An intelligent diagnosis method for motor faults based on digital twins

Abstract

The present invention discloses an intelligent motor fault diagnosis method based on digital twins, belonging to the field of motor fault diagnosis. In the model training stage, the method selects feature quantities through the random forest method to train an extreme learning machine model with an autoregressive cycle function as a digital twin motor. In the open circuit fault diagnosis stage, the method collects the physical entity motor feature quantity and inputs it into the digital twin virtual motor to obtain the digital twin current, and makes a difference between the actual winding current data and the digital twin winding current to obtain the current residual, and uses the difference value of the current residual as a threshold to perform open circuit fault diagnosis, and adopts the sliding average window method to locate the fault; in the turn-to-turn short circuit fault diagnosis stage, the current symmetrical mean absolute error is used as a judgment standard to diagnose the short circuit fault. The present invention does not require precise design parameters and motor fault data of the motor, but only requires data when the motor is in normal operation to quickly and real-time diagnose the motor open circuit fault and turn-to-turn short circuit fault, and has strong robustness.

Description

Technical Field

[0001] The present invention relates to the field of motor fault diagnosis, and in particular to a motor fault intelligent diagnosis method based on digital twins. Background Art

[0002] The faults of motor drive systems mainly include open circuit faults and turn-to-turn short circuit faults. Open circuit faults are usually caused by winding faults or power tube faults, while turn-to-turn short circuit faults manifest as local short circuits in the motor windings, which may cause serious degradation of motor performance or even failure. For these faults, current diagnostic methods are mainly divided into three categories: signal processing-based methods, analytical model-based methods, and data-driven methods. However, these methods still have their own shortcomings in practical applications.

[0003] Signal processing-based methods rely on extracting fault features from complex signals, and have low diagnostic accuracy in high-noise environments. In addition, due to the large amount of calculations, real-time performance is limited. Analytical model-based methods require an accurate mathematical model of the motor, but under actual working conditions, the mathematical model of the motor and the actual physical model parameters often have irreconcilable errors. Therefore, analytical model-based methods have poor adaptability to parameter changes and complex systems, and are difficult to apply to multi-parameter fault conditions. In addition, the model building process is complex and relies on a high level of professional knowledge. Existing data-driven methods often require the collection of a large amount of motor fault data to establish a diagnostic database, such as a dual three-phase permanent magnet motor open circuit and sensor fault diagnosis method based on a digital twin model (patent number CN118519029A, publication date 2024-08-20). It constructs a digital twin motor model in a healthy state through a neural network, and constructs a diagnostic database for diagnosis by collecting stator current residuals under various fault conditions of the digital twin motor and the physical motor. The disadvantage is that this method relies on a large amount of different types of motor fault data for model training. These fault data are often difficult to obtain and the accuracy of the data is difficult to guarantee. In addition, it is difficult to obtain high-precision diagnostic effects when the data is insufficient, and the generalization ability for different working conditions is limited. The diagnostic speed of this diagnostic method is not reflected, and the diagnostic speed is difficult to guarantee, and it is impossible to diagnose open circuit faults and turn-to-turn short circuit faults at the same time.

[0004] In addition, complex data-driven algorithms also require high computing resources. Therefore, the existing fault diagnosis methods need to be further improved in terms of accuracy, adaptability, and real-time performance to meet the diverse fault diagnosis needs of motor systems. Summary of the invention

[0005] The purpose of the present invention is to overcome the above-mentioned defects existing in the background technology and provide an intelligent motor fault diagnosis method based on digital twins, so that it does not require the motor parameters at the time of failure, does not need to add additional hardware circuits, has a simple diagnostic algorithm, a fast diagnostic speed, and a strong ability to prevent misdiagnosis. It can quickly and accurately diagnose the open circuit fault of the power tube, the open circuit fault of the motor winding and the inter-turn short circuit fault of the motor drive system.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] A motor fault intelligent diagnosis method based on digital twins, the method comprises a model training stage and a fault diagnosis stage, and specifically comprises the following steps:

[0008] Step S1: In the model training stage, synchronous sampling technology is used to collect various timing motor parameters when the motor is operating normally under different working conditions to build a database;

[0009] Step S2: In the model training phase, the random forest model is trained to sort the importance of all motor parameters, and the electrical angle at the same time in the database is selected according to the importance. , torque and phase voltage , and add the physical entity motor winding current at the previous moment , where n represents each phase of the motor, which is used as the input for training the R-ELM model; the R-ELM model is an improved extreme learning machine model with autoregressive cycle function; the R-ELM model extracts features by the extreme learning machine, and then performs time series modeling and predicts output through a recurrent neural network structure;

[0010] Step S3: In the model training phase, online training is adopted to use the electrical angle of the normal operation in the diagnostic database. , torque and phase voltage , and add the motor winding current at the previous moment As the training set input, the winding current data corresponding to the feature quantity in time sequence is used as the training set output, and the preliminary R-ELM model is obtained through training;

[0011] Step S4: In the fault diagnosis stage, save the parameters of the R-ELM model trained in step S3, and use the trained model as a digital twin virtual motor; use synchronous sampling technology to collect the electrical angle of the physical entity motor at the same time when it is running in real time , torque and phase voltage , and add the motor winding current at the previous moment , input the data into the digital twin virtual motor and obtain the winding current of the digital twin virtual motor through online training , and update the parameters of the digital twin virtual motor in real time;

[0012] Step S5: In the fault diagnosis stage, the actual winding current of the physical motor when it is running is calculated. and digital twin virtual motor winding current Subtract and get the current residual ;

[0013] Step S6: In the fault diagnosis stage, the current residual The absolute value of the difference The size is used as a criterion for judging whether the motor has a winding open circuit fault or a power tube open circuit fault;

[0014] Step S7: In the fault diagnosis stage, the current residual Use the sliding average window method to obtain the residual mean of each phase current ; Secondly, the absolute value of the current of each phase winding of the physical motor Calculate the sliding average window average and obtain the average of the absolute values ​​of each phase current ; Finally, three diagnostic variables are obtained by calculation D n Conduct fault location.

