Motor transfer learning fault diagnosis method based on digital-analog driving roller blade moisture regaining machine
Through the transfer learning technology combined with digital-analog driven multi-physics coupled model and deep convolutional neural network (DCNN), the problem of difficulty in extracting complex vibration signal features in roller motor fault diagnosis is solved, and the fault diagnosis effect with high accuracy and robustness is achieved.
Patent Information
- Application Number
- CN202411812575.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-10
- Publication Date
- 2025-02-28
AI Technical Summary
The prior art is difficult to effectively extract key features in the complex vibration signals of the drum motor, resulting in insufficient accuracy and robustness of fault diagnosis.
By establishing a digital-analog-driven multi-physics coupled model, high-precision simulation data is generated, and combined with deep convolutional neural network (DCNN) models are trained, and knowledge transfer is transferred from the simulation environment to the actual working conditions using transfer learning technology.
It significantly improves the accuracy and robustness of fault diagnosis, reduces dependence on large amounts of labeled data, reduces model training costs, and improves diagnostic efficiency and engineering application adaptability.
Smart Images
Figure QLYQS_1 
Figure QLYQS_2 
Figure QLYQS_6
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of motor fault diagnosis, and is used for motor health monitoring and fault predictive maintenance, and specifically relates to a motor fault diagnosis method based on transfer learning of digital-analog driven drum blade rehumidifiers. Background Art
[0002] With the development of industrial automation, motors, as important power equipment, have a running state that is directly related to the stability and safety of production. However, in actual working scenarios, due to long-term operation of motors, external load changes, power supply deviation, overheating, insufficient lubrication, poor sealing, product defects and natural aging, the motors will face many unavoidable stresses such as mechanical, electrical, thermal and environmental stresses during operation. Ultimately, these problems will cause the motors to fail, which will not only affect the normal operation of the equipment, but may also reduce the economic benefits of the enterprise. In addition, motor failures may cause a series of safety hazards and even cause safety accidents such as personal injury. Therefore, it is very important to ensure the normal operation of the motors at work.
[0003] The drum motor is the core driving component of the drum blade rehumidifier. Its performance directly determines the stability and production efficiency of the entire equipment. In actual operation, due to long-term high load, humid environment and variable working conditions, the drum motor is prone to various faults such as winding short circuit, bearing damage and rotor imbalance. These faults usually manifest as complex vibration signals, temperature rise changes and abnormal electrical characteristics. Traditional rule-based diagnostic methods are difficult to effectively extract key features and accurately classify them.
[0004] In recent years, deep learning methods have gradually become a research hotspot due to their excellent feature extraction and fault identification capabilities. However, relying solely on actual data for model training faces problems such as high sample collection costs and incomplete coverage of operating conditions. Although the method of generating data based on motor simulation modeling can solve the problem of insufficient experimental samples, the deviation of the simulation model may lead to a decrease in the performance of the diagnostic model in actual applications. Therefore, combining simulation and actual data, high-precision fault diagnosis from digital model to actual is achieved through transfer learning technology. Summary of the invention
[0005] The purpose of the present invention is to provide a motor fault diagnosis method for a drum blade rehumidifier based on digital-analog drive, so as to improve the motor operation reliability and solve the problems existing in the above-mentioned background technology.
[0006] To achieve the above objectives, this application is implemented through the following technical solutions:
[0007] The fault diagnosis method of the motor of the drum blade rehumidifier based on the digital analog drive is composed of the following steps:
[0008] S1: Obtain key parameters and working condition signals of the drum motor, and pre-process the acquired signals;
[0009] S2: Establish a mathematical model of the motor;
[0010] S3: Establish a multi-physics coupling model of the drum motor through MATLAB / Simulink to generate high-precision simulation data under multiple working conditions;
[0011] S4: Build a deep convolutional neural network model;
[0012] S5: Use the simulation data generated by S3 as the data set for the model built in S4 for model training, and use the actual data to correct the simulation model error, and continuously improve the performance of the model through interactive iterative optimization algorithms;
[0013] S6: Use transfer learning methods to achieve knowledge transfer from the simulation environment to the actual working conditions. By freezing the feature extraction layer of the pre-trained model and only fine-tuning the classification layer for the actual data, the training time and data requirements are reduced.
[0014] Furthermore, in step S1, the steps of obtaining the key parameters and operating condition signals of the drum motor and preprocessing the acquired signals are as follows:
[0015] S11: Obtain actual geometric parameters of the drum motor by checking relevant drawings and instructions;
[0016] S12: Obtain the electrical and mechanical parameters of the drum motor by checking the brand name and manual;
[0017] S13: Obtaining actual vibration data and actual current data of the motor during operation through sensors, and preprocessing the signals.
