A method and device for diagnosing faults of inverter motor control

Abstract

The invention relates to the field of motor control technology, and discloses a method and device for diagnosing faults in frequency converter motor control, wherein the method comprises: collecting stator three-phase current signals, inverter output voltage signals and speed signals of a frequency converter drive motor system, and establishing a state space model; constructing an energy function coefficient matrix, and obtaining training energy function features; performing Clarke transformation on the stator three-phase current signals, and performing difference to obtain training current residual vectors; performing sample entropy calculation to obtain entropy value feature sequences, and training a deep neural network according to the entropy value feature sequences to obtain a fault classification model; inputting real-time energy function features and real-time current residual vectors into the fault classification model, and performing fault identification in combination with the distribution characteristics of the real-time current residual vectors in an αβ plane, and obtaining fault type discrimination results and fault location information, wherein the invention realizes accurate discrimination of fault types and accurate positioning of locations.

Description

Technical Field

[0001] This invention relates to the field of motor control technology, and in particular to a method and device for diagnosing motor control faults in frequency converters. Background Art

[0002] Variable frequency drive (VFD) motor systems are widely used in industrial automation, and their control performance and operational reliability directly affect production efficiency and product quality. However, due to their complex structure and variable operating conditions, VFD motor systems are prone to various control faults, such as sensor failures, switching device failures, and parameter deviations. These faults can lead to system performance degradation or even shutdown.

[0003] Currently, common fault diagnosis methods are mainly based on deterministic models or observer designs. These methods require the establishment of accurate mathematical models and have high computational complexity, making them difficult to meet the requirements of real-time diagnosis. Furthermore, traditional fault diagnosis methods often focus only on single features, such as current harmonics or voltage fluctuations, lacking consideration of the overall dynamic characteristics of the system, resulting in insufficient reliability of diagnostic results. In addition, existing fault diagnosis methods generally suffer from insufficient feature extraction and inaccurate fault location. Especially when system parameters and operating conditions change, diagnostic performance significantly degrades. Summary of the Invention

[0004] This invention provides a method and device for diagnosing motor control faults in frequency converters. This invention achieves accurate identification of fault types and accurate location of faults.

[0005] In a first aspect, the present invention provides a method for diagnosing motor control faults in a frequency converter, the method comprising: Collect the stator three-phase current signal, inverter output voltage signal and speed signal of the inverter drive motor system, and establish a state-space model including the stator current dynamic equation, rotor flux dynamic equation and speed dynamic equation. A quadratic Lyapunov function is established for the state variables in the state-space model, and an energy function coefficient matrix is ​​constructed based on the derivative terms of the state variables to obtain the training energy function characteristics of the inverter drive motor system. The stator three-phase current signal is subjected to Clarke transformation to obtain the current components in the αβ coordinate system, and the predicted current value is calculated according to the state space model. The difference between the predicted current value and the measured current value is used to obtain the training current residual vector. The sample entropy is calculated on the training energy function features and the training current residual vector to obtain an entropy feature sequence, and a deep neural network is trained based on the entropy feature sequence to obtain a fault classification model; The real-time energy function features and real-time current residual vector are input into the fault classification model. Fault identification is performed by combining the distribution characteristics of the real-time current residual vector in the αβ plane, and the fault type discrimination result and fault location information are obtained.

[0006] Secondly, the present invention provides a frequency converter motor control fault diagnosis device, the frequency converter motor control fault diagnosis device comprising: The acquisition module is used to acquire the stator three-phase current signal, inverter output voltage signal and speed signal of the inverter drive motor system, and to establish a state-space model including the stator current dynamic equation, rotor flux linkage dynamic equation and speed dynamic equation. A module is established to establish quadratic Lyapunov functions for the state variables in the state space model, and to construct an energy function coefficient matrix based on the derivative terms of the state variables, thereby obtaining the training energy function characteristics of the inverter drive motor system. The calculation module is used to perform Clarke transform on the stator three-phase current signal to obtain the current components in the αβ coordinate system, and calculate the predicted current value according to the state space model. The difference between the predicted current value and the measured current value is used to obtain the training current residual vector. The training module is used to calculate the sample entropy of the training energy function features and the training current residual vector to obtain an entropy feature sequence, and to train a deep neural network based on the entropy feature sequence to obtain a fault classification model. The fault identification module is used to input the real-time energy function features and the real-time current residual vector into the fault classification model, and to identify the fault by combining the distribution characteristics of the real-time current residual vector in the αβ plane, thereby obtaining the fault type discrimination result and fault location information.

[0007] A third aspect of the present invention provides a frequency converter motor control fault diagnosis device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the frequency converter motor control fault diagnosis device to execute the above-described frequency converter motor control fault diagnosis method.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described inverter motor control fault diagnosis method.