[0015] Step S8: In the fault diagnosis stage, the difference between the relative mean square error of a phase of the motor and the relative mean square error of other phases is calculated. The absolute value of the relative mean square error of a phase of the motor is used as the diagnostic variable. When the absolute value of the difference between the relative mean square error of the other phases is greater than the set threshold When , it is determined that a turn-to-turn short circuit fault occurs in this phase.

[0016] Furthermore, when the step S1 uses synchronous sampling technology to collect motor parameter data to build a database, the collected data is preprocessed, null values ​​in the database are deleted, and obvious abnormal values ​​in the database are removed using the Z-score method.

[0017] Furthermore, the random forest model in step S2 is composed of T decision trees, the number of T is determined according to the actual working conditions, and step S2 includes the following steps:

[0018] Step S2-1: Randomly select a subset from the database as a training set for training, and the motor parameters in the subset are used as the input set , the corresponding winding current is used as output;

[0019] Step S2-2: By predicting each decision tree, the predicted values ​​of all decision trees are Take the average value to get the predicted value of winding current for:

[0020] ;

[0021] Step S2-3: Randomly select a subset as the test set and calculate the baseline mean square error of the model on the test set , and then randomly select a subset of the training set to calculate the new predicted value and its corresponding new mean square error ;

[0022] Step S2-4, calculate the feature importance by the difference between the baseline mean square error and the new mean square error:

[0023] ,

[0024] in is the importance score of each motor parameter, and MSE is the mean square error.

[0025] Furthermore, the R-ELM model in step S2 is a three-layer neural network with a single hidden layer and an autoregressive loop structure, wherein the three layers are an input layer, a hidden layer, and an output layer, wherein the input layer has four input ports, a is the number of data for each motor parameter in the database, and the input layer ports are electrical angles, respectively. , torque and phase voltage and the motor winding current at the previous moment ; The output layer has 1 output port, and the output is the winding current of the digital twin virtual motor ; The number of neurons in its hidden layer is M, and the specific value of M is adjusted according to the actual working conditions. The hidden layer is composed of a cyclic structure, and the output of the hidden layer depends on the current input and the hidden layer output at the previous moment.

[0026] Furthermore, the output of the R-ELM model in step S3 is for:

[0027] ,

[0028] in, and b are the weights and biases from the input layer to the hidden layer of the R-ELM network, respectively; U, V, and c are the weights and biases of the RNN layer. is a matrix with 4 rows and M columns, b is a column vector with M rows; g is the output from the input layer to the hidden layer of the R-ELM network, which is a column vector with M rows, and h is the output of the hidden layer; H is the activation function of the neurons in the hidden layer, and the activation function uses the Sigmoid function; is the output of the hidden layer, which is the column vector of row a; is the generalized inverse matrix of H, β is the weight from the hidden layer to the output layer of the R-ELM network, which is M rows and 1 column, and is solved by the least squares method in the model training stage of step S3.

[0029] Furthermore, in step S3, the R-ELM model is trained by online training, and the feature quantity input at time t for:

[0030] ,

[0031] in The electrical angle in the database , torque , Phase voltage and the data set of the motor winding current at the previous moment; Represents the electrical angle of the motor at time t in the database; represents the electrical angle of the motor at time t, represents the phase voltage of the motor at time t, Represents the motor winding current at time t-1; the feature quantities are input into the R-ELM model in chronological order for model training.

[0032] Furthermore, in step S4, the trained R-ELM model is used as a digital twin virtual motor, and synchronous sampling technology is used to collect the electrical angle, torque and phase voltage data of the physical entity motor at the same time when it is running in real time. The collected real-time data is input into the digital twin virtual motor model in the form of streaming data to obtain the digital twin virtual motor winding current. And process historical data through the sliding window method. The model only retains the current data points of the physical entity motor in the most recent fundamental cycle at each time step as the real-time training set for online training.

[0033] Furthermore, the current residual of the motor winding current at time t in step S5 is for:

[0034] ,

[0035] in is the actual current of the physical motor winding at time t, is the digital twin virtual motor winding current at time t;

[0036] The difference value of the current residual of the motor winding in step S6 for:

[0037] ,

[0038] in for the latter moment;

[0039] The judgment basis for open circuit fault diagnosis is:

[0040] ,

[0041] in d 1 The open circuit fault diagnosis threshold is set according to the specific working conditions;

[0042] The formula of the sliding average window method in step S7 is:

[0043] ,

[0044] Where N is the number of sampling points in one cycle, is the angular frequency;

[0045] The diagnostic variable for fault location in step S7 D n for:

[0046] ,

[0047] Where n is A, B, C, representing the A, B, C phases of the physical motor. is the residual mean of each phase current, is the average absolute value of the current of each phase winding of the digital twin virtual motor;

[0048] The judgment basis for locating the open circuit fault is:

[0049] ,

[0050] in d 2 is the threshold for open circuit fault diagnosis and positioning, which is determined according to specific working conditions.