[0018] Furthermore, the preprocessing method is CEEMDAN or FIR filter, which reduces the noise in the signal through data preprocessing.
[0019] Furthermore, in step S2, a mathematical model of the motor is established, the rotor voltage equation is subjected to Park transformation, and the following assumptions are made: (1) the resistance in the stator and rotor windings is constant; (2) there is no saturation phenomenon in the magnetic circuit, and hysteresis and iron loss are ignored; (3) the three-phase windings are star-connected, and the spatial distribution is completely symmetrical, and the synthetic magnetomotive force is sinusoidally distributed in the air gap direction; (4) the influence of spatial harmonics is ignored;
[0020] The stator flux equation is:
[0021] (1)
[0022] The rotor flux equation is:
[0023] (2)
[0024] Where: L s , L r are the equivalent self-inductance of the stator and rotor, H; L m is the equivalent mutual inductance between the stator and rotor, , ; , H;
[0025] The electromagnetic torque equation is:
[0026] (3)
[0027] Where: p is the pole pair number;
[0028] The rotor angular velocity equation is:
[0029] (4)
[0030] Where: T e 、T m They are electromagnetic torque and load torque, N·m; n p is the rotor speed, r / min; J is the moment of inertia, kg·m².
[0031] Furthermore, in step S3, a multi-physics field coupling model of the drum motor is established through MATLAB / Simulink to generate high-precision simulation data under multiple working conditions. The specific steps are as follows:
[0032] S31: Input the key parameters of the drum motor in MATLAB / Simulink and build a modular multi-physics simulation framework based on electromagnetic, mechanical, and thermal mathematical models;
[0033] S32: Use Simulink and Simscape modules to achieve coupling between electromagnetic, mechanical, and thermal fields, and set key boundary conditions to ensure accurate interaction of physical fields;
[0034] S33: Run the simulation model for different working conditions, collect multi-physics field output data such as vibration signals and current signals, and improve the simulation accuracy through high-precision sampling and optimization algorithms.
[0035] Furthermore, in step S4, a network model for deep convolutional neural motor fault diagnosis is constructed, including the following four modules:
[0036] S41: a sample division module, which divides the total samples of the collected vibration signals and current signals into a training set, a verification set, and a test set according to a ratio of 7:2:1;
[0037] S42: branch feature extraction module, which is equivalent to extracting features in parallel with two convolution channels in structure, and can simultaneously extract features between the vibration signal and the current signal and save the features;
[0038] S43: Feature fusion module, which uses the features through the Lambda layer Functions are merged to complete feature weighted fusion, where It can be expressed as:
[0039] ,
[0040] in represents the tensor to be concatenated, Indicated in Dimensional splicing, A tensor representing the output;
[0041] S44: Fault diagnosis module, which uses the features fused in step S33 to complete the fault diagnosis of the motor, uses the test set to perform fault diagnosis on the trained model, inputs the data in the test set into the model, and observes the output results of the model.
[0042] Furthermore, in step S5, the specific steps are:
[0043] S51: Use the constructed data set to train the deep convolutional neural network model, and introduce actual operation data to correct the deviation of the simulation model output to improve the true reflection ability of the data set;
[0044] S52: Iteratively tune the simulation model and deep learning model based on the optimization algorithm, comprehensively analyze the training loss and verification error, continuously optimize the simulation model parameters and deep learning model hyperparameters, and achieve gradual performance improvement.
[0045] Furthermore, in step S6, the specific steps are:
[0046] S61: Load the deep convolutional neural network model pre-trained on the simulation data, freeze its feature extraction layer, and retain its learned multi-operating condition feature representation capability as the basic model for the actual operating condition data;
[0047] S62: According to the actual operating conditions, vibration signals and current signals are collected and used as inputs to fine-tune the classification layer of the pre-trained model, and the classification weights are re-optimized with less actual data;
[0048] S63: Verify the performance of the transfer learning model on the actual working condition data set. By analyzing the classification accuracy and model error, further adjust the classification layer parameters or unfreeze some high-level feature extraction layers to improve the model's adaptability and generalization ability.
