Numerical simulation driven motor inter-turn short circuit depth migration fault diagnosis method

Abstract

This invention discloses a numerical simulation-driven method for deep migration fault diagnosis of inter-turn short circuits in motors, belonging to the field of fault diagnosis technology. The method includes the following steps: constructing a simulation model of inter-turn short circuit faults in the motor stator; constructing a parameter migration diagnostic network and a feature migration diagnostic network; obtaining fault classification results based on the simulation model and the parameter migration diagnostic model to complete fault diagnosis; and obtaining further fault classification results based on the simulation model and the feature migration diagnostic model to complete fault diagnosis. This invention introduces a simulation model of inter-turn short circuit faults in the motor stator and utilizes the simulation data generated by the simulation model, combined with the parameter migration diagnostic network and the feature migration diagnostic network, to achieve diagnosis of the degree of inter-turn short circuit faults in motors when there are few or no actual fault samples.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, and in particular relates to a method for diagnosing inter-turn short-circuit depth migration faults in motors driven by numerical simulation. Background Technology

[0002] Rotating machinery is widely used in industrial, military, and civilian fields, and rotating components are crucial parts that directly affect the operating efficiency and working condition of the equipment. However, prolonged exposure to harsh, high-load environments makes rotating components highly susceptible to damage. This can range from reduced operating efficiency to complete shutdowns and even injuries or fatalities. Therefore, research on condition monitoring and fault diagnosis of rotating machinery is of significant theoretical and engineering value for improving production efficiency and ensuring production safety.

[0003] Deep learning, originating from artificial neural networks, aims to extract deep-dimensional feature representations of fault samples by stacking multiple layers of neural networks. Its performance in fields such as image recognition and speech recognition far surpasses other methods. Deep transfer learning is a branch of deep learning. With sufficient samples, deep learning methods achieve high accuracy in fault diagnosis, meeting application requirements. However, with limited or even missing samples, the accuracy of fault diagnosis drops significantly, necessitating appropriate methods to improve the fault diagnosis rate.

[0004] Transfer learning is an effective way to address the problem of data scarcity. It has the ability to apply knowledge learned in a source domain to a target domain, which can help improve the prediction accuracy of unlabeled data. Using a simulated sample dataset as the source domain and a real-world sample dataset as the target domain, knowledge corresponding to the degree of faults learned from the simulated sample dataset is applied to the real-world sample dataset. Transfer learning achieves fault diagnosis under conditions of missing samples by mining commonalities between different data sets.

[0005] There are two main approaches to implementing transfer learning. The first approach involves using source domain samples to achieve source domain classification and target domain samples to achieve target domain classification. It's assumed that these two mapping methods share a hidden commonality. Transferring this mapping method allows for the classification of the target domain. A representative method for this approach is parameter transfer. The second approach assumes a hidden commonality between the distributions of the source and target domain samples. A neural network is trained to project samples from both domains into the same feature space. The classification of the target domain is achieved through the commonalities between these features. A representative method is feature-based transfer. Parameter-based transfer methods train the model using source domain samples until the source domain classification task is completed. Some parameters of the trained network model are fixed, and the network parameters are fine-tuned using a small number of target domain samples. Transfer learning is divided into two steps to train the network, and the fine-tuned network is used for actual fault diagnosis tasks. Feature-based transfer methods first assume that samples from both data domains can be projected into the same high-dimensional space, where the features of samples with the same fault severity are close to each other; and that the features of different types of fault samples in the source domain dataset can be accurately classified. Transfer learning involves training a network to solve for this hidden common feature space and then directly applying it to the classification task of target domain samples.

[0006] Scholars have conducted extensive research on intelligent fault diagnosis models for rotating machinery based on deep learning. Deep learning-based diagnostic models require a large number of training samples; insufficient training samples can negatively impact fault identification accuracy. In real-world production, mechanical equipment cannot operate under faulty conditions, resulting in a limited number of fault samples. Therefore, improving the accuracy of fault diagnosis for actual motors under conditions of limited existing samples is a crucial issue. Summary of the Invention

[0007] The technical problem this invention aims to solve is as follows: To address the technical issues mentioned in the background art, this invention proposes a numerical simulation-driven method for diagnosing deep migration faults in motor inter-turn short circuits. It establishes a simulation model of inter-turn short circuit faults in the motor stator and designs parameter migration diagnostic networks and feature migration diagnostic networks. This invention introduces a simulation model of inter-turn short circuit faults in the motor stator and utilizes the simulation data generated by the model, combined with the parameter migration diagnostic network and feature migration diagnostic network, to achieve diagnosis of the degree of inter-turn short circuit faults in motors when actual fault samples are few or missing.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A numerical simulation-driven method for diagnosing inter-turn short-circuit depth migration faults in motors includes the following steps:

[0010] Step 1: Construct a simulation model of inter-turn short-circuit fault in the motor stator;

[0011] Step 2: Construct a parameter transfer diagnostic model;

[0012] Step 3: Construct a feature transfer diagnostic model;

[0013] Step 4: Based on the simulation model of the inter-turn short circuit fault of the motor stator and the parameter migration diagnosis model, obtain the fault classification results and complete the fault diagnosis;

[0014] Step 5: Based on the simulation model of the inter-turn short circuit fault of the motor stator and the feature transfer diagnosis model, obtain the fault classification result and complete the fault diagnosis.