[0051] Furthermore, the calculation time The symmetrical mean absolute error of each phase current at all sampling points in is:

[0052] ,

[0053] in is the sampling frequency, is a sampling cycle time;

[0054] Relative symmetric mean absolute error As an indicator of motor turn-to-turn short-circuit fault, the relative symmetric mean absolute error is the difference between the actual symmetric mean absolute error and the initial symmetric mean absolute error, and the formula is:

[0055] ,

[0056] in is the symmetric mean absolute error of the R-ELM model, that is, the initial symmetric mean absolute error;

[0057] During operation, the relative symmetric mean absolute error of one phase of the motor is subtracted from the relative symmetric mean absolute error of other phases to obtain the difference , the formula is:

[0058] ,

[0059] in, For motor n The relative symmetric mean absolute error of the phase, For motor m The relative symmetric mean absolute error of the phase, m For n Other phases besides

[0060] Will As a diagnostic variable, the diagnostic threshold of the turn-to-turn short-circuit fault is set to , the threshold is determined by the specific working conditions, then the diagnosis basis of turn-to-turn short-circuit fault is:

[0061] ,

[0062] in d 3>0;

[0063] The fault degree of inter-turn short-circuit fault is divided into 4 levels: 15%, 35%, 60%, and 75%, which represent the percentage of short-circuited coils in all winding coils. The magnitude of the short circuit fault is used to determine the degree of the short circuit fault. The basis for determination is:

[0064] ,

[0065] in, To determine the threshold value of the inter-turn short-circuit fault degree set according to the working conditions, n =1,2,3,4,5, the threshold is determined by the specific working conditions.

[0066] Furthermore, the motor drive system is composed of a permanent magnet fault-tolerant motor of each phase and a set of three-phase half-bridge inverters. The power switch tubes T1, T2, T3, T4, T5, and T6 constitute the upper and lower bridge arms of the three-phase half-bridge inverter. The inverter bus terminal is DC power supply, each phase of the permanent magnet fault-tolerant motor adopts a 12-slot 8-pole fractional slot double-layer concentrated winding structure, the rotor structure is a surface-mounted permanent magnet rotor, the permanent magnet structure is a centrifugal tile structure, the permanent magnet is divided into a two-stage structure with an angle of 3°, and the rotor core is made of multi-layer insulated cold-rolled silicon steel sheets stacked along the axial direction of the motor rotor. The digital twin virtual motor consists of three parallel R-ELM networks; when a winding open circuit fault occurs in the permanent magnet motor drive system and the power device of a phase winding bridge arm is open, the fault phase PWM wave signal is blocked and Isolate and turn on the bidirectional thyristor TR. At this time, the motor drive system is a four-switch two-phase fault-tolerant topology; when the permanent magnet motor drive system is in a healthy state, the bidirectional thyristor is in a closed state, and the motor drive system is a half-bridge inverter topology for each phase; when an open-circuit fault occurs in the winding or an open-circuit fault occurs in a phase power tube, the motor drive system blocks the faulty bridge arm, and the motor works in a four-switch two-phase drive system. The neutral current in the inverter system is eliminated by the bidirectional thyristor connected to the neutral point of the capacitor and the neutral point of each phase of the permanent magnet fault-tolerant motor, and the neutral current remains zero.

[0067] Compared with the prior art, the present invention adopts the above technical solution and has the following beneficial effects:

[0068] (1) The present invention provides an intelligent motor fault diagnosis method based on digital twins. It does not need to collect data on various motor faults. Only the normal operation data of the motor is needed to diagnose the fault, which greatly reduces the difficulty of diagnostic data collection. It can also diagnose the open circuit fault of the power tube of the motor drive system, the open circuit fault of the motor winding and the inter-turn short circuit fault at the same time.

[0069] (2) The present invention provides an intelligent motor fault diagnosis method based on digital twins, which can obtain accurate motor winding current without knowing the precise parameters of the motor in advance.

[0070] (3) The present invention provides an intelligent motor fault diagnosis method based on digital twins. The proposed R-ELM model has both the speed of an extreme learning machine and the accuracy of an autoregressive recurrent neural network.

[0071] (4) The present invention provides an intelligent motor fault diagnosis method based on digital twins. The synchronous sampling technology used can very accurately collect different parameter data of the motor at the same time, ensuring the time consistency of each parameter data and the validity of the training data.

[0072] (5) The present invention provides an intelligent motor fault diagnosis method based on digital twins. The R-ELM model is trained using health data during actual operation, so it can replicate the state of the target motion more accurately than the observer-based mathematical model.

[0073] (6) The present invention provides a motor fault intelligent diagnosis method based on digital twins, which uses a random forest model to extract the feature quantities required for training. It does not need to take all feature parameters as input, thereby reducing the model training time and ensuring the accuracy of the model.

[0074] (7) Compared with other machine learning and neural network methods, the R-ELM model training time is significantly shortened, the algorithm implementation is easy, the learning rate is fast, the prediction accuracy is high, it is easy to use, and it can guarantee a unique optimal solution while having good generalization performance.

[0075] (8) The present invention provides an intelligent motor fault diagnosis method based on digital twins. The proposed permanent magnet fault-tolerant motor for each phase not only has the advantages of high output torque, high power density and high efficiency of the permanent magnet synchronous motor, but also has the fault tolerance capabilities of magnetic isolation, thermal isolation and electrical isolation.

[0076] (9) The present invention provides an intelligent motor fault diagnosis method based on digital twins, which has an extremely fast diagnosis speed and a simple diagnostic algorithm, is superior to general fault diagnosis methods, and does not require the formulation of complex thresholds.

