Permanent magnet synchronous motor fault diagnosis method

Abstract

The invention discloses a fault diagnosis method for a magnetic synchronous motor. The method comprises the following steps: converting currently collected three-phase current and three-phase voltage into current and voltage on a d-axis and a q-axis; respectively calculating theoretical voltage values and theoretical current values on the d axis and the q axis based on the current and the voltage on the d axis and the q axis; obtaining a voltage prediction value at the current moment based on the theoretical voltage value at the current moment and the operation parameters, and obtaining a current prediction value at the current moment based on the theoretical current value at the current moment and the operation parameters; and performing anomaly detection based on the difference values between the voltage predicted value and the current predicted value and the corresponding acquisition values, and determining the fault type of the motor when the motor is detected to be abnormal. The difference index fully considers the difference of the current and the voltage on the d axis and the q axis, and can more comprehensively reflect the deviation degree between the motor operation state and model prediction, thereby improving the fault diagnosis precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of motor control, and more specifically, the present invention relates to a permanent magnet synchronous motor fault diagnosis method. Background Art

[0002] Permanent magnet synchronous motor (PMSM) is highly favored in the industrial field due to its excellent performance, stability and compact design. However, PMSM may encounter various faults during operation, such as rotor eccentricity, stator winding short circuit, permanent magnet demagnetization, etc. These problems will not only interfere with the normal operation of the motor, but also may cause serious safety risks.

[0003] Traditional fault diagnosis technology often relies on experience and data from specific sensors. The above diagnosis has the following defects:

[0004] (1) High dependence: Traditional methods are highly dependent on the detection signals of sensors. When the sensor detection shows a mutation or abnormality, the reliability of the entire diagnostic system will be greatly reduced;

[0005] (2) Lack of accuracy: Experience-based diagnostic methods are difficult to fully cover all possible failure modes, resulting in a high misdiagnosis rate;

[0006] (3) Poor adaptability: Existing technologies lack sufficient adaptability to PMSMs of different models and under different operating conditions, and diagnostic strategies need to be designed separately for each situation. Summary of the invention

[0007] The present invention provides a permanent magnet synchronous motor fault diagnosis method, aiming to improve at least one of the above problems.

[0008] The present invention is implemented as follows: a permanent magnet synchronous motor fault diagnosis method, the method comprising the following steps:

[0009] (1) Convert the currently collected three-phase current and three-phase voltage into the current and voltage on the d-axis and q-axis;

[0010] (2) Based on the current and voltage on the d-axis and q-axis, the theoretical voltage values ​​on the d-axis and q-axis are calculated respectively. and Theoretical current value and

[0011] (3) Based on the theoretical voltage value at the current moment and and operating parameters to obtain the voltage forecast value at the current moment and Based on the theoretical current value at the current moment and and operating parameters to obtain the current forecast value at the current moment and

[0012] (4) Anomaly detection is performed based on the difference between the voltage prediction value and the current prediction value and the corresponding collected value. When an abnormality is detected in the motor, the fault type of the motor is determined.

[0013] Furthermore, the current theoretical current value The current operating data includes three-phase voltage, three-phase current and motor temperature T, which form an input vector and are input into the trained first LSTM network. The first LSTM network outputs the current prediction value at the current moment. and

[0014] Furthermore, the current theoretical voltage value and The current operating data includes three-phase voltage, three-phase current and motor temperature T, which form an input vector and are input into the trained second LSTM network. The second LSTM network outputs the voltage prediction value at the current moment. and

[0015] Furthermore, the threshold update formula of the first LSTM network or the second LSTM network is as follows:

[0016] θ b =βθ n +(1-β)α[D n -θ n ]

[0017] Among them, θ n is the threshold formed by the last threshold update, and is also the threshold before this update. n is the difference index updated based on the running data collected at the current time interval, θ b This is the threshold formed after this update, α is the learning rate, and β is the forgetting factor.

[0018] Furthermore, the initial value determination process of the threshold is as follows:

[0019] Based on the difference indicator sequence {D1, D2, ..., D m}Construct the normal distribution probability density function f(D) of the difference index D: According to the principle of maximum likelihood estimation, construct the likelihood function L(μ,σ), and determine the estimated values ​​of μ and σ based on the likelihood function L(μ,σ). The initial value of the threshold is θ0=μ+3σ, where μ and σ are parameters of the normal distribution.