[0015] As a further preferred embodiment of the numerical simulation-driven motor inter-turn short-circuit depth migration fault diagnosis method of the present invention, step 1, constructing a motor stator inter-turn short-circuit fault simulation model, specifically includes the following steps:

[0016] Step 1.1: Construct a normal motor simulation model in Matlab;

[0017] Step 1.2, Particle swarm optimization algorithm for motor parameter identification;

[0018] Step 1.3: Construct a fault simulation model with neutral point voltage correction;

[0019] Step 1.4: Import the motor parameters identified by particle swarm optimization into the fault simulation model with neutral point voltage correction.

[0020] As a further preferred embodiment of the numerical simulation-driven motor inter-turn short circuit depth migration fault diagnosis method of the present invention, in step 1.1, the normal motor simulation model includes the establishment of a mathematical model of a normal motor and the construction of a model in Matlab.

[0021] The mathematical model of a normal motor includes voltage equation, flux linkage equation, torque equation, and speed equation.

[0022] The voltage equation is used to solve the voltage constraint relationship of the stator winding, linking current, magnetic flux and voltage.

[0023] The flux linkage equation is the core of the motor model, used to determine the mathematical relationship between flux linkage and current through the law of electromagnetic induction;

[0024] The torque equation is used to determine the mathematical relationship between electromagnetic torque, magnetic flux, and current through Ampere's law, converting the electromagnetic field into torque for mechanical rotation.

[0025] The speed equation is used to convert torque into the angular velocity of the rotor rotation.

[0026] As a further preferred embodiment of the numerical simulation-driven fault diagnosis method for inter-turn short circuit depth migration in motors according to the present invention, the expression of the voltage equation is shown in Equation 1:

[0027]

[0028] Among them, R sa =R sb =R sc =R s R ra =R rb =R rc =R r ;

[0029] The expression for the magnetic flux linkage equation is shown in Equation 2:

[0030]

[0031] Let the mutual inductance of the coils be l. m The stator leakage inductance is l s The stator leakage inductance is l r Then the self-inductance of each stator winding coil is equivalent to The stator winding coils are spatially arranged at a phase difference of 120°, and the equivalent mutual inductance between the coils is L. sasb =L sasc =L sbsa =L sbsc =L scsa =L scsb =cos(120°)*l m =-0.5*l m The self-inductance of each phase of the rotor is equivalent to The rotors are spatially arranged at a 120° angle, and their mutual inductance is equivalent to L. rarb =L rarc =L rbra =L rbrc =L rcra =L rcrb =-0.5*l m θ is the spatial angle through which the rotor rotates, and the mutual inductance between the stator and rotor is L. sara =L sbrb =L scrc =L rasa =L rbsb =cos(θ)*l m L sarb =L sbrc =L scra =L rasc =L rbsa =L rcsb=cos(θ+120°)*l m L sarc =L sbra =L scrb =L rasb =L rbsc =L rcsa =cos(θ-120°)*l m ;

[0032] The torque equation is shown in Equation 3:

[0033]

[0034] The speed equation is shown in Equation 4:

[0035] ω r =∫[(T) e -T L )*np / J]dt (4)

[0036] Where, ω r It is the rotor rotational angular velocity, T L J is the load torque, and J is the moment of inertia.

[0037] As a further preferred embodiment of the numerical simulation-driven motor inter-turn short-circuit depth migration fault diagnosis method of the present invention, in step 1.2, the motor parameter identification using the particle swarm optimization algorithm is as follows:

[0038] The particle swarm optimization algorithm for motor parameter identification estimates the physical parameters of the motor by measuring physical quantities during operation. It uses the actual motor as a reference, the motor's state observation equation as an adjustable model, and adjusts the motor parameters of the model by using the deviation between the actual stator current and the simulated stator current.

[0039] The formulas for identifying motor parameters using the particle swarm optimization algorithm are shown in Formulas 5 and 6:

[0040] Particle position x i That is, the parameter value, velocity v i This is the direction for parameter optimization;

[0041] v i =w*v i +c1*rand()*(pbest-x i )+c2*rand()*(gbest-x i (5)

[0042] x i =v i +x i (6)

[0043] w is the inertia factor that determines the bias of particle optimization, whether it is global or local optimization; c1 and c2 are learning factors that are constants; rand() is a random number; pbest is the individual optimal solution of the particle, and gbest is the global optimal solution of the particle.

[0044] As a further preferred embodiment of the numerical simulation-driven fault diagnosis method for inter-turn short-circuit depth migration in motors according to the present invention, in step 1.3, a fault simulation model with neutral point voltage correction is constructed, as follows:

[0045] The fault simulation model for neutral point voltage correction also includes the establishment of a mathematical model and the construction of the model in Matlab.

[0046] Compared with the mathematical model of a normal motor, the mathematical model of the fault simulation model with neutral point voltage correction only requires modification to the voltage equation and the flux linkage equation.

[0047] The voltage equation is shown in Equation 7, assuming the offset voltage at the neutral point of the motor is U. k :

[0048]

[0049] Among them, R 11 =u*R s +R f R 22 = (1-u)*R s +R f U k =-1 / 3*(u*R s *(I sa2 -I sa1 )+u 2 *l s *p(I sa2 )+(u 2 -2*u)*l s *p(I sa1 ));

[0050] The magnetic flux linkage equation is shown in Equation 8:

[0051]

[0052] The self-inductance of the short-circuit coil in the faulted phase is equivalent to L. sa2sa2 =u 2 *(l m +l s The self-inductance of the remaining coil in the faulty phase is L. sa2sa1 =L sa1sa2 =u*(1-u)*l m The mutual inductance between the faulty phase winding coil and other phase windings is equivalent to L. sa2sb =Lsa2sc =L sbsa2 =L scsa2 =-0.5*u*l m L sa2ra =L rasa2 =u*cos(θ)*l m L sa2rb =L rbsa2 =u*cos(θ+120°)*l m L sa2rc =L rcsa2 =u*cos(θ-120°)*l m L salsa1 =(1-u) 2 *(l m +l s ), L sa1sb =L sa1sc =L sbsa1 =L scsa1 = -0.5*(1-u)*l m L sa1ra =L rasa1 = (1-u)*cos(θ)*l m L sa1rb =L rbsa1 = (1-u)*cos(θ+120°)*l m L sa1rc =L rcsa1 = (1-u)*cos(θ-120°)*l m Other inductance parameters are the same as those of a normal motor model.