[0077] (10) The present invention provides an intelligent motor fault diagnosis method based on digital twins. For an open circuit fault, the differential value of the current residual is used as the standard for judging whether a fault occurs. For a turn-to-turn short circuit fault, the relatively symmetrical mean absolute error is used as the standard for judging whether a fault occurs. The method has strong robustness and can effectively prevent misdiagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 This is a flowchart of a method for intelligent diagnosis of motor faults based on digital twins of the present invention;

[0079] Figure 2 It is a control and fault diagnosis strategy block diagram of the online fault diagnosis stage of the present invention;

[0080] Figure 3 is the A-phase winding current when the physical entity motor of the present invention is operating normally;

[0081] Figure 4 is the A-phase winding current when the digital twin motor of the present invention is operating normally;

[0082] Figure 5 is the current residual value of the physical entity motor and the digital twin motor of the present invention during normal operation;

[0083] Figure 6 It is a waveform diagram of currents of each phase when an open circuit fault occurs in the A-phase winding of the physical entity motor of the present invention at 0.05s;

[0084] Figure 7 is the speed of the physical entity motor of the present invention when an open circuit fault occurs in the A-phase winding at 0.05s;

[0085] Figure 8 is the torque of the physical entity motor of the present invention when an open circuit fault occurs in the A-phase winding at 0.05s;

[0086] Fig. 9 is the electrical angle of the physical entity motor of the present invention when an open circuit fault occurs in the A-phase winding at 0.05s;

[0087] Fig.10 The actual waveform and digital twin waveform diagram of the current of the A-phase winding of the physical entity motor of the present invention when an open circuit fault occurs in the A-phase winding at 0.08s;

[0088] Fig.11 is a fault diagnosis identifier for a physical entity motor of the present invention when an open circuit fault occurs in the A-phase winding at 0.08s;

[0089] Fig.12 The actual waveform of the A-phase winding current and the digital twin waveform diagram of the physical entity motor of the present invention when an open circuit fault occurs on the power tube on the A-phase inverter at 0.08s;

[0090] Fig.13 is a fault diagnosis identifier for a physical entity motor of the present invention when an open circuit fault occurs on the power tube of the A-phase inverter at 0.08s;

[0091] Fig.14 The actual current of the A-phase winding and the digital twin current diagram of the motor of the present invention before and after the A-phase winding open circuit fault under the condition of torque mutation;

[0092] Fig.15 It is a diagnostic identification diagram when the A-phase winding of the motor of the present invention is open-circuited under the condition of torque mutation;

[0093] Fig.16 A new extreme learning machine with an autoregressive cycle structure proposed by the present invention. DETAILED DESCRIPTION

[0094] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings of the specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work are within the scope of protection of the present invention.

[0095] Example:

[0096] As attached Figure 1As shown, it is a flowchart of a motor fault intelligent diagnosis method based on digital twins of the present invention, which specifically includes the following steps:

[0097] Step S1: In the model training stage, when synchronous sampling technology is used to collect motor parameter data to build a database, the collected data is preprocessed to delete null values ​​in the database and use the Z-score method to remove obvious outliers in the database.

[0098] Step S2: In the model training phase, the random forest method is used to sort the motor operation data in the database by importance, and the most important parameters are selected to train the R-ELM model to reduce the training cost and training time of the model. The random forest model consists of T decision trees, and the number of T is determined according to the actual working conditions, including the following steps:

[0099] Step S2-1: Randomly select a subset from the database as a training set for training, and the motor parameters in the subset are used as the input set , the corresponding winding current is used as output;

[0100] Step S2-2: By predicting each decision tree, the predicted values ​​of all decision trees are Take the average value to get the predicted value of winding current for:

[0101] ,

[0102] Step S2-3: Randomly select a subset as the test set and calculate the baseline mean square error of the model on the test set , and then randomly select a subset of the training set to calculate the new predicted value and its corresponding new mean square error ;

[0103] Step S2-4, calculate the feature importance by the difference between the baseline mean square error and the new mean square error:

[0104] ,

[0105] in is the importance score of each motor parameter, and MSE is the mean square error.

[0106] Step S3: In the model training phase, the electrical angle at the same time in the database is used , torque and phase voltage and the motor winding current at the previous moment The data is input as a training set, as shown in the attached Fig.16The R-ELM model shown in the figure is a three-layer neural network with a single hidden layer and an autoregressive loop structure. The three layers are the input layer, the hidden layer and the output layer. The input layer has 4 input ports, a is the number of data for each motor parameter in the database, and the input layer ports are the electrical angle , torque and phase voltage and the motor winding current at the previous moment ; The output layer has 1 output port, and the output is the winding current of the digital twin virtual motor ; The number of neurons in its hidden layer is M, and the specific value is adjusted according to the actual working conditions. The hidden layer is composed of a cyclic structure, and the output of the hidden layer depends on the current input and the hidden layer output at the previous moment.

[0107] The output of the R-ELM model for:

[0108] ,

[0109] in, and b are the weights and biases from the input layer to the hidden layer of the R-ELM network, respectively; U, V, and c are the weights and biases of the RNN layer. is a matrix with 4 rows and M columns, b is a column vector with M rows; g is the output from the input layer to the hidden layer of the R-ELM network, which is a column vector with M rows, and h is the output of the hidden layer; H is the activation function of the neurons in the hidden layer, and the activation function uses the Sigmoid function; is the output of the hidden layer, which is the column vector of row a; is the generalized inverse matrix of H, β is the weight from the hidden layer to the output layer of the R-ELM network, which is M rows and 1 column, and is solved by the least squares method in the model training stage of step S3.