[0009] The technical solution provided by this invention establishes a state-space model including stator current, rotor flux linkage, and rotational speed, and analyzes it using the Lyapunov energy function, comprehensively considering the dynamic characteristics of the system and improving the accuracy of fault diagnosis. It employs Clark transform to convert the three-phase current to the αβ coordinate system and uses the state-space model for current prediction. By calculating the current residual vector, it effectively captures the abnormal characteristics of the system caused by the fault. It introduces a sample entropy calculation method to analyze the entropy values ​​of the energy function characteristics and the current residual vector, extracting more discriminative fault features. A deep neural network with a multilayer perceptron structure is designed, and the feature learning ability of the model is enhanced through multilayer nonlinear transformation and batch normalization. Fault verification is performed by combining the distribution characteristics of the current residual vector in the αβ plane. Through comprehensive analysis of fault direction and amplitude characteristics, accurate fault type identification and precise location are achieved.

[0010] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or understood by practicing the present invention. The objectives and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0011] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of one embodiment of the inverter motor control fault diagnosis method of the present invention; Figure 2 This is a schematic diagram of one embodiment of the inverter motor control fault diagnosis device of the present invention; Figure 3 This is a schematic diagram of one embodiment of the inverter motor control fault diagnosis device in this invention. DETAILED DESCRIPTION

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0015] To facilitate understanding of this embodiment, a detailed description of a frequency converter motor control fault diagnosis method disclosed in this embodiment of the invention will be provided first. For example... Figure 1 As shown, this method includes the following steps: 101. Collect the stator three-phase current signal, inverter output voltage signal and speed signal of the frequency converter drive motor system, and establish a state-space model including the stator current dynamic equation, rotor flux dynamic equation and speed dynamic equation. It is understood that the executing entity of this invention can be a frequency converter motor control fault diagnosis device, a terminal, or a server; no specific limitation is made here. This embodiment of the invention will be described using a server as an example.

[0016] Specifically, three-phase current signals of the inverter-driven motor system are acquired using current sensors. The raw current data acquired by the current sensors contains high-frequency noise and aliasing effects. Anti-aliasing filtering is applied to remove high-frequency noise components, yielding the stator three-phase current signals. Simultaneously, the DC bus voltage and inverter output voltage of the inverter-driven motor system are sampled using voltage sensors to obtain raw voltage data. Due to the operating characteristics of the inverter, the sampling process is affected by dead-time effects, leading to errors in the measured voltage signals. To compensate for this error, the effective value of each phase output voltage is calculated based on the conduction time of the inverter switching devices, resulting in an inverter output voltage signal closer to the true value. Simultaneously, the rotor speed of the motor is acquired using a speed sensor. The speed signal is affected by measurement noise; therefore, mean filtering is applied to the raw speed data. Mean filtering reduces the influence of random noise, making the filtered speed signal more stable and more consistent with the actual operating state of the motor. A dynamic equation for the stator current is established based on the stator three-phase current signals, the inverter output voltage signal, and the speed signals. The equation is in the form of the relationship between the derivative of the stator current and the stator voltage, stator resistance, and stator inductance. In this process, the stator resistance and stator inductance are used as equation parameters, and differential equations reflecting the stator current variation characteristics are constructed using collected current and voltage data. The stator current dynamic equation is extended, and the rotor flux linkage dynamic equation is established by combining the physical definition and relationships of flux linkage. The change in rotor flux linkage is not only related to the stator current but also affected by the rotor inductance and mutual inductance parameters. Accurate calculation of these parameters is necessary during modeling to ensure that the flux linkage dynamic equation can truly reflect the dynamic change behavior of the rotor flux linkage. This step further refines the dynamic description of the electromagnetic part in the state-space model. Based on the stator current dynamic equation and the rotor flux linkage dynamic equation, and combined with the rotational inertia parameters of the mechanical system, the speed dynamic equation is established. The speed dynamic equation describes the state of rotor mechanical motion, and its main variables include the rate of change of torque and speed. When establishing this equation, the interaction between electromagnetic torque and load torque is considered simultaneously to fully describe the dynamic behavior of the motor system. The stator current dynamic equation, rotor flux linkage dynamic equation, and speed dynamic equation are combined into a single state-space model. This model uses state vectors to uniformly describe the electrical and mechanical dynamic behavior of the system. To solve this complex set of nonlinear equations, a fourth-order Runge-Kutta method is employed for numerical computation. The fourth-order Runge-Kutta method, with its high accuracy and stability, can effectively discretize the state equations, thus accurately calculating the evolution of the system state over time in the time domain. Through this process, a state-space model is established, comprising the stator current dynamic equation, the rotor flux linkage dynamic equation, and the rotational speed dynamic equation.