[0020] Furthermore, the collected operation data is subjected to anomaly detection. The anomaly detection process is as follows:

[0021] Determine the i-th running data x based on historical data i The mean μ i and standard deviation σ i , for the currently collected running data x i , judge |x i -μ i |>3σ i Is it true? If so, the currently collected running data x i For outliers, replace them with the average of their neighboring normal values.

[0022] Furthermore, the difference index D corresponding to the current running data is calculated based on the following formula:

[0023]

[0024] in, It represents the d-axis current and q-axis current obtained by converting the three-phase current collected at the kth moment. represents the d-axis voltage and q-axis voltage obtained by converting the three-phase voltage collected at the kth moment, They respectively represent the predicted values ​​of d-axis current, q-axis current, d-axis voltage, and q-axis voltage output by the abnormality detection model at the kth moment, and n represents the total number of the nearest historical sampling points.

[0025] Furthermore, the abnormality determination process of the motor is as follows:

[0026] When the current difference index D>θ1, and the number of times that the difference index D corresponding to the past 10 sampling points is greater than the primary threshold θ1 exceeds 60%, it is determined that the motor is abnormal, where θ1 is set to 0.7 times the current threshold;

[0027] When the current difference index D>θ2, and the number of times that the difference index D corresponding to the past 20 sampling points is greater than the intermediate threshold θ2 exceeds 70%, and the average value of the difference index shows an increasing trend, θ2 is set to 1.1 times the current threshold;

[0028] When the current difference index D>θ3, and in the past 30 sampling points, the number of times the difference index D is greater than the advanced threshold θ3 exceeds 80%, and the average value of the difference index continues to increase and the increase exceeds a set growth threshold, where θ3 is set to 1.3 times the current threshold.

[0029] Furthermore, after being determined to be abnormal, the mean, variance and spectral characteristics of the current current and voltage of the motor are input into the fault prediction models of various types of faults, and the current fault type of the motor is determined based on the prediction results output by each fault prediction model.

[0030] Furthermore, a fault prediction model is constructed for each type of fault. The fault prediction model is constructed using a convolutional neural network, and all fault prediction models are placed in a fault prediction model library.

[0031] The permanent magnet synchronous motor fault diagnosis method provided by the present invention has the following beneficial technical effects:

[0032] (1) Improving diagnostic accuracy: The current and voltage prediction model provided by the present invention comprehensively considers the physical characteristics of the motor and the nonlinear relationship in the actual operating data. Compared with the traditional method, it is more accurate in judging the motor abnormality. It adopts a detailed difference index calculation method, which fully considers the difference between the current and voltage on the d-axis and q-axis, and can more comprehensively reflect the degree of deviation between the motor operating state and the model prediction, thereby improving the detection accuracy of the motor abnormality.

[0033] (2) Reduce false alarm rate: The combination of motor physical model and machine learning model (such as long short-term memory network LSTM) avoids the misjudgment caused by the limitations of a single model. The physical model is based on the basic principles of the motor and has basic constraints on the normal operating state; the machine learning model can supplement the situation that is difficult for the physical model to handle by learning the complex relationships in a large amount of data. The two work together to effectively reduce the false alarm caused by a single factor.

[0034] (3) Improve response speed: Real-time monitoring of motor operating parameters, short data collection interval (data collection every 5 milliseconds), can timely obtain the latest operating status of the motor. Once the difference between the measured data and the model prediction value exceeds the abnormal judgment standard, fault detection is immediately started, and the fault can be quickly diagnosed in a short time. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flow chart of a permanent magnet synchronous motor fault diagnosis method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The specific implementation modes of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0037] The present invention provides a permanent magnet synchronous motor fault diagnosis method based on model difference analysis. First, a physical model of the PMSM is established. The theoretical values ​​corresponding to the current operating parameters are calculated based on the current physical model. The current and voltage of the motor are predicted based on the monitoring of the actual operating parameters and the corresponding theoretical values. Then, the difference index between the predicted value and the measured value is calculated. The abnormality detection of the motor is performed based on the difference index. After the current abnormality of the permanent magnet synchronous motor is detected, it is matched with the fault detection models of various faults in the fault model library to determine the fault type of the motor, such as Figure 1 shown.