[0049] The beneficial effects of the present invention are:
[0050] The present invention proposes a transfer learning fault diagnosis method based on a digital-analog driven drum blade rehumidifier motor, which includes four stages: signal acquisition and preprocessing, mathematical modeling and simulation, deep learning model construction, and transfer learning implementation. By generating and efficiently utilizing simulation data, this method effectively solves the problem that fault data is difficult to obtain under actual working conditions. By combining multi-physics field coupling simulation with deep convolutional neural network (DCNN), the feature correlation between vibration signals and current signals is fully explored, which significantly improves the accuracy and robustness of fault diagnosis. In addition, by freezing the feature extraction layer and only fine-tuning the classification layer, the transfer learning strategy enables the model to better adapt to the complexity of actual working conditions, reduces dependence on a large amount of labeled data, and reduces the cost of model training. The overall method has significant advantages in diagnostic efficiency and adaptability to engineering applications. DETAILED DESCRIPTION
[0051] The technical solution of the present invention is described in detail below through examples. The following examples are merely exemplary and can only be used to explain and illustrate the technical solution of the present invention, but cannot be construed as limiting the technical solution of the present invention.
[0052] The present application provides a fault diagnosis method for a motor of a drum blade rehumidifier based on digital-analog drive transfer learning, the method comprising the following steps:
[0053] S1: Obtain the key parameters and working condition signals of the drum motor and pre-process the collected signals. The specific steps are as follows:
[0054] S11: Obtain the actual geometric parameters of the drum motor by checking the relevant drawings and instructions.
[0055] S12: Obtain the electrical and mechanical parameters of the drum motor by checking the brand name and manual.
[0056] S13: Obtain the actual vibration data z(t) and the actual current data d(t) of the motor during operation through sensors, and pre-process the signals.
[0057] In the present application, the preprocessing method is CEEMDAN or FIR filter, and the purpose of data preprocessing is to reduce the noise in the signal.
[0058] S2: Establish a mathematical model of the drum motor. Due to the relative motion between the stator and the rotor, the dynamic mathematical model of the squirrel cage rotor is very complex. In order to simplify the mathematical model of the motor, the rotor voltage equation is transformed by Park and the following assumptions are made: (1) The resistance in the stator and rotor windings is a constant value; (2) There is no saturation phenomenon in the magnetic circuit, and hysteresis and iron loss are ignored; (3) The three-phase windings are star-connected and the spatial distribution is completely symmetrical. The synthetic magnetomotive force is sinusoidally distributed in the air gap direction; (4) The influence of spatial harmonics is ignored.
[0059] The stator flux equation is:
[0060] (1)
[0061] The rotor flux equation is:
[0062] (2)
[0063] Where: L s , L r are the equivalent self-inductance of the stator and rotor, H; L m is the equivalent mutual inductance between the stator and rotor, , ; , H.
[0064] The electromagnetic torque equation is:
[0065] (3)
[0066] Where: p is the pole pair number.
[0067] The rotor angular velocity equation is:
[0068] (4)
[0069] Where: T e 、T m They are electromagnetic torque and load torque, N·m; n p is the rotor speed, r / min; J is the moment of inertia, kg·m².
[0070] S3: Use MATLAB / Simulink to establish a multi-physics coupling model of the drum motor and generate high-precision simulation data under multiple working conditions. The specific steps are as follows:
[0071] S31: Input the key parameters of the drum motor in MATLAB / Simulink and build a modular multi-physics simulation framework based on electromagnetic, mechanical, and thermal mathematical models.
[0072] S32: Use Simulink and Simscape modules to achieve coupling between electromagnetic, mechanical, and thermal fields, and set key boundary conditions to ensure accurate interaction of physical fields.
[0073] S33: Run the simulation model for different working conditions, collect multi-physics field output data such as vibration signals and current signals, and improve the simulation accuracy through high-precision sampling and optimization algorithms.
[0074] S4: Construct a network model for deep convolutional neural motor fault diagnosis. The model contains the following four modules, and the specific steps are as follows:
[0075] S41: Sample division module. This module divides the total samples of vibration signals and current signals collected into training set, verification set, and test set in a ratio of 7:2:1. The training set is used to train the model and is the main source of data features for model learning. The verification set is used to adjust the hyperparameters of the model and evaluate the performance of the model on unseen data to prevent overfitting. The test set is used to finally evaluate the performance of the model and is a test of the model's practical application capabilities.
[0076] S42: a multi-branch feature extraction module, which is structurally equivalent to two convolution channels extracting features in parallel, and can simultaneously extract features between vibration signals and current signals and save the features.