[0053] As a further preferred embodiment of the numerical simulation-driven deep migration fault diagnosis method for inter-turn short circuits in motors according to the present invention, in step 2, the parameter migration diagnosis model includes a deep feature extraction network and an equipment state classification network.

[0054] The deep feature extraction network is used to extract the depth features of the simulated fault current samples.

[0055] The equipment status classification network is used to classify the degree of inter-turn short-circuit faults in motors;

[0056] The deep feature extraction network is a shallow ResNet structure. The ResNet structure has one convolutional layer with a kernel size of 3×3, a stride of 2, padding of 1, eight residual blocks, and one fully connected layer.

[0057] The device status classification network is a two-layer linear fully connected network.

[0058] As a further preferred embodiment of the numerical simulation-driven deep migration fault diagnosis method for inter-turn short circuits in motors according to the present invention, in step 3, the feature migration diagnosis model includes a feature extraction network, a distance metric network, and a classification network.

[0059] The feature extraction network is used to extract deep features from simulated fault samples and actual fault samples;

[0060] The distance metric network is used to reduce the difference between the depth features of simulated fault samples and the depth features of actual fault samples;

[0061] The classification network is used to simulate the classification of deep features of fault samples;

[0062] The feature extraction network uses one convolutional layer and eight residual blocks.

[0063] The distance metric network selects the maximum mean deviation;

[0064] The classification network is a two-layer fully connected network.

[0065] As a further preferred embodiment of the numerical simulation-driven deep migration fault diagnosis method for inter-turn short circuits in motors according to the present invention, the parameter migration diagnosis model further includes a classification loss function for backpropagation optimization of the deep feature extraction network.

[0066] The classification loss function is used for backpropagation to optimize the deep feature extraction network;

[0067] The expression for the classification loss function is:

[0068]

[0069] Where L is the classification loss and P is the true label of the sample. Predict labels for samples.

[0070] As a further preferred embodiment of the numerical simulation-driven motor inter-turn short circuit depth migration fault diagnosis method of the present invention, in step 3, the feature migration diagnosis model further includes two loss functions for backpropagation optimization of the network model.

[0071] The two loss functions are: device status classification loss and depth feature distribution difference loss between source domain and target domain data, respectively.

[0072] The device state classification loss is used for backpropagation to optimize the deep feature extraction network and the device state classification network;

[0073] The expression for the device state classification loss is:

[0074]

[0075] Among them, L v The loss is defined as the equipment status classification loss, where P is the true label of the sample. Predict labels for the samples;

[0076] The deep feature distribution difference loss is used to minimize the difference between source domain features and target domain features;

[0077] The expression for the depth feature distribution difference loss between the source and target domains is:

[0078]

[0079] Among them, L mmd The loss is the distribution difference between the source domain and the target domain. sup is the maximum upper bound. There are m source domain samples and n target domain samples. g(x) is the mapping function.

[0080] The expression for the total loss function of the feature transfer diagnostic network is:

[0081] L lossall =L y +λ1L mmd

[0082] Among them, L lossall Let λ1 be the total loss function, and L be the loss function. mmd The weight.

[0083] Compared with the prior art, the present invention has the following advantages and technical effects:

[0084] 1. This invention proposes a numerical simulation-driven method for fault diagnosis of inter-turn short circuit depth migration in motors, enabling fault diagnosis of motors under small sample conditions;

[0085] 2. This invention introduces a simulation model for inter-turn short-circuit faults in motor stators. By utilizing the simulation data generated by the simulation model and combining it with parameter transfer diagnostic networks and feature transfer diagnostic networks, the degree of inter-turn short-circuit faults in motors can be diagnosed when there are few or no actual fault samples.

[0086] 3. This invention designs a parameter transfer diagnostic network, uses source domain samples for model training, fixes some parameters of the model, uses a small number of labeled samples from the target domain to fine-tune the rest of the network, and uses the fine-tuned network for actual fault diagnosis tasks.

[0087] 4. This invention designs a feature transfer diagnostic network, which uses a network trained in the source domain to predict the target domain. It also introduces MMD to reduce the probability distribution difference of deep features between the source and target domains. Attached Figure Description

[0088] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0089] Figure 1 This is a schematic diagram of the fault diagnosis method for inter-turn short circuit depth migration of motor driven by numerical simulation according to an embodiment of the present invention.

[0090] Figure 2 This is a block diagram of the stator inter-turn short-circuit fault modeling structure in an embodiment of this civilization;

[0091] Figure 3 This is a Matlab simulation diagram of a normal motor according to an embodiment of the present invention;

[0092] Figure 4 The simulation results of current, torque, speed, etc., of a normal motor simulation model according to an embodiment of the present invention are shown in the figure.