[0110] The R-ELM model is trained by online training. The feature quantity input at time t for:

[0111] ,

[0112] in The electrical angle in the database , torque and phase voltage and a data set of the winding current of the physical motor at the previous moment; Represents the electrical angle of the motor at time t in the database; represents the electrical angle of the motor at time t, represents the phase voltage of the motor at time t, Represents the motor winding current at time t-1; the feature quantities are input into the R-ELM model in chronological order for model training.

[0113] Step S4: In the fault diagnosis stage, the trained R-ELM model is used as a digital twin virtual motor, and the synchronous sampling technology is used to collect the electrical angle, torque and phase voltage data of the physical motor at the same time when it is running in real time. The collected real-time data is input into the digital twin virtual motor model in the form of streaming data to obtain the digital twin virtual motor winding current. The model processes historical data through a sliding window method, and only retains the physical entity motor current data points of the most recent fundamental cycle as a real-time training set at each time step.

[0114] Step S5: In the fault diagnosis stage, the actual winding current of the physical motor when it is running is calculated. and digital twin virtual motor winding current Subtract and get the current residual of the motor winding current at time t for:

[0115] ,

[0116] in is the actual current of the physical motor winding at time t, is the digital twin virtual motor winding current at time t.

[0117] Step S6: In the fault diagnosis stage, the current residual The difference value The magnitude is used as the criterion for judging whether the motor has an open-circuit fault in the winding or an open-circuit fault in the power tube. The differential value of the residual current of the motor winding for:

[0118] ,

[0119] in for the latter moment;

[0120] The basis for fault diagnosis is:

[0121] ,

[0122] in d 1 is the set fault diagnosis threshold.

[0123] Step S7: In the fault diagnosis stage, the current residual Use the sliding average window method to obtain the residual mean of each phase current ; Secondly, the absolute value of the current of each phase winding of the physical motor Calculate the sliding average window average and obtain the average of the absolute values ​​of each phase current ; Finally, three diagnostic variables are obtained by calculation D n Conduct fault location.

[0124] The formula of the sliding average window method is:

[0125] ,

[0126] Where N is the number of sampling points in one cycle, is the angular frequency;

[0127] The diagnostic variables for the fault location D n for:

[0128] ,

[0129] Where n is A, B, C, representing the A, B, and C phases of the physical motor. is the residual mean of each phase current, is the average absolute value of the current of each phase winding of the digital twin virtual motor;

[0130] The basis for locating the fault is:

[0131] ,

[0132] in d 2 is the threshold value set for fault diagnosis and positioning.

[0133] Step S8: To eliminate the influence of fluctuation and noise in the residual vector, calculate the time The symmetrical mean absolute error of each phase current at all sampling points in is:

[0134] ,

[0135] in is the sampling frequency, is a sampling cycle time;

[0136] Since the training error of the neural network cannot be reduced to 0, in order to eliminate the interference of the initial error and show the fault characteristics more clearly, the relatively symmetrical mean absolute error is used. As an indicator of motor turn-to-turn short-circuit fault, the relative symmetric mean absolute error is the difference between the actual symmetric mean absolute error and the initial symmetric mean absolute error, and the formula is:

[0137] ,

[0138] in is the symmetrical mean absolute error of the R-ELM model, that is, the initial symmetrical mean absolute error. During normal operation, the relative mean square error of the three-phase current fluctuates very little around 0. If the relative symmetrical mean absolute error of the phase current exceeds the upper limit of the normal value, it indicates that the motor is faulty. In addition, if the relative symmetrical mean absolute error of a phase is much larger than the relative symmetrical mean absolute error of other phases, it indicates that the motor is currently experiencing a turn-to-turn short circuit fault.

[0139] During operation, the relative symmetric mean absolute error of one phase of the motor is subtracted from the relative symmetric mean absolute error of other phases to obtain the difference , the formula is:

[0140] ,

[0141] in, For motor n The relative symmetric mean absolute error of the phase, For motor m The relative symmetric mean absolute error of the phase, m For n Other phases besides

[0142] Will As a diagnostic variable, the diagnostic threshold of the turn-to-turn short-circuit fault is set to , the threshold is determined by the specific working conditions, then the diagnosis basis of turn-to-turn short-circuit fault is:

[0143] ,

[0144] in d 3>0;

[0145] The fault degree of inter-turn short-circuit fault is divided into 4 levels: 15%, 35%, 60%, and 75%, which represent the percentage of short-circuited coils in all winding coils. The magnitude of the short-circuit fault is used to determine the degree of the short-circuit fault. The basis for determination is:

[0146] ,

[0147] in, To determine the threshold value of the inter-turn short-circuit fault degree set according to the working conditions, n =1,2,3,4,5, the threshold is determined by the specific working conditions.