[0077] S43: Feature fusion module, which uses the features through the Lambda layer Functions are combined to complete feature weighted fusion. It can be expressed as:
[0078] ,
[0079] in represents the tensor to be concatenated, Indicated in Dimensional splicing. A tensor representing the output.
[0080] S44: Fault diagnosis module, which uses the features fused by S3-3 to complete the fault diagnosis of the motor. Use the test set to perform fault diagnosis on the trained model, input the data in the test set into the model, and observe the output of the model. By comparing the results predicted by the model and the actual labels, the accuracy and performance of the model are evaluated. And the classification results of each type of fault can be obtained through SoftMax. Among them, SoftMax can be expressed as:
[0081] ,
[0082] in Refers to the category of classification, refers to the output of the classification, Refers to exponential function. Thus, motor compound fault diagnosis and classification can be realized.
[0083] S5: Use the simulation data generated by S3 as the data set for constructing the model in S4 for model training, and use the actual data to correct the simulation model error, and continuously improve the performance of the model through interactive iterative optimization algorithm. The specific steps are as follows:
[0084] S51: Use the constructed data set to train the deep convolutional neural network model, and introduce actual operation data to correct the deviation of the simulation model output to improve the true reflection ability of the data set.
[0085] S52: Iteratively tune the simulation model and deep learning model based on the optimization algorithm, comprehensively analyze the training loss and verification error, continuously optimize the simulation model parameters and deep learning model hyperparameters, and achieve gradual performance improvement.
[0086] S6: Use transfer learning methods to achieve knowledge transfer from simulation environment to actual working conditions. By freezing the feature extraction layer of the pre-trained model and only fine-tuning the classification layer for actual data, the training time and data requirements are reduced. The specific steps are as follows:
[0087] S61: Load the deep convolutional neural network model pre-trained on simulation data, freeze its feature extraction layer, and retain its learned multi-operating condition feature representation capability as the basic model for actual operating condition data.
[0088] S62: According to the actual operating conditions, vibration signals and current signals are collected and used as input to fine-tune the classification layer of the pre-trained model, and the classification weights are re-optimized with less actual data.
[0089] S63: Verify the performance of the transfer learning model on the actual working condition data set. By analyzing the classification accuracy and model error, further adjust the classification layer parameters or unfreeze some high-level feature extraction layers to improve the model's adaptability and generalization ability.
[0090] The present invention proposes a transfer learning fault diagnosis method based on a digital-analog driven drum blade rehumidifier motor, which includes four stages: signal acquisition and preprocessing, mathematical modeling and simulation, deep learning model construction, and transfer learning implementation. By generating and efficiently utilizing simulation data, this method effectively solves the problem that fault data is difficult to obtain under actual working conditions. By combining multi-physics field coupling simulation with deep convolutional neural network (DCNN), the feature correlation between vibration signals and current signals is fully explored, which significantly improves the accuracy and robustness of fault diagnosis. In addition, by freezing the feature extraction layer and only fine-tuning the classification layer, the transfer learning strategy enables the model to better adapt to the complexity of actual working conditions, reduces dependence on a large amount of labeled data, and reduces the cost of model training. The overall method has significant advantages in diagnostic efficiency and adaptability to engineering applications.
[0091] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and improved concepts of the present invention within the technical scope disclosed by the present invention, and they should be covered by the protection scope of the present invention.
Claims
1. A fault diagnosis method for a motor of a drum blade rehumidifier based on digital analog drive, characterized in that: The following steps are involved: S1: Obtain key parameters and working condition signals of the drum motor, and pre-process the acquired signals; S2: Establish a mathematical model of the motor; S3: Establish a multi-physics coupling model of the drum motor through MATLAB / Simulink to generate high-precision simulation data under multiple working conditions; S4: Build a deep convolutional neural network model; S5: Use the simulation data generated by S3 as the data set for the model built in S4 for model training, and use the actual data to correct the simulation model error, and continuously improve the performance of the model through interactive iterative optimization algorithms; S6: Use transfer learning methods to achieve knowledge transfer from the simulation environment to the actual working conditions. By freezing the feature extraction layer of the pre-trained model and only fine-tuning the classification layer for the actual data, the training time and data requirements are reduced.
2. The fault diagnosis method based on digital-analog driven drum blade rehumidifier motor transfer learning according to claim 1 is characterized in that: In step S1, the steps of obtaining the key parameters and working condition signals of the drum motor and preprocessing the acquired signals are as follows: S11: Obtain actual geometric parameters of the drum motor by checking relevant drawings and instructions; S12: Obtain the electrical and mechanical parameters of the drum motor by checking the brand name and manual; S13: Obtain the actual vibration data and the actual current data of the motor during operation through sensors, and pre-process the signals.