[0093] Figure 5 This is a schematic diagram of the particle swarm identification motor parameter process according to an embodiment of the present invention;

[0094] Figure 6 This is a simulation diagram of an inter-turn short-circuit fault in a motor according to an embodiment of the present invention;

[0095] Figure 7 This is a schematic diagram of a network framework based on parameter migration according to an embodiment of the present invention;

[0096] Figure 8 This is a schematic diagram of a feature transfer-based network framework according to an embodiment of the present invention;

[0097] Figure 9 This is a motor integrated test bench according to an embodiment of the present invention;

[0098] Figure 10 This is a schematic diagram of the actual current signal of the motor in an embodiment of the present invention;

[0099] Figure 11 This is a schematic diagram comparing the simulated and actual current of the F4 type of fault in this embodiment of the invention.

[0100] Figure 12 This is a schematic diagram showing the difference between the simulated and actual current for the fourth, fifth, and sixth types of faults in this embodiment of the invention;

[0101] Figure 13 This is a schematic diagram of the source domain training loss function and target domain test accuracy of the parameter transfer diagnostic network according to an embodiment of the present invention;

[0102] Figure 14 This is a schematic diagram of the target domain fine-tuning loss function and test accuracy of the parameter migration diagnostic network according to an embodiment of the present invention;

[0103] Figure 15 This is a schematic diagram illustrating the test accuracy of a network fine-tuning based on a parameter transfer diagnostic network with different numbers of target domain samples according to an embodiment of the present invention.

[0104] Figure 16 This is a schematic diagram of the overall loss function and MMD loss function for training the feature transfer diagnostic network according to an embodiment of the present invention;

[0105] Figure 17 This is a schematic diagram illustrating the source domain classification accuracy and target domain classification accuracy based on the feature transfer diagnostic network in an embodiment of the present invention.

[0106] Figure 18 This is a schematic diagram of the confusion matrix of the target domain diagnosis results based on the feature transfer diagnostic network in an embodiment of the present invention;

[0107] Figure 19 This is a schematic diagram showing a visual comparison of source and target domain features based on a feature transfer diagnostic network according to an embodiment of the present invention.

[0108] Figure 20 This is a schematic diagram comparing the test accuracy of different numbers of fault samples in an embodiment of the present invention;

[0109] Figure 21 This is a schematic diagram comparing the test accuracy at different rotational speeds according to an embodiment of the present invention. Detailed Implementation

[0110] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0111] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0112] Example:

[0113] like Figure 1 As shown, this embodiment provides a numerical simulation-driven method for diagnosing inter-turn short-circuit depth migration faults in motors, including the following steps:

[0114] S1. Based on different degrees of inter-turn short circuit, a simulation model of inter-turn short circuit fault in motor stator is constructed to obtain simulation data of inter-turn short circuit fault;

[0115] The stator inter-turn short-circuit fault modeling structure block diagram is as follows: Figure 2As shown, it includes normal motor modeling, particle swarm parameter identification, and inter-turn short-circuit fault model.

[0116] Furthermore, a normal simulation model of the motor is established. By simultaneously solving the motor's voltage equation, flux linkage equation, torque equation, and speed equation, the entire process of the motor consuming electrical energy from the application of voltage to the rotor rotating and gaining mechanical energy can be expressed. Using the four equations mentioned above, a simulation model of the normal motor is constructed in Matlab. The simulation model of the normal motor is as follows: Figure 3 As shown, the simulation results are as follows: Figure 4 As shown;

[0117] Furthermore, the Particle Swarm Optimization (PSO) algorithm is used to identify motor parameters. To accurately build a simulation model of the motor, it is essential to obtain its physical parameters. The main unknown physical parameters in the motor model include: stator resistance R. s Rotor resistance R r Mutual inductance of coils m stator leakage inductance s Rotor leakage inductance l r , moment of inertia J. Under laboratory conditions, R s J is more convenient to measure directly, R r l m l s l r The parameters can be indirectly measured through no-load and stall tests, but the accuracy of the measured parameters is relatively low.

[0118] A schematic diagram of the particle swarm identification motor parameter process is shown below. Figure 5 As shown, it includes the initialization part and the update optimization part.

[0119] The initialization part refers to randomly generating particles within a limited space to calculate fitness. The more random particles there are, the higher the probability of algorithm convergence and the more accurate the identified parameters. However, too many particles can also lead to excessive computational cost in the update and optimization iteration. In this case, a manual selection part is added to reduce the computational cost.

[0120] The update optimization process moves all particles closer to the particle with the lowest fitness, which is called the optimal particle. The higher the iteration number, the closer the optimal particle's solution is to the expected value. However, due to computational constraints, the optimal particle's solution is only an approximate solution.

[0121] In this embodiment, the motor parameters after particle swarm optimization are as follows: stator winding resistance is 7.15Ω, rotor winding resistance is 5.10Ω, stator winding leakage inductance is 0.0007H, rotor winding leakage inductance is 0.0003H, winding mutual inductance is 0.1439H, and moment of inertia is 0.0027Kg / m. 2 .

[0122] Furthermore, the inter-turn short-circuit fault model is established. The established inter-turn short-circuit fault model is as follows: Figure 6 As shown, it is mainly divided into four parts. The first part, represented by the red dashed line, represents the neutral point voltage correction equation. After an inter-turn short-circuit fault occurs, the neutral point of the star winding will shift. The second part, represented by the yellow dashed line, represents the seven voltage equations. The neutral point voltage deviation only affects the voltage equations of the three-phase stator windings and does not affect the voltage equations of the faulty windings or the rotor voltage equations. The third part, represented by the blue dashed line, represents the flux linkage calculation matrix. The fourth part, represented by the brown line, is the rotational speed equation.

[0123] S2. Construct a diagnostic model based on parameter transfer;

[0124] S3. Construct a diagnostic model based on feature transfer.