[0148] As attached Figure 2The figure shows a motor fault diagnosis block diagram of the present invention. The permanent magnet fault-tolerant motor drive system of each phase is composed of a permanent magnet fault-tolerant motor of each phase and a set of half-bridge inverters of each phase. The power switch tubes T1, T2, T3, T4, T5, and T6 constitute the upper and lower bridge arms of the half-bridge inverter of each phase. The inverter bus terminal is DC power supply, the error of the PI controller of the outer speed loop is calculated by the reference speed and the differential value of the position sensor of each phase FTPMM. The PI controller output is q Axis reference current i q * At the same time, each phase FTPMM current i A , i B 、i C Rotating the coordinate system by coordinate transformation d Shaft current i d and q Shaft current i q , the vector control method adopted d Reference current of the axis i d * =0, respectively dq The axis reference current and actual current are input to the current PI controller, and the output is dq Axis reference voltage d Axis reference voltage V d * and q Axis reference voltage V q * , perform coordinate transformation to obtain the voltage of the stationary coordinate system V a *、V β * , perform voltage space vector pulse width modulation, output six PWM waves of each phase winding, control the on and off of the power tube, and realize the drive control of each phase FTPMM.

[0149] The permanent magnet fault-tolerant motor of each phase adopts a 12-slot 8-pole fractional slot double-layer concentrated winding structure, the rotor structure is a surface-mounted permanent magnet rotor, the permanent magnet structure is a centrifugal tile structure, the permanent magnet is divided into a two-stage structure with an angle of 3°, and the rotor core is made of multi-layer insulated cold-rolled silicon steel sheets stacked along the axial direction of the motor rotor. The digital twin virtual motor consists of three parallel R-ELM networks. When a permanent magnet motor drive system has a winding open circuit fault and a power device of a phase winding bridge arm is open, the fault phase PWM wave signal is blocked and isolated, and at the same time, the bidirectional thyristor TR is turned on. The motor drive system is now a four-switch two-phase fault-tolerant topology. When the permanent magnet motor drive system is in a healthy state, the bidirectional thyristor is in the off state, and the motor drive system is a half-bridge inverter topology for each phase; when an open-circuit fault occurs in the winding or an open-circuit fault occurs in a phase power tube, the motor drive system blocks the faulty bridge arm, and the motor works in a four-switch two-phase drive system. The neutral current of the inverter system is not zero, but can be eliminated by the bidirectional thyristor connected to the neutral point of the capacitor and the neutral point of each phase of the permanent magnet fault-tolerant motor, and the neutral current remains zero.

[0150] The diagnostic effect of the present invention is taken as an example of phase A of a permanent magnet fault-tolerant motor of each phase, and the physical entity motor runs at a speed of 1000rpm, a load torque of 6N.m and a DC power supply of 270V. Figure 3 and attached Figure 4 As shown, attached Figure 3 is the current of the A phase winding of the physical motor and attached Figure 4 Predict the current of the digital twin of the A phase winding of the digital twin virtual motor , as attached Figure 5 Shown and From the current residual value, we can see that the current residual value is around 1A, which is approximately 5% of the current amplitude of the physical motor winding. This shows that the prediction accuracy of the digital twin motor is very high and it can accurately predict the operating status of the physical motor.

[0151] As attached Figure 6 The figure shows the current diagram of each phase winding of a physical motor, that is, each phase permanent magnet fault-tolerant motor when an A-phase winding open circuit fault occurs at 0.05s. It can be seen that when an open circuit fault occurs in the motor, the A-phase current suddenly changes to 0, and the amplitude of the remaining healthy phase winding current increases and remains equal, showing a fault-tolerant operation state. Figure 7 The speed of the physical motor shown in the figure does not fluctuate drastically after the open circuit fault and still runs at a speed of 1000 rpm; as shown in the attached figure Figure 8 The load torque of the physical motor shown in the figure does not fluctuate drastically before and after the open circuit fault, and remains at 6N.m. Fig. 9As shown in the figure, the electrical angle of the physical motor did not fluctuate dramatically after the open circuit fault occurred at 0.05s. In summary, the motor has a good fault tolerance effect. Because the physical motor is in a fault-tolerant operating state after the fault, the torque, speed and electrical angle data of the motor can still be consistent with those before the motor fault, so the digital twin motor is still running in a healthy state.

[0152] As attached Fig.10 The figure shows that the physical motor has an open-circuit fault in the A-phase winding at 0.08s. The current of the A-phase winding of the physical motor suddenly changes to 0 due to the open-circuit fault at 0.08s, while the current of the digital twin still runs in a normal state, as shown in the attached figure. Fig.11 As shown in FIG. 1 , the motor fault identifier changes within about 10 μs to diagnose an open-circuit fault in the motor winding, and the diagnosis speed is extremely fast.

[0153] As attached Fig.12 The physical motor has an open circuit fault on the power tube of phase A at 0.05s. The power tube current of phase A of the physical motor changes suddenly due to the open circuit fault at 0.05s, and the current is negative half-wave. The digital twin current still runs in a normal state, as shown in the attached figure. Fig.13 As shown, the motor fault identifier jumps in about 10μs to diagnose an open-circuit fault in the motor winding, and the diagnosis speed is extremely fast.