3. The fault diagnosis method based on digital analog drive drum blade rehumidifier motor transfer learning according to claim 2 is characterized in that: The preprocessing method is CEEMDAN or FIR filter, which reduces the noise in the signal through data preprocessing.
4. The fault diagnosis method based on digital-analog driven drum blade rehumidifier motor transfer learning according to claim 1 is characterized in that: In step S2, a mathematical model of the motor is established, the rotor voltage equation is transformed by Park transformation, and the following assumptions are made: (1) the resistance in the stator and rotor windings is constant; (2) there is no saturation phenomenon in the magnetic circuit, and hysteresis and iron loss are ignored; (3) the three-phase windings are star-connected, and the spatial distribution is completely symmetrical, and the synthetic magnetomotive force is sinusoidally distributed in the air gap direction; (4) the influence of spatial harmonics is ignored; The stator flux equation is: （1） The rotor flux equation is: （2） Where: L s , L r are the equivalent self-inductance of the stator and rotor, H; L m is the equivalent mutual inductance between the stator and rotor, , ; , H; The electromagnetic torque equation is: （3） Where: p is the pole pair number; The rotor angular velocity equation is: （4） Where: T e , T m They are electromagnetic torque and load torque, N·m; n p is the rotor speed, r / min; J is the moment of inertia, kg·m².
5. The fault diagnosis method based on digital-analog driven drum blade rehumidifier motor transfer learning according to claim 1 is characterized in that: In step S3, a multi-physics field coupling model of the drum motor is established through MATLAB / Simulink to generate high-precision simulation data under multiple working conditions. The specific steps are as follows: S31: Input the key parameters of the drum motor in MATLAB / Simulink and build a modular multi-physics simulation framework based on electromagnetic, mechanical, and thermal mathematical models; S32: Use Simulink and Simscape modules to achieve coupling between electromagnetic, mechanical, and thermal fields, and set key boundary conditions to ensure accurate interaction of physical fields; S33: Run the simulation model for different working conditions, collect multi-physics field output data of vibration signals and current signals, and improve the simulation accuracy through high-precision sampling and optimization algorithms.
6. The fault diagnosis method based on digital-analog driven drum blade rehumidifier motor transfer learning according to claim 5 is characterized in that: In step S4, a network model for deep convolutional neural motor fault diagnosis is constructed, including the following four modules: S41: a sample division module, which divides the total samples of the collected vibration signals and current signals into a training set, a verification set, and a test set according to a ratio of 7:2:1; S42: a branch feature extraction module, which is structurally equivalent to extracting features in parallel with two convolution channels, and can simultaneously extract features between the vibration signal and the current signal, and save the features; S43: Feature fusion module, which uses the features through the Lambda layer Functions are merged to complete feature weighted fusion, where It can be expressed as: ， in represents the tensor to be concatenated, Indicated in Dimensional splicing, A tensor representing the output; S44: Fault diagnosis module, which uses the features fused in step S33 to complete the fault diagnosis of the motor, uses the test set to perform fault diagnosis on the trained model, inputs the data in the test set into the model, and observes the output results of the model.
7. The fault diagnosis method based on digital-analog driven drum blade rehumidifier motor transfer learning according to claim 1 is characterized in that: In step S5, the specific steps are: S51: Use the constructed data set to train the deep convolutional neural network model, and introduce actual operation data to correct the deviation of the simulation model output to improve the true reflection ability of the data set; S52: Iteratively tune the simulation model and deep learning model based on the optimization algorithm, comprehensively analyze the training loss and verification error, continuously optimize the simulation model parameters and deep learning model hyperparameters, and achieve gradual performance improvement.
8. The fault diagnosis method based on digital-analog driven drum blade rehumidifier motor transfer learning according to claim 1 is characterized in that: In step S6, the specific steps are: S61: Load the deep convolutional neural network model pre-trained on simulation data and freeze its feature extraction layer. The learned multi-operating condition feature representation capability is retained as the basic model for actual operating condition data; S62: According to the actual operating conditions, vibration signals and current signals are collected and used as inputs to fine-tune the classification layer of the pre-trained model, and the classification weights are re-optimized with less actual data; S63: Verify the performance of the transfer learning model on the actual working condition data set. By analyzing the classification accuracy and model error, further adjust the classification layer parameters or unfreeze some high-level feature extraction layers to improve the model's adaptability and generalization ability.