[0125] S4. Based on the simulation model of inter-turn short circuit fault in the motor stator and the parameter migration diagnosis model, obtain the fault classification results and complete the fault diagnosis.

[0126] The structure of the parameter transfer diagnostic model is as follows: Figure 7 As shown, it includes: a deep feature extraction network, a fine-tuning network, and a device status classification network;

[0127] The basic workflow of a parameter transfer diagnostic network is as follows:

[0128] (1) Collect signals from rotating machinery under different operating conditions and divide the data into a source domain and a target domain. Perform linear normalization on the source domain and target domain datasets. In this embodiment of the invention, the experimental source domain training dataset is data simulating various fault degrees. The target domain fine-tuning dataset is selected from actual data on various fault degrees, and the target domain test dataset is also selected from actual data on various fault degrees.

[0129] (2) Model training: The network is trained using source domain samples to ensure that it can accurately classify the source domain targets. The shallow network uses the ResNet network structure, and the deep network uses linear connection layers.

[0130] (3) The overall network loss function can be the cross-entropy loss function L, and the loss function is calculated.

[0131]

[0132] Where L is the equipment status classification loss, and P is the true label of the sample. Predict labels for samples.

[0133] (4) Iterate through steps (2)-(3) until the loss function is less than the set value or the number of iterations reaches the set requirement, and the trained network is obtained.

[0134] (3) Network fine-tuning: The parameters of the trained shallow network are fixed, and the deep network is fine-tuned by using 1, 5, 10, 20 and 30 labeled samples of each target domain.

[0135] (4) Testing: Compare the required number of labeled samples in the target domain and test the network performance using the target domain test set respectively.

[0136] S5. Based on the simulation model of inter-turn short circuit fault in the motor stator and the feature transfer diagnosis model, obtain the fault classification results and complete the fault diagnosis.

[0137] The structure of the feature transfer-based diagnostic model is as follows: Figure 8 As shown, it includes: a feature extraction network, a distance metric network, and a classification network;

[0138] The basic workflow of a feature transfer diagnostic network is as follows:

[0139] (1) Collect signals from rotating machinery under different operating conditions and divide the data into a source domain and a target domain. Perform linear normalization on the source domain and target domain datasets. In this embodiment of the invention, the experimental source domain training dataset is current data simulating various fault degrees. The target domain training dataset is selected from actual fault current data of various types. The target domain test dataset is selected from actual fault current data of various types.

[0140] (2) Data Loading: Due to the severe imbalance between the samples in the simulation dataset and the actual dataset, a data loading method needs to be set. In this embodiment of the invention, the source domain simulation dataset has 100 samples in each of 7 classes, for a total of 700 training samples. The target domain actual dataset has 1, 5, 10, 20, and 30 samples for each fault severity class to gradually test the impact of the number of samples on the diagnostic accuracy.

[0141] (3) Feature Extraction Network: A ResNet-based feature extraction network was constructed. Feeding samples into the feature extraction network maps the samples to a completely new dimensional space, obtaining deep features representing the samples. The feature extraction network was trained using the source domain training set and the target domain training dataset.

[0142] (4) Distance metric: MMD distance is selected to measure the difference between the depth features of the simulated dataset and the depth features of the actual dataset. The calculation formula is as follows:

[0143]

[0144] In the formula, sup represents finding the maximum upper bound. This is achieved by solving a function g() across all function sets G, such that the features of the source domain and the target domain have maximum values. Due to the imbalance between the source and target domain samples, the distribution of faulty samples in the source domain in the feature space can be calculated for each sample, and the distribution of normal samples in the target domain can also be calculated. However, the distribution of faulty samples in the target domain can only be represented by the mean of one or a few samples. Each time the network returns for optimization, the mean of the target domain features needs to be recalculated.

[0145] (5) Classification Network: The deep features of the source domain are classified using a linear mapping relationship, mapping the 256×1 dimension features to 7×1 to obtain the predicted class probability of the samples. The label with the highest probability is selected as the prediction result. The cross-entropy function is used to calculate the difference between the predicted result and the actual label as the classification loss function. The overall loss function of the network is L... lossall The calculation formula is as follows:

[0146] L lossall =L y +λ1L mmd

[0147] In the formula L y It is the classification loss of the source domain, L mmd λ is the MMD distance between the source domain features and the target domain features. λ is a constant, and different values ​​of λ will directly affect the quality of network training.

[0148] Furthermore, in this embodiment, the fault diagnosis and analysis method based on data-driven parameter migration diagnostic network and feature migration diagnostic network proposed in this invention is also used to perform fault analysis on the motor integrated test bench.

[0149] 1. Motor Comprehensive Test Bench

[0150] All experiments were conducted on a comprehensive motor test bench, such as... Figure 9 As shown. The test motor is a Marathon three-phase induction motor. The current signal passes through a current transformer and is then connected to an NI current acquisition card. Fault settings: Two short-circuit fault settings are configured by connecting terminals with turns ratios of 0.05 and 0.1 on the stator windings of the motor. Simultaneously, an external sliding rheostat is connected to simulate the short-circuit resistance of the motor, with resistance values ​​of 0, 0.5, and 1 Ω.

[0151] Different scenarios are simulated by setting the short-circuit turns ratio and short-circuit resistance, as shown in Table 1. Table 1 shows seven types of fault states. In the table, N represents the current of the normal motor, and F1 to F6 are the fault currents under different combinations of short-circuit turns and short-circuit resistance. By changing the short-circuit turns ratio and the short-circuit resistance, different degrees of inter-turn short circuits are simulated. The larger the inter-turn short circuit turns ratio, the smaller the circuit resistance, and the higher the degree of fault.