[0154] As attached Fig.14 As shown in the figure, the physical motor has a sudden torque change at 0.03s, and the digital twin virtual motor can still predict the motor winding current well. And the physical motor has a winding open circuit fault at 0.08s. The fault diagnosis identifier can still jump normally, as shown in the attached figure. Fig.15 The differential value of the motor shown in the figure has an obvious mutation at 0.08s when the winding is open, which is 6 times the normal state. Therefore, using the differential value as the judgment standard will not cause misdiagnosis, has strong robustness and fast diagnosis speed.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A motor fault intelligent diagnosis method based on digital twin, characterized in that: The method comprises a model training phase and a fault diagnosis phase, and specifically comprises the following steps: Step S1: In the model training stage, synchronous sampling technology is used to collect various timing motor parameters when the motor is operating normally under different working conditions to build a database; Step S2: In the model training phase, the random forest model is trained to sort the importance of all motor parameters, and the electrical angle θ(t), torque T at the same time in the database are selected according to the importance. e (t) and phase voltage u dc (t), and add the physical entity motor winding current i at the previous moment rn (t-1), where n represents each phase of the motor, which is used as the input for training the R-ELM model; the R-ELM model is an improved extreme learning machine model with autoregressive loop function; the R-ELM model extracts features by the extreme learning machine, and then performs time series modeling and predicts output through a recurrent neural network structure; Step S3: In the model training stage, an online training method is used to train the model using the electrical angle θ(t) and torque T during normal operation in the diagnostic database. e (t) and phase voltage u dc (t), and add the motor winding current i at the previous moment rn (t-1) is used as the training set input, and the winding current data corresponding to the feature quantity in time sequence is used as the training set output, and the preliminary R-ELM model is obtained through training; Step S4: In the fault diagnosis stage, save the parameters of the R-ELM model trained in step S3, and use the trained model as a digital twin virtual motor; use synchronous sampling technology to collect the electrical angle θ(t) and torque T at the same time when the physical entity motor is running in real time. e (t) and phase voltage u dc (t), and add the motor winding current i at the previous moment rn (t-1), input the data into the digital twin virtual motor and obtain the digital twin virtual motor winding current i through online training pn , and update the parameters of the digital twin virtual motor in real time; Step S5: In the fault diagnosis stage, the actual winding current i of the physical motor when it is running is calculated. rn and the digital twin virtual motor winding current i pn Subtract and get the current residual i error_n ; Step S6: In the fault diagnosis stage, the current residual i error_n The absolute value of the difference value |Δi is used as the criterion for judging whether the motor has a winding open circuit fault or a power tube open circuit fault; Step S7: In the fault diagnosis stage, the current residual i error_n Use the sliding average window method to obtain the residual mean of each phase current Secondly, the absolute value of the current of each phase winding of the physical motor |i rn |Calculate the sliding average window average to obtain the average of the absolute values ​​of each phase current Finally, three diagnostic variables D are obtained by calculation n Conduct fault location. Step S8: In the fault diagnosis stage, the difference Q between the relative symmetric mean absolute error of a certain phase of the motor and the relative symmetric mean absolute error of other phases is calculated. n The absolute value of is used as the diagnostic variable. When the absolute value of the difference between the relative symmetrical mean absolute error of a phase of the motor and the relative symmetrical mean absolute error of other phases is greater than the set threshold d3, it is determined that a turn-to-turn short circuit fault occurs in this phase.

2. According to claim 1, a motor fault intelligent diagnosis method based on digital twins is characterized in that: When the step S1 uses synchronous sampling technology to collect motor parameter data to build a database, the collected data is preprocessed to delete null values ​​in the database and use the Z-score method to remove obvious abnormal values ​​in the database.

3. The method for intelligent motor fault diagnosis based on digital twin according to claim 1 is characterized in that: The random forest model in step S2 is composed of T decision trees, where the number of T is determined according to the actual working conditions. Step S2 includes the following steps: Step S2-1: Randomly select a subset from the database as a training set for training, and the motor parameters in the subset are used as the input set X n , the corresponding winding current is used as output; Step S2-2: By predicting each decision tree, the predicted value f of all decision trees is t Take the average value to get the predicted value of winding current for: Step S2-3: Randomly select a subset as the test set and calculate the baseline mean square error (MSE) of the model on the test set. original , and then randomly select a subset of the training set to calculate the new predicted value and its corresponding new mean square error MSE shuffled ; Step S2-4, calculate the feature importance by the difference between the baseline mean square error and the new mean square error: Importance(X n )=MSE original -MSE shuffled Importance(X n ) is the importance score of each motor parameter, and MSE is the mean square error.

4. The method for intelligent diagnosis of motor faults based on digital twins according to claim 1 is characterized in that: The R-ELM model in step S2 is a three-layer neural network with a single hidden layer and an autoregressive loop structure, wherein the three layers are respectively an input layer, a hidden layer and an output layer, wherein the input layer has four input ports, a is the number of data for each motor parameter in the database, and the input layer ports are respectively electrical angle θ(t), torque T e (t) and phase voltage u dc (t) and the motor winding current i at the previous moment rn (t-1); the output layer has one output port, and the output is the digital twin virtual motor winding current i pn ; The number of neurons in its hidden layer is M, and the specific value of M is adjusted according to the actual working conditions. The hidden layer is composed of a cyclic structure, and the output of the hidden layer depends on the current input and the hidden layer output at the previous moment.

5. The method for intelligent diagnosis of motor faults based on digital twins according to claim 1 is characterized in that: The output of the R-ELM model in step S3 is q for: Among them, ω and b are the weights and biases from the input layer to the hidden layer of the R-ELM network, respectively; U, V and c are the weights and biases of the RNN layer; ω is a matrix with 4 rows and M columns, and b is a column vector with M rows; g is the output from the input layer to the hidden layer of the R-ELM network, which is a column vector with M rows; h is the output of the hidden layer; H is the activation function of the hidden layer neurons, and the activation function uses the Sigmoid function; I q (x) is the output of the hidden layer, which is the column vector of row a; H + is the generalized inverse matrix of H, β is the weight from the hidden layer to the output layer of the R-ELM network, which is M rows and 1 column, and is solved by the least squares method in the model training stage of step S3.