[0038] (1) Physical model: A physical model of the motor during normal operation is constructed based on the basic physical equations of the motor. The calculation formula for the d-axis (direct-axis) voltage is as follows:

[0039]

[0040] The calculation model of the q-axis (quadrature-axis) voltage is as follows:

[0041]

[0042] Among them, V d 、V q are the voltages on the d-axis and q-axis respectively, R s is the stator resistance, L d , L q are the d-axis and q-axis inductances, ω e is the angular velocity of the motor, I d ,I q are the currents of the d-axis and q-axis respectively, λ r is the flux linkage generated by the permanent magnet.

[0043] In order to determine the parameters in the physical model (R s , L d , L q , r The method of combining experimental test with theoretical calculation is adopted. First, the basic data such as voltage, current and speed of the motor under different working conditions are measured through no-load test and short-circuit test. Then, according to these experimental data, the motor equivalent circuit theory and electromagnetic field theory are used to calculate the parameters. For example, the voltage and speed data measured by no-load test are combined with V q =ω e λ r (No-load I q ≈0) to calculate the permanent magnet flux; the stator resistance R is calculated by combining the short-circuit current and voltage data measured by the short-circuit experiment with the short-circuit equivalent circuit equation s and inductance (L d , L q ) and other parameters.

[0044] The detailed no-load test steps are as follows: connect the motor to the driver, make the motor run in a no-load state, gradually increase the motor speed to the rated speed, and run stably for a period of time. During this process, use high-precision voltage sensors and current sensors to measure the phase voltage and phase current of the motor, and use speed sensors to measure the motor speed. Record a set of data at a certain time interval (10s), and collect N sets of data in total (N≥100). For the short-circuit test, short-circuit one phase winding of the motor (such as short-circuit the U phase winding), apply a certain DC voltage (the voltage value should be lower than the rated voltage of the motor to avoid excessive current damaging the motor) when the motor is stationary, and measure the short-circuit current and short-circuit voltage. Repeat the short-circuit test multiple times (3 to 5 times), take the average value as the final short-circuit test data, and use it to calculate the stator resistance and inductance.

[0045] (2) Data collection: The three-phase current, three-phase voltage and motor temperature of the permanent magnet synchronous motor are collected at fixed time intervals (5ms) during operation. High-precision sensors (current sensor accuracy reaches ±0.2%, voltage sensor accuracy reaches ±0.1%, and temperature sensor accuracy reaches ±0.5℃) are used to ensure data accuracy. The collected data is transmitted to the computer through a high-speed data acquisition card for subsequent processing. The sampling frequency of the data acquisition card is set to 20kHz to meet the requirements of high-frequency signal acquisition.

[0046] The hardware architecture of the data acquisition system includes sensor modules, signal conditioning circuits, data acquisition cards, and computers. The sensor module is responsible for converting the motor's operating parameters (current, voltage, temperature, etc.) into electrical signals, and the signal conditioning circuit amplifies, filters, and performs analog-to-digital conversion on the electrical signals output by the sensor to meet the input requirements of the data acquisition card. The data acquisition card converts analog signals into digital signals and transmits them to the computer memory through a high-speed bus for storage and processing.

[0047] (3) Data preprocessing: A Kalman filter is used to remove noise from the collected data. The state equation and observation equation of the Kalman filter are determined based on the dynamic model and measurement model of the motor. The recursive algorithm is used to estimate and correct the data in real time, effectively filtering out measurement noise and system noise. For current and voltage data, normalization processing is also required to make their value range between 0 and 1 to facilitate subsequent model calculations.

[0048] (4) Current and voltage prediction: In order to reduce the impact of sudden changes in the collected values ​​of operating parameters, a neural network-long short-term memory network (LSTM) is used to predict the current current and voltage of the motor. The LSTM network structure includes an input layer, a hidden layer (containing 50 LSTM units) and an output layer. The preprocessed normal operation data is divided into a training set (80%), a validation set (10%) and a test set (10%). The LSTM network is trained using the training set. The training algorithm uses the adaptive moment estimation (Adam) optimization algorithm. The initial learning rate is set to 0.001, β1=0.9, β2=0.999, and the number of training iterations is 2000. Through training, the LSTM network can capture the temporal dependencies and nonlinear characteristics in the motor operation data.

[0049] The training process of the LSTM network is as follows: the training set data is input into the LSTM network in time series order. The input data of each time step includes three-phase current, three-phase voltage and motor temperature. For the input data, necessary format conversion and dimension adjustment are performed to make it meet the input requirements of the LSTM network. At each time step, the LSTM unit generates a hidden state h according to the input data and the previous moment. t-1 and cell state c t-1 , calculate the hidden state h at the current moment t and cell state c t , and get the predicted value through the output layer. Calculate the loss function between the predicted value and the actual value, and calculate the gradient through back propagation of the Adam optimization algorithm to update the weight and bias parameters of the LSTM network. Repeat the above process until the set number of training iterations is reached or the loss function converges below a certain threshold.