[0152] Table 1

[0153]

[0154]

[0155] Experimental setup: The motor was set to open-loop control, with speeds of 30, 35, and 40 r / s. Circuit data for seven states were collected at each speed, as shown in Table 1. Data was collected three times for each state, with each collection lasting 100 seconds and a sampling frequency of 16 kHz.

[0156] The actual rotational speeds under various conditions are shown in Table 2. The experimentally measured speeds show a significant decrease after a fault occurs; the higher the degree of the fault, the more pronounced the decrease. Under the same load torque, when an inter-turn short-circuit fault occurs in the motor, the effective value of the fundamental magnetomotive force decreases. To generate sufficient electromagnetic torque, the rotor speed will inevitably decrease slightly.

[0157] Table 2

[0158]

[0159] Dataset setup: Separate training and test sets were set up for each rotational speed. 100 samples of current were collected for each state, totaling 700 samples across all states. Each sample was 3×4096 pixels long. Actual current signal of the motor. Figure 10 As shown. Due to the disassembly and reassembly of the motor windings, the waveform of the actual normal current is not a strictly symmetrical three-phase current. There is a slight difference in the peak values ​​between the three-phase currents, and each phase current has a slight DC component. Affected by the load and frequency converter, the actual current peak value will show slight periodic fluctuations.

[0160] 2. Simulation Model Analysis

[0161] After determining the parameters of a normal motor, simulation samples are generated by setting the short-circuit turns ratio and short-circuit resistance values. The values ​​of u are 0.05 and 0.1, and R... f The values ​​are 0 ohms, 0.5 ohms, and 1 ohm, for a total of 6 fault states, corresponding to actual fault states. Real-world scenarios are simulated by adding random disturbances of 0 to 0.05 ohms to the load torque.

[0162] Actual and simulation samples of F4 type fault severity, such as Figure 11 As shown in the figure, the peak current values ​​during the fault vary considerably, but the overall trend of the fault current is the same. When a fault occurs in phase C, the current amplitude in phase C is the largest, the current in phase A also increases, while the current in phase B remains essentially unchanged. The phase change trends are also the same. During a fault, the phase difference between A and B in the actual sample is 87.1°, and the phase difference between A and C is -145.1°. In the simulation sample, the phase difference between A and B is 101.6°, and the phase difference between A and C is -142.3°.

[0163] The differences in phase current values ​​for fault types F4, F5, and F6 are as follows: Figure 12 As shown in the figure, the simulated fault current and the actual fault current exhibit the same trend. Especially at lower fault severity, the similarity between the two currents is very high. The higher the fault severity, the larger the fault phase current value and the greater the phase imbalance. The greater the fault severity, the greater the difference.

[0164] The sample discrepancies for all types of faults are shown in Table 3. The discrepancy calculation method involves selecting samples of each type of fault, calculating the distance between each point of the simulated sample and the actual sample, and then calculating the average value of each point. The table shows that the distance between the simulated sample and the actual sample is related to the degree of fault; the higher the degree of fault, the greater the average deviation per point.

[0165] Table 3

[0166]

[0167] 3. Analysis of Experimental Results Based on Simulation Model and Parameter Transfer Diagnosis Model of Inter-turn Short Circuit Fault in Motor Stator

[0168] The method of constructing a parameter transfer diagnostic model in S2 is adopted, and the simulation dataset and the actual fault dataset are combined to complete the diagnosis of the degree of inter-turn short circuit fault in the motor.

[0169] (1) Train the network using the source domain training data and test it directly using the target domain dataset. Experiment

[0170] The results are as follows Figure 13 As shown.

[0171] As shown in the figure, training the network using the source domain training dataset results in rapid convergence, with the loss function converging after 20 training epochs, because the network at this point only uses simulated samples for training. If the source domain samples and target domain samples have no similarity in distribution, the classification accuracy should be around 1 / 7. In this case, directly testing the network using the actual target domain dataset yields a classification accuracy of 45.7143%.

[0172] (2) The network was trained using the target domain dataset, with 30 actual samples per class, and tested using the target domain test dataset. The experimental results are as follows: Figure 14 As shown.

[0173] As shown in the figure, the network achieved a test accuracy of 75% after the first round of training, far exceeding the probability of 1 / 7. After 30 rounds of fine-tuning, the loss function of the fine-tuned network converged. At this point, the test accuracy stabilized at around 97.8571%, which can accurately classify almost all faults. Compared to training the network with only 30 actual fault samples, the fault diagnosis accuracy based on parameter transfer improved by about 50%.

[0174] (3) Comparative analysis of the number of samples of different actual faults:

[0175] Using 30 real-world samples for each type of fault to fine-tune the network to achieve an accuracy of 97.8571%, which meets the needs of practical applications. However, the required number of real-world fault samples is also relatively large. The test results for 1, 5, 10, 20, and 30 real-world samples per type are as follows: Figure 15 As shown, the accuracy of the test significantly improved with the increase in the number of actual fine-tuning samples. The final test accuracies were 55.7143%, 75.8571%, 85.4286%, 94.2857%, and 97.8571%, respectively. The number of actual fault samples required to achieve the actual fault diagnosis task through parameter transfer is approximately 20.

[0176] Parameter-based transfer learning is a fundamental method of transfer learning. Its network structure and loss function are relatively simple, and training is relatively easy. However, due to its unique training method, problems may arise.