6. The method for intelligent diagnosis of motor faults based on digital twins according to claim 1 is characterized in that: In step S3, the R-ELM model is trained by online training. The feature value x input at time t n for: x n =[θ(t),T e (t),u dc (t),i rn (t-1)] where x n The electrical angle θ and torque T in the database e , phase voltage u dc and the data set of the motor winding current at the previous moment; θ(t) represents the electrical angle of the motor at time t in the database; T e (t) represents the electrical angle of the motor at time t, u dc (t) represents the phase voltage of the motor at time t, i rn (t-1) represents the motor winding current at time t-1; the feature quantities are input into the R-ELM model in chronological order for model training.

7. The method for intelligent diagnosis of motor faults based on digital twins according to claim 1 is characterized in that: In step S4, the trained R-ELM model is used as a digital twin virtual motor, and the synchronous sampling technology is used to collect the electrical angle, torque and phase voltage data of the physical entity motor at the same time when it is running in real time. The collected real-time data is input into the digital twin virtual motor model in the form of streaming data to obtain the digital twin virtual motor winding current i pn And process historical data through the sliding window method. The model only retains the current data points of the physical entity motor in the most recent fundamental cycle at each time step as the real-time training set for online training.

8. The method for intelligent diagnosis of motor faults based on digital twins according to claim 1 is characterized in that: The current residual i of the motor winding current at time t in step S5 is error_n for: i error_n (t)=i rn (t)-i pn (t) where i rn (t) is the actual current of the physical motor winding at time t, i pn (t) is the winding current of the digital twin virtual motor at time t; The differential value Δi of the current residual of the motor winding in step S6 is: where t n t n-1 the latter moment; The judgment basis for open circuit fault diagnosis is: Where d1 is the set open circuit fault diagnosis threshold, which is determined according to the specific working conditions; The formula of the sliding average window method in step S7 is: Where N is the number of sampling points in one cycle, ω s is the angular frequency; The diagnostic variable D of the fault location in step S7 n for: Where n is A, B, C, representing the A, B, C phases of the physical motor. is the residual mean of each phase current, is the average absolute value of the current of each phase winding of the digital twin virtual motor; The judgment basis for locating the open circuit fault is: Wherein d2 is the threshold for open circuit fault diagnosis and positioning, which is determined according to specific working conditions.

9. The method for intelligent diagnosis of motor faults based on digital twins according to claim 1 is characterized in that: Calculation time T s The symmetrical mean absolute error of each phase current at all sampling points in is: where f s is the sampling frequency, T s is a sampling cycle time; Relative symmetric mean absolute error As an indicator of motor turn-to-turn short-circuit fault, the relative symmetric mean absolute error is the difference between the actual symmetric mean absolute error and the initial symmetric mean absolute error, and the formula is: where e smpae0 is the symmetric mean absolute error of the R-ELM model, that is, the initial symmetric mean absolute error; During operation, the relative symmetric mean absolute error of one phase of the motor is subtracted from the relative symmetric mean absolute error of other phases to obtain the difference Q n-m , the formula is: in, is the relative symmetric mean absolute error of the nth phase of the motor, is the relative symmetric mean absolute error of the mth phase of the motor, where m is the phase other than n; Q n-m As a diagnostic variable, the diagnostic threshold of the turn-to-turn short-circuit fault is set to d3. The threshold is determined by the specific working conditions. The diagnosis basis of the turn-to-turn short-circuit fault is: Where d3>0; The fault degree of inter-turn short-circuit fault is divided into 4 levels: 15%, 35%, 60%, and 75%, which represent the percentage of short-circuited coils in the total winding coils. n-m The magnitude of the short circuit fault is used to determine the degree of the short circuit fault. The basis for determination is: Among them, μ n The threshold value for determining the degree of inter-turn short-circuit fault is set according to the working conditions, n=1, 2, 3, 4, 5, and the threshold value is determined by the specific working conditions.

10. The method for intelligent diagnosis of motor faults based on digital twins according to claim 1, characterized in that: The motor drive system consists of a permanent magnet fault-tolerant motor of each phase and a three-phase half-bridge inverter. The power switch tubes T1, T2, T3, T4, T5, and T6 form the upper and lower bridge arms of the three-phase half-bridge inverter. The inverter bus terminal is U dc DC power supply, each phase of the permanent magnet fault-tolerant motor adopts a 12-slot 8-pole fractional slot double-layer concentrated winding structure, the rotor structure is a surface-mounted permanent magnet rotor, the permanent magnet structure is a centrifugal tile structure, the permanent magnet is divided into a two-stage structure with an angle of 3°, and the rotor core is made of multi-layer insulated cold-rolled silicon steel sheets stacked along the axial direction of the motor rotor. The digital twin virtual motor consists of three parallel R-ELM networks; when a winding open circuit fault occurs in the permanent magnet motor drive system and the power device of a phase winding bridge arm is open, the fault phase PWM wave signal is blocked and Isolate and turn on the bidirectional thyristor TR. At this time, the motor drive system is a four-switch two-phase fault-tolerant topology; when the permanent magnet motor drive system is in a healthy state, the bidirectional thyristor is in a closed state, and the motor drive system is a half-bridge inverter topology for each phase; when an open-circuit fault occurs in the winding or an open-circuit fault occurs in a phase power tube, the motor drive system blocks the faulty bridge arm, and the motor works in a four-switch two-phase drive system. The neutral current in the inverter system is eliminated by the bidirectional thyristor connected to the neutral point of the capacitor and the neutral point of each phase of the permanent magnet fault-tolerant motor, and the neutral current remains zero.

Images