of the features; 3) Based on the distance metric method, select the parameters that need to be optimized in the virtual model from the measured values ​​and the virtual values, and then update the virtual model using the particle swarm optimization algorithm; The specific steps of step 3) are as follows: The root mean square (RMS), kurtosis, peak value, crest factor, and skewness are directly extracted from the time domain; the average frequency, frequency center, RMS of the frequency distribution, standard deviation of the frequency distribution, and power envelope spectrum are extracted from the frequency domain; and the effective value and average envelope spectrum of the frequency distribution are extracted from the time-scale domain. Then, based on the distance metric method, the similarity between the measured feature values ​​obtained from the physical system and the feature values ​​output by the dynamic model is analyzed, and the top 6 parameters are selected. The virtual model is then updated using the particle swarm optimization algorithm. 4) Run the updated virtual model, construct a training set based on the data generated by the virtual model, and train a convolutional neural network using the training set; 5) Construct a test set based on the data obtained from the physical model; analyze the test set using a trained convolutional neural network and output the diagnostic effect on equipment faults; The process of constructing the virtual model is as follows: 1) Determine the equations of motion of the physical entity The rotor is considered as a typical multi-node Timoshenko beam, and two points are selected to represent the dynamic characteristics of the physical system; each point is restricted to three degrees of freedom. in, and Represent each direction and The curvature, represent The direction of torsion is considered, and the cross-sectional changes of the shaft are reflected by taking into account the shear force effect; The governing equations of the physical system that takes into account inertial force, restoring force, damping force, and constant vertical force acting on the inner ring of the bearing are shown in equations (1) and (2); + + (1) (2) in, This indicates the mass of the rotor supported by the bearing and the mass of the inner ring. Represents the equivalent viscous damping coefficient. It refers to the number of balls. Represents the Hertzian contact elastic deformation constant. Indicates internal radial clearance. Indicates the first The amplitude of the surface ripples at each ball position Indicates a radially controllable load. This indicates the force generated by rotor imbalance. It is the first The angular position of each rolling element It is the main displacement of the inner circle center. This indicates the angular velocity of the cage-type guide bar / outer ring / inner ring. Indicates time; The subscript + sign in equations (1) and (2) indicates that... when When, it indicates the loading angle position. The rolling elements generate restoring force; when When, it indicates that the angular position is not loaded. The rolling element has a restoring force of 0; The equation is solved using Newton's method or implicit Newmark's method to obtain the displacement, velocity, and acceleration of each point in the system. 2) Bearing fault modeling The bearing inner and outer ring faults are modeled as small segments with a sinusoidal half-wave shape and an angular width of . Depth is When each ball passes through the defect area, it is introduced into the virtual model of the motor fault test platform by increasing the radial clearance; the instantaneous restoring force when the bearing fails is calculated by formulas (3) and (4); (3) (4) in, The extra gap is represented by formula (5); (5) in, Indicates the angular location of the defect. Indicates the angular width of the defect. The relative angle between ball i and the defect location is a function of the bearing cage angular velocity as expressed in formula (6); 。 2. The fault diagnosis method based on digital twins according to claim 1, characterized in that, The physical model is a three-phase asynchronous motor fault test platform, which includes a controllable load motor, a rotor connected to the drive end of the controllable load motor, and two ball bearings mounted on the rotor. There is a fault point on the outer ring and inner ring of each ball bearing. Sensors for acquiring data are provided in the horizontal, vertical, and axial directions at the drive end of the controllable load motor.

Images