[0177] 1) Parameter transfer assumes that the previous network is used to extract features, and the parameters are directly used in the test network. There is no explicit formula to prove that the previous network is used for feature extraction, and since the source and target domains extract features in the same way, the accuracy of the transferred network may be low. The accuracy of the 10 samples in the figure is typical. Using actual samples to fine-tune the network, theoretically, it should improve steadily and gradually, especially after the network stabilizes, without significant fluctuations.

[0178] 2) Parameter-based transfer learning networks require a small number of labeled samples from the target domain for fine-tuning, but the exact number of labeled samples is not specified. In this experiment, two methods were used to optimize the fault simulation model, improving the similarity between simulated and actual samples, resulting in an accuracy of approximately 45% even without actual samples. If the difference between the source and actual samples is significant, fine-tuning the network using only 1 or 5 fault samples will lead to even lower accuracy. Although parameter-based transfer learning fault diagnosis methods still have many shortcomings, they perform well in practical applications, maintaining high test accuracy even with a limited number of samples.

[0179] 4. Analysis of Experimental Results Based on Simulation Model and Feature Transfer Diagnostic Model of Inter-turn Short Circuit Fault in Motor Stator

[0180] The method of constructing a feature transfer-based diagnostic model in S3 is adopted, and the simulation dataset and the actual fault dataset are combined to complete the diagnosis of the degree of inter-turn short circuit fault in the motor.

[0181] Let's take the results of training the network with 30 real samples per class at 40 r / s as an example. Figure 16 The diagram shows the loss function for training the network, as follows: Figure 17 The figure shows the classification accuracy of the trained network. The confusion matrix of the target domain diagnosis results is shown below. Figure 18 As shown.

[0182] Visualization of source and target domain features, for example Figure 19 As shown, each color represents a feature of a type of fault sample. Figure a) shows the fault features of the source domain dataset extracted by the neural network, and Figure b) shows the fault features of the target domain dataset extracted by the same neural network. The distribution of features for the same fault type samples in the source and target domains is quite similar, but the features in the source domain are more compact and the classification is more accurate. Fault diagnosis methods based on feature transfer reduce the feature differences between the source and target domains. The results of feature visualization indicate to some extent that the feature transfer-based method can effectively reduce the differences in feature space between samples from two different data domains.

[0183] 5. Comparative Analysis of Parameter-Based and Feature-Based Transfer Learning Diagnostic Networks: Two transfer learning methods were tested at rotational speeds of 30, 35, and 40 r / s. The experimental results for fault diagnosis, and the test accuracy after training the network with different target domain samples, are shown in Table 4.

[0184] Table 4

[0185]

[0186] (1) The impact of different numbers of fault samples

[0187] At a rotational speed of 40 r / s, the results of the two transfer learning algorithms are compared as follows: Figure 20 As shown in the figure, with the increase in the number of actual fault samples, the more information is extracted from the actual samples, and the higher the diagnostic accuracy of both transfer algorithms becomes. However, the result of parameter transfer is slightly better than that of feature transfer by 5%. Parameter transfer requires fine-tuning the network using a small number of labeled actual fault samples. These labeled actual samples are used not only in the forward training of the network but also to calculate the classification loss of the actual samples for the backward optimization of the network. Feature transfer requires a small number of actual fault samples in the forward training of the network to estimate the distribution of the actual samples. It does not require labels or to calculate the classification loss of the actual samples. As can be seen from the figure, when the number of actual fault samples exceeds 10, both transfer learning methods perform well, with accuracy rates above 80%.

[0188] (2) Effect of different rotation speeds

[0189] At a rotational speed of 40 r / s, the results of the two transfer learning algorithms are compared as follows: Figure 21 As shown, both transfer learning methods performed well at the three tested speeds, with accuracy rates generally above 80%. Experimental results at different speeds demonstrate the good applicability of the method. The simulated fault dataset contains some knowledge from actual fault data. Transfer learning can be used to solve the fault diagnosis of inter-turn short circuits in motors under conditions of limited fault sample size.

[0190] (3) Comparison of computational complexity

[0191] The parameter transfer-based method requires a two-step experiment: first, a source domain classification network is trained using simulated data; then, real-world data is used to fine-tune the network. Each network converges after approximately 50 iterations, with each iteration taking about 10 seconds. The feature transfer-based method trains only one network, and each iteration requires calculating the distance between simulated and real-world data in the feature space, with each iteration taking about 60 seconds.