[0017] 102. Establish quadratic Lyapunov functions for the state variables in the state-space model, and construct the energy function coefficient matrix based on the derivative terms of the state variables to obtain the training energy function characteristics of the inverter drive motor system. Specifically, a coordinate transformation is performed on the state-space model. By acquiring the stator three-phase current and rotational speed signals, a Park transformation matrix is ​​constructed to map the current and flux linkage signals from the three-phase stationary coordinate system to the dq rotating coordinate system. This aligns the stator current variation with the rotating magnetic field, providing a more intuitive description of its dynamic characteristics. After the coordinate transformation, the state variables in the dq rotating coordinate system are decomposed into stator current components, rotor flux linkage components, and mechanical angular velocity components, which together constitute the state variable set. Based on the information in the state variable set, a quadratic function of the stator current error term is constructed. The current energy term is obtained by summing the squares of the d-axis and q-axis current errors. This energy term reflects the dynamic characteristics of the current error and can quantify the deviation between the actual and expected current values. Simultaneously, a quadratic function of the flux linkage error term is constructed based on the state variable set. The flux linkage energy term is obtained by calculating the sum of the squares of the d-axis and q-axis flux linkage errors. This energy term characterizes the degree of dynamic flux linkage error, thus describing the deviation characteristics of the rotor flux linkage. A weighted combination of the current energy term and the flux linkage energy term is used to construct a complete combined energy function. The current energy term and the flux linkage energy term are multiplied by their respective weighting coefficients. The current weighting coefficient is set to the reciprocal of the motor's rated current, and the flux linkage weighting coefficient is set to the reciprocal of the rated flux linkage. This weighting design ensures a reasonable balance between the dimensions and influence of the current and flux linkage energy terms, thereby ensuring that the combined energy function accurately reflects the overall dynamic characteristics of the system. Based on the state variable set, the derivative expression of the combined energy function is calculated to obtain the energy change rate equation. By differentiating the combined energy function with respect to time, the dynamic characteristic equation of energy change is obtained, reflecting the trend of energy function change over time. To ensure system stability, a coefficient matrix is ​​designed for the energy change rate equation. By constructing a positive definite matrix, the rate of change of the energy function is always less than zero. The introduction of the positive definite matrix ensures that the system satisfies the Lyapunov stability condition under normal operating conditions, forming a complete Lyapunov stability criterion. Based on the above Lyapunov stability criterion, the combined energy function is calculated online. During normal system operation, the energy function value is calculated in real time and its changing characteristics are recorded to form the energy function features for the training phase. This process captures the dynamic energy characteristics of the system during normal operation, forming baseline data and providing a reference standard for fault diagnosis. By recording the energy function values ​​under normal operating conditions, the training energy function features of the inverter drive motor system are finally extracted.

[0018] 103. Perform Clarke transform on the stator three-phase current signal to obtain the current components in the αβ coordinate system, and calculate the predicted current value according to the state space model. Subtract the predicted current value from the measured current value to obtain the training current residual vector. Specifically, the stator three-phase current signal is normalized, mapping the actual operating current signal of the motor to a standardized unit interval, thereby reducing the impact of different dimensions and numerical ranges on the calculation. Using the motor's rated current value as the normalization reference, the three-phase current signal is divided by the rated current value to obtain the normalized current signal. By establishing a Clarke transform matrix, the normalized three-phase stationary coordinate system current signal is converted into current components in a two-phase orthogonal coordinate system. After the conversion, the originally complex three-phase system is simplified to two current components in the αβ stationary coordinate system. A current prediction model is constructed based on the state-space model. Combining the voltage equation and flux linkage equation, the prediction model of the current components is derived by analyzing the voltage balance relationship and the dynamic coupling characteristics of the flux linkage. The prediction model aims to describe the current variation law of the system and provide a theoretical basis for subsequent error analysis. To adapt to the operating characteristics of the digital control system, the prediction model is transformed from the continuous time domain to the discrete time domain. This transformation is achieved through an approximation method, converting the original continuous state equation into a discrete time-stepping model, forming a discrete prediction model that can be calculated in each sampling period. The α-axis and β-axis current prediction values ​​for the next sampling period are calculated using a discrete prediction model. These prediction values ​​form a prediction vector representing the theoretical operating state of the system. Simultaneously, within each sampling period, the actual current components in the αβ coordinate system are sampled, and the current true current value is recorded as a measured vector. To analyze the system's operating state, the difference between the predicted vector and the measured vector is calculated. By comparing components one by one, the current residuals in the α-axis and β-axis are obtained. These residuals reflect the deviation between the actual current and the predicted current and are important indicators of whether there are any abnormalities in the system's operation. The α-axis residuals and β-axis residuals are combined into a complete training current residual vector. This residual vector contains all the information about the current error in the αβ coordinate system and can keenly capture potential fault characteristics in the motor system.