[0050] The above physical model is combined with the trained LSTM network model. The physical model provides constraints on the basic operating principles of the motor, and the LSTM network model supplements the nonlinear and dynamic characteristics that are difficult to accurately describe with the physical model.

[0051] For current prediction, the d-axis voltage and q-axis voltage at the current moment are calculated based on the three-phase voltage collected at the current moment, and the d-axis voltage and q-axis voltage are input into the physical model, and the physical model calculates the theoretical current value based on the current moment. and The current theoretical current value The collected current operation data includes three-phase voltage, three-phase current and motor temperature, which constitute the input vector Among them, V a 、V b 、V c is the three-phase voltage of the motor, I a ,I b ,I cis the three-phase voltage of the motor, T is the temperature of the motor, and is input into the trained first LSTM network. The first LSTM network outputs the current prediction value at the current moment according to the input vector and the previously learned timing dependency relationship for the current and

[0052] For voltage prediction, the d-axis current and q-axis current at the current moment are calculated based on the three-phase current collected at the current moment, and the d-axis current and q-axis current are input into the physical model, and the physical model calculates the theoretical voltage value at the current moment. and The current theoretical voltage value The input vector is composed of the currently collected operating data including three-phase voltage, three-phase current and motor temperature. The second LSTM network outputs the current voltage prediction value based on the input vector and the previously learned timing dependency relationship for voltage. and

[0053] The prediction ability of the hybrid model under normal operation is evaluated by leave-one-out cross-validation. For each sample, it is used as a test sample, and the remaining samples are used as training samples for model training and prediction. The mean absolute error (MAE) between the predicted value and the actual value is calculated. The formula is:

[0054]

[0055] Among them, m is the number of samples, y i is the actual value of the i-th sample, is the predicted value of the i-th sample. If MAE is less than the set threshold, it is considered that the model prediction ability meets the requirements; otherwise, adjust the parameters of the LSTM network model (such as increasing the number of hidden layer units or adjusting the learning rate), retrain and verify until the requirements are met.

[0056] During the leave-one-out cross-validation process, the results of each training and prediction are recorded in detail, including the MAE value, the deviation distribution between the predicted value and the actual value, and other information. By analyzing these results, the generalization ability and stability of the model on different samples are evaluated. If it is found that some samples cause large model prediction errors, further analysis of the reasons may be caused by factors such as data anomalies and special motor operating conditions. In these cases, further data processing or adjustment of the model structure can be considered to improve the robustness of the model.

[0057] (5) Difference analysis: Input the measured data (current, voltage and other data after preprocessing) into the LSTM network and calculate the corresponding predicted value. and And the voltage prediction value and Calculate the difference between the predicted value and the measured value and define the difference index D:

[0058]

[0059] in, It represents the d-axis current and q-axis current obtained by converting the three-phase current collected at the kth moment. It represents the d-axis voltage and q-axis voltage obtained by converting the three-phase voltage collected at the kth moment. They represent the predicted values ​​of d-axis current, q-axis current, d-axis voltage, and q-axis voltage output by the anomaly detection model at the kth moment, respectively, and n represents the total number of the nearest historical sampling points. In order to avoid the excessive influence of individual abnormal data points on the difference index, the measured values ​​and predicted values ​​are detected and processed before calculating the difference. The 3-times standard deviation method is used to identify the abnormal values ​​and replace them with the average value of the adjacent normal data points.

[0060] The collected operation data is subjected to abnormality detection. The abnormal value detection process is as follows: First, the i-th operation data x of the motor is obtained. i The mean μ i and standard deviation σ i , for the currently collected running data x i , judge |x i -μ i |>3σ i Is it true? If so, it is considered that the currently collected running data x i For outliers, replace the outliers with the average of the adjacent normal values. When calculating the difference index D, the processed measured values ​​and predicted values ​​are used for calculation to improve the reliability of the difference index. i It is three-phase voltage, three-phase current and motor temperature.

[0061] (6) Threshold update: Based on a large amount of historical data (covering data under normal operation and various fault conditions of the motor, with a data volume of no less than 5,000 sets) and experimental results, an empirical threshold θ0 is preliminarily set.