[0192] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A numerical simulation-driven method for diagnosing inter-turn short-circuit depth migration faults in motors, characterized in that, Specifically, the following steps are included: Step 1: Construct a simulation model of inter-turn short-circuit fault in the motor stator; Step 2: Construct a parameter transfer diagnostic model; Step 3: Construct a feature transfer diagnostic model; Step 4: Based on the simulation model of the inter-turn short circuit fault of the motor stator and the parameter migration diagnosis model, obtain the fault classification results and complete the fault diagnosis; Step 5: Based on the simulation model of the inter-turn short circuit fault of the motor stator and the feature transfer diagnosis model, obtain the fault classification result and complete the fault diagnosis; In step 1, a simulation model of inter-turn short-circuit fault in the motor stator is constructed, which specifically includes the following steps: Step 1.1: Construct a normal motor simulation model in Matlab; Step 1.2, Particle swarm optimization algorithm for motor parameter identification; Step 1.3: Construct a fault simulation model with neutral point voltage correction; Step 1.4: Import the motor parameters identified by particle swarm optimization into the fault simulation model with neutral point voltage correction. The expression for the voltage equation is shown in Equation 1: （1） in, , ; The expression for the flux linkage equation is shown in Equation 2: （2） Let the mutual inductance of the coils be stator leakage inductance stator leakage inductance Then the self-inductance of each stator winding coil is equivalent to The stator winding coils are spatially arranged at a phase difference of 120°, and the mutual inductance between the coils is equivalent to... The self-inductance of each phase of the rotor is equivalent to The rotors are spatially arranged at a 120° angle, and their mutual inductance is equivalent to... , It is the spatial angle through which the rotor rotates, and the mutual inductance between the stator and rotor is... ; The torque equation is shown in Equation 3: （3） The speed equation is shown in Equation 4: （4） in, It is the rotor's angular velocity. It is the load torque. It is the moment of inertia; In step 2, the parameter transfer diagnostic model includes a deep feature extraction network and a device state classification network; The deep feature extraction network is used to extract the depth features of the simulated fault current samples. The equipment status classification network is used to classify the degree of inter-turn short-circuit faults in motors; The deep feature extraction network is a shallow ResNet structure. The ResNet structure has one convolutional layer with a kernel size of 3×3, a stride of 2, padding of 1, eight residual blocks, and one fully connected layer. The device status classification network is a two-layer linear fully connected network. In step 3, the feature transfer diagnostic model includes a feature extraction network, a distance metric network, and a classification network; The feature extraction network is used to extract deep features from simulated fault samples and actual fault samples; The distance metric network is used to reduce the difference between the depth features of simulated fault samples and the depth features of actual fault samples; The classification network is used to simulate the classification of deep features of fault samples; The feature extraction network uses one convolutional layer and eight residual blocks. The distance metric network selects the maximum mean deviation; The classification network is a two-layer fully connected network.

2. The numerical simulation-driven method for diagnosing inter-turn short-circuit depth migration faults in motors according to claim 1, characterized in that, In step 1.1, the normal motor simulation model includes the establishment of a mathematical model of the normal motor and the construction of the model in Matlab; The mathematical model of a normal motor includes voltage equation, flux linkage equation, torque equation, and speed equation. The voltage equation is used to solve the voltage constraint relationship of the stator winding, linking current, magnetic flux and voltage. The flux linkage equation is the core of the motor model, used to determine the mathematical relationship between flux linkage and current through the law of electromagnetic induction; The torque equation is used to determine the mathematical relationship between electromagnetic torque, magnetic flux, and current through Ampere's law, converting the electromagnetic field into torque for mechanical rotation. The speed equation is used to convert torque into the angular velocity of the rotor rotation.

3. The numerical simulation-driven method for diagnosing inter-turn short-circuit depth migration faults in motors according to claim 1, characterized in that, In step 1.2, the particle swarm optimization algorithm motor parameter identification is as follows: The particle swarm optimization algorithm for motor parameter identification estimates the physical parameters of the motor by measuring physical quantities during operation. It uses the actual motor as a reference, the motor's state observation equation as an adjustable model, and adjusts the motor parameters of the model by using the deviation between the actual stator current and the simulated stator current. The formulas for identifying motor parameters using the particle swarm optimization algorithm are shown in Formulas 5 and 6: Particle position That is, parameter value, speed This is the direction for parameter optimization; (5) (6) The inertia factor determines the bias of particle optimization, whether it is global or local optimization; and The learning factor is a constant; It is a random number; It is the optimal solution for an individual particle. It is the global optimal solution for the particle.

4. The numerical simulation-driven method for diagnosing inter-turn short-circuit depth migration faults in motors according to claim 1, characterized in that, In step 1.3, a fault simulation model with neutral point voltage correction is constructed, as follows: The fault simulation model for neutral point voltage correction also includes the establishment of a mathematical model and the construction of the model in Matlab. Compared with the mathematical model of a normal motor, the mathematical model of the fault simulation model with neutral point voltage correction only requires modification to the voltage equation and the flux linkage equation. The voltage equation is shown in Equation 7, assuming the offset voltage at the neutral point of the motor is... : （7） in, , , ; The magnetic flux linkage equation is shown in Equation 8: （8） The self-inductance of the short-circuit coil in the faulted phase is equivalent to: The self-inductance of the remaining coil in the faulty phase is The mutual inductance between the faulty phase winding coil and other phase windings is equivalent to , , , , , , , , Other inductance parameters are the same as those of a normal motor model.

5. The numerical simulation-driven method for diagnosing inter-turn short-circuit depth migration faults in motors according to claim 1, characterized in that, The parameter transfer diagnostic model also includes a classification loss function for backpropagation optimization of the deep feature extraction network; The classification loss function is used for backpropagation to optimize the deep feature extraction network; The expression for the classification loss function is: Where L is the classification loss and P is the true label of the sample. Predict labels for samples.

6. The numerical simulation-driven method for diagnosing inter-turn short-circuit depth migration faults in motors according to claim 1, characterized in that, In step 3, the feature transfer diagnostic model also includes two loss functions for backpropagation optimization of the network model; The two loss functions are: device status classification loss and depth feature distribution difference loss between source domain and target domain data, respectively. The device state classification loss is used for backpropagation to optimize the deep feature extraction network and the device state classification network; The expression for the device state classification loss is: ;in, The loss is defined as the equipment status classification loss, where P is the true label of the sample. Predict labels for the samples; The deep feature distribution difference loss is used to minimize the difference between source domain features and target domain features; The expression for the depth feature distribution difference loss between the source and target domains is: ;in, The sum is the loss function for the distribution difference between the source and target domains. It finds the maximum upper bound, given m source domain samples and n target domain samples. For mapping functions; The expression for the total loss function of the feature transfer diagnostic network is: ;in, For the total loss function, for The weight.

Images