[0062] The specific operations are as follows;

[0063] First, we conduct a statistical analysis on the difference index D in the historical data and calculate its mean μ and standard deviation σ. We assume that the difference index D follows a normal distribution N(μ, σ 2). In order to ensure the rationality of the normal distribution assumption, the Shapiro-Wilk normality test method was used to test the data. If the test results showed that the data did not obey the normal distribution, the data were logarithmically transformed or square root transformed to make it approximately obey the normal distribution before subsequent calculations.

[0064] Then, the parameters μ and σ of the normal distribution are estimated by the maximum likelihood estimation method. For the difference indicator D sequence {D1, D2, …, D m} (m is the number of data points), construct the probability density function f(D) of the normal distribution:

[0065]

[0066] According to the maximum likelihood estimation principle, construct the likelihood function L(μ,σ):

[0067]

[0068] Taking the logarithm of the likelihood function yields:

[0069]

[0070] Taking partial derivatives of μ and σ and setting them equal to 0, we get the following system of equations:

[0071]

[0072] Solve this system of equations to get the estimated values ​​of μ and σ. Take the initial value of the threshold θ0 = μ + 3σ.

[0073] In this way, under normal operation, the probability that the difference index exceeds the threshold is about 0.3%, which effectively reduces the false alarm rate during normal operation. At the same time, in order to further improve the accuracy of the initial threshold setting, the cross-validation method can also be used to divide the historical data into a training set and a validation set, calculate the initial threshold on the training set, and then verify its effectiveness on the validation set, and fine-tune the threshold according to the verification results. The adaptive threshold adjustment formula is as follows:

[0074] θ b =θ n +α[D n -θ n ];

[0075] Among them, θ n is the threshold formed by the last threshold update, and is also the threshold before this update. n is the difference index updated based on the currently collected operating data, θ bThis is the threshold formed after this update, α is the learning rate (the value range is between 0.05-0.2, and the initial setting is 0.1), which controls the speed of threshold update.

[0076] After a certain time interval (30 seconds) or a certain amount of new data (500 new data points) is collected, the difference index is recalculated and the threshold is updated according to the above formula. In the process of updating the threshold, in order to prevent the threshold from being adjusted too frequently or fluctuating too much, a threshold adjustment step limit is set (each adjustment does not exceed 10% of the previous threshold). At the same time, the forgetting factor β (with a value range of 0.8-0.99) is introduced to weight the impact of historical data so that the threshold can better reflect the current system operation status. The threshold update formula is as follows:

[0077] θ b =βθ n +(1-β)α[D n -θ n ];

[0078] In order to ensure the effectiveness of dynamic threshold adjustment, in actual applications, the fault detection effect after threshold adjustment can be monitored in real time, such as changes in false alarm rate and missed alarm rate. If the false alarm rate or missed alarm rate is found to increase significantly, the threshold adjustment is suspended and the cause is analyzed. It may be that the motor operating conditions have changed significantly (such as sudden load changes, sudden changes in ambient temperature, etc.). At this time, the threshold can be manually adjusted or the learning rate α and forgetting factor β can be adjusted according to the specific situation, and then the dynamic threshold adjustment can be continued.

[0079] (7) Abnormality judgment: Primary threshold θ1 (value is 0.7θ b ), when D>θ1, it is considered that the motor may be abnormal. In order to avoid misjudgment due to fluctuations in individual data points, when judging that the motor is abnormal, it is required that not only the difference index D at the current moment is greater than the primary threshold θ1, but also the number of times that the difference index D corresponding to the past 10 sampling points is greater than the primary threshold θ1 exceeds 60%.

[0080] Intermediate threshold θ2 (valued at 1.1θ b ), when D>θ2, the suspicion of fault is further increased. When judging whether the motor is abnormal, the change of the difference index in the past period of time is also considered. It is required that the number of times the difference index D corresponding to the past 20 sampling points is greater than the intermediate threshold θ2 exceeds 70%. The 20 sampling points are grouped, and the average value of the difference index of each group is required to show an upward trend.

[0081] Advanced threshold θ3 (valued at 1.3θ b)When D>θ3, it is basically determined that the motor is faulty. In order to avoid misjudgment due to fluctuations in individual data points, when judging that the motor is abnormal, it is required that in the past 30 sampling points, the number of times that the difference index D is greater than the advanced threshold θ3 exceeds 80%, and the average value of the difference index continues to rise and exceeds a set growth threshold. The 30 sampling points are grouped, and the average value of the difference index of the next group is required to be greater than 0.1 from the average value of the difference index of the previous group.

[0082] The duration judgment of each level of threshold is based on the understanding of the development process of motor faults. Motor faults usually do not occur instantly, but have a development process. By setting the duration threshold (primary threshold duration exceeds 10 seconds, intermediate threshold duration exceeds 5 seconds, and advanced threshold duration exceeds 2 seconds), misjudgment caused by short-term interference signals can be avoided. At the same time, in actual applications, the values ​​of each level of threshold and the duration threshold can be adjusted and optimized according to different motor types and operating conditions to improve the accuracy and reliability of fault diagnosis.

[0083] (8) Fault detection: When a motor abnormality is detected, the fault detection process is started, and the difference index D and related measured data (such as current, voltage, etc.) are input into various fault prediction models. Before inputting the data, feature engineering is performed on the data to extract statistical features and frequency domain features such as the mean, variance, and spectrum characteristics of current and voltage as the input feature vector of the fault state model to improve the accuracy of fault classification.

[0084] A fault prediction model is built for each type of fault. The fault prediction model is built using a convolutional neural network (CNN). Fault types include rotor eccentricity, stator winding short circuit, permanent magnet demagnetization, bearing fault, and power supply fault. All fault prediction models are placed in the fault model library. The number of training samples for each fault type is not less than 1,000) for training. The CNN network structure includes convolutional layers, pooling layers, and fully connected layers. The convolution kernel size is set to 3×3, and the pooling layer uses the maximum pooling operation with a step size of 2.

[0085] The local features and abstract features of the input feature vector are extracted through the convolution layer and the pooling layer, and then the fault classification is performed through the fully connected layer. The cross entropy loss function and the stochastic gradient descent (SGD) optimization algorithm are used in the training process. The learning rate is initially set to 0.01, the momentum factor is 0.9, and the number of training iterations is 5000. The input feature vector is passed through the trained CNN classifier to determine the fault type of the motor and output the fault diagnosis result.

[0086] When training the CNN model, data enhancement technology is used to increase the number of training samples and improve the generalization ability of the model. At the same time, in order to prevent overfitting, an L2 regularization term (regularization coefficient λ = 0.01) is added to the fully connected layer to constrain the weight of the model. In practical applications, the model is regularly updated and optimized using newly collected fault data to adapt to changes in the motor's operating status and the emergence of new fault types.

[0087] The permanent magnet synchronous motor fault diagnosis method provided by the present invention has the following beneficial technical effects:

[0088] (1) Improving diagnostic accuracy: The current and voltage prediction model provided by the present invention comprehensively considers the physical characteristics of the motor and the nonlinear relationship in the actual operating data. Compared with the traditional method, it is more accurate in judging the motor abnormality. It adopts a detailed difference index calculation method, which fully considers the difference between the current and voltage on the d-axis and q-axis, and can more comprehensively reflect the degree of deviation between the motor operating state and the model prediction, thereby improving the detection accuracy of the motor abnormality.

[0089] (2) Reduce false alarm rate: The combination of motor physical model and machine learning model (such as long short-term memory network LSTM) avoids the misjudgment caused by the limitations of a single model. The physical model is based on the basic principles of the motor and has basic constraints on the normal operating state; the machine learning model can supplement the situation that is difficult for the physical model to handle by learning the complex relationships in a large amount of data. The two work together to effectively reduce the false alarm caused by a single factor.

[0090] (3) Improve response speed: Real-time monitoring of motor operating parameters, short data collection interval (data collection every 5 milliseconds), can timely obtain the latest operating status of the motor. Once the difference between the measured data and the model prediction value exceeds the abnormal judgment standard, fault detection is immediately started, and the fault can be quickly diagnosed in a short time.

[0091] The present invention has been described exemplarily. Obviously, the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A permanent magnet synchronous motor fault diagnosis method, characterized in that: The method comprises the following steps: (1) Convert the currently collected three-phase current and three-phase voltage into the current and voltage on the d-axis and q-axis; (2) Based on the current and voltage on the d-axis and q-axis, the theoretical voltage values ​​on the d-axis and q-axis are calculated respectively. and Theoretical current value and (3) Based on the theoretical voltage value at the current moment and and operating parameters to obtain the voltage forecast value at the current moment and Based on the theoretical current value at the current moment and and operating parameters to obtain the current prediction value at the current moment and (4) Anomaly detection is performed based on the difference between the voltage prediction value and the current prediction value and the corresponding collected value. When an abnormality is detected in the motor, the fault type of the motor is determined.

2. The permanent magnet synchronous motor fault diagnosis method according to claim 1, characterized in that: The current theoretical current value The current operating data includes three-phase voltage, three-phase current and motor temperature T, which form an input vector and are input into the trained first LSTM network. The first LSTM network outputs the current prediction value at the current moment. and 3. The permanent magnet synchronous motor fault diagnosis method according to claim 1, characterized in that: The current theoretical voltage value and The current operating data includes three-phase voltage, three-phase current and motor temperature T, which form an input vector and are input into the trained second LSTM network. The second LSTM network outputs the voltage prediction value at the current moment. and 4. The permanent magnet synchronous motor fault diagnosis method according to claim 2 or 3, characterized in that: The threshold update formula of the first LSTM network or the second LSTM network is as follows: i b =βθ n +(1-β)α[D n -θ n ] Among them, θ n is the threshold formed by the last threshold update, and is also the threshold before this update. n is the difference index updated based on the running data collected at the current time interval, θ b This is the threshold formed after this update, α is the learning rate, and β is the forgetting factor.

5. The permanent magnet synchronous motor fault diagnosis method according to claim 4, characterized in that: The process of determining the initial value of the threshold is as follows: Based on the difference indicator sequence {D1, D2, ..., D m }Construct the normal distribution probability density function f(D) of the difference index D: According to the principle of maximum likelihood estimation, construct the likelihood function L(μ,σ), and determine the estimated values ​​of μ and σ based on the likelihood function L(μ,σ). The initial value of the threshold is θ0=μ+3σ, where μ and σ are parameters of the normal distribution.

6. The permanent magnet synchronous motor fault diagnosis method according to claim 1, characterized in that: Perform anomaly detection on the collected operating data. The anomaly detection process is as follows: Determine the i-th running data x based on historical data i The mean μ i and standard deviation σ i , for the currently collected running data x i , judge |x i -μ i |>3σ i Is it true? If so, the currently collected running data x i For outliers, replace them with the average of their neighboring normal values.

7. The permanent magnet synchronous motor fault diagnosis method according to claim 6, characterized in that: The difference index D corresponding to the current running data is calculated based on the following formula: in, It represents the d-axis current and q-axis current obtained by converting the three-phase current collected at the kth moment. represents the d-axis voltage and q-axis voltage obtained by converting the three-phase voltage collected at the kth moment, They respectively represent the predicted values ​​of d-axis current, q-axis current, d-axis voltage, and q-axis voltage output by the abnormality detection model at the kth moment, and n represents the total number of the nearest historical sampling points.

8. The permanent magnet synchronous motor fault diagnosis method according to claim 7, characterized in that: The specific process of determining the abnormality of the motor is as follows: When the current difference index D>θ1, and the number of times that the difference index D corresponding to the past 10 sampling points is greater than the primary threshold θ1 exceeds 60%, it is determined that the motor is abnormal, where θ1 is set to 0.7 times the current threshold; When the current difference index D>θ2, and the number of times that the difference index D corresponding to the past 20 sampling points is greater than the intermediate threshold θ2 exceeds 70%, and the average value of the difference index shows an increasing trend, θ2 is set to 1.1 times the current threshold; When the current difference index D>θ3, and in the past 30 sampling points, the number of times the difference index D is greater than the advanced threshold θ3 exceeds 80%, and the average value of the difference index continues to increase and the increase exceeds a set growth threshold, where θ3 is set to 1.3 times the current threshold.

9. The permanent magnet synchronous motor fault diagnosis method according to claim 1, characterized in that: After being determined to be abnormal, the mean, variance and spectral characteristics of the current current and voltage of the motor are input into the fault prediction models of various types of faults, and the current fault type of the motor is determined based on the prediction results output by each fault prediction model.

10. The permanent magnet synchronous motor fault diagnosis method according to claim 9, characterized in that: A fault prediction model is constructed for each type of fault. The fault prediction model is constructed using a convolutional neural network. The fault types include rotor eccentricity, stator winding short circuit, permanent magnet demagnetization, bearing failure, and power supply failure. All fault prediction models are placed in the fault prediction model library.

Images