[0019] 104. Calculate the sample entropy of the training energy function features and the training current residual vector to obtain the entropy value feature sequence, and train a deep neural network based on the entropy value feature sequence to obtain a fault classification model; Specifically, a sliding time window is set for the training energy function features and the training current residual vector to capture the dynamic changes in the data. Within the sliding time window, the time series data is reconstructed into m-dimensional vectors. This process is achieved by embedding dimension m and time delay τ, aiming to segment the original one-dimensional time series data into a sequence of reconstructed vectors in a high-dimensional space. After vector reconstruction, the similarity between vectors in the reconstructed vector sequence is calculated. By setting a similarity tolerance r, Euclidean distance is used to measure the similarity between two reconstructed vectors, and the number of reconstructed vectors similar to the template vector is counted. For a given template vector, similar reconstructed vectors are those whose Euclidean distance is less than the tolerance r. Counting these vectors reflects the repetition pattern of the data sequence. This statistical process provides a basis for similar vector counting to calculate sample entropy. Conditional probabilities are calculated based on the similar vector counts. The conditional probability is obtained by dividing the number of similar vectors by the total number of reconstructed vectors. Its calculation needs to be performed separately for template lengths of m and m+1. This yields two sets of conditional probability values, reflecting the similarity of the data sequences under different template lengths. Logarithmic operations are performed on the two sets of conditional probability values, and their average is calculated. The sample entropy value is obtained by subtracting the average log-conditional probability of template length m from the average log-conditional probability of template length m+1. Sample entropy is an effective measure of complexity, quantitatively reflecting the dynamic characteristics and disorder of a time series. As the time window slides, the sample entropy values ​​of the training energy function features and the training current residual vector are calculated window by window. All calculated sample entropy values ​​are arranged in chronological order to form an entropy feature sequence, reflecting the dynamic changes of the training data in the time domain. After the entropy feature sequence is extracted, a deep neural network with a multilayer perceptron structure is constructed for training the fault classification model. This deep neural network includes an input layer, three hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the entropy feature sequence. The number of neurons in the three hidden layers are set to 128, 64, and 32, respectively, to achieve stepwise feature extraction and dimensionality reduction. The number of neurons in the output layer is equal to the number of fault types, used to output the probability distribution of faults. The ReLU activation function is used in the hidden layers for nonlinear transformation to enhance the network's expressive power. Batch normalization is applied to the output of each layer to accelerate convergence and prevent gradient vanishing. The output layer uses the Softmax activation function to transform the network output into a probability distribution of fault types, providing a direct basis for fault classification. After inputting the entropy feature sequence into the deep neural network, the gradient of the network parameters is calculated using the backpropagation algorithm, and the Adam optimizer is used to optimize and update the network parameters. The Adam optimizer combines the advantages of momentum and adaptive learning rate, quickly converging to the optimal solution without requiring manual adjustment of the learning rate.Through multiple iterations of training, the network gradually adjusts its parameters, enabling the model to accurately classify different fault types, ultimately resulting in a fault classification model.

[0020] 105. Input the real-time energy function features and real-time current residual vector into the fault classification model, and combine the distribution characteristics of the real-time current residual vector in the αβ plane to identify the fault, and obtain the fault type discrimination result and fault location information.

[0021] Specifically, real-time operating data of the inverter drive motor system is collected and processed. The collected data includes stator current signals, inverter output voltage signals, and speed signals. The energy function value at the current moment is calculated using a state-space model to generate real-time energy function features, characterizing the dynamic stability of the system during operation. Sample entropy is calculated on the real-time energy function features. By setting the embedding dimension *m* and similarity tolerance *r*, a sliding window is applied to the real-time energy function features to extract the complexity information of the sequence. Sample entropy calculation reflects the dynamic changes of the real-time energy function and generates real-time entropy value features, further characterizing the system's operating state. These real-time entropy value features serve as input to a deep neural network, effectively improving the diagnostic capability of the fault classification model. Simultaneously, the orientation angle of the real-time current residual vector in the αβ coordinate system is calculated. The spatial angle of the current residual vector in the αβ plane is determined by the ratio of the α-axis component to the β-axis component, obtaining the fault orientation feature. This reflects the directional distribution of the current residual and has a specific correspondence with different types of faults, providing a basis for fault type verification. Simultaneously with the calculation of the orientation feature, the amplitude of the real-time current residual vector is calculated. The magnitude of the residual vector is calculated by taking the square root of the sum of the squares of the α-axis and β-axis components, representing the absolute size of the current residual. This magnitude is then compared with a preset threshold to generate fault amplitude features. Changes in these features reflect the severity of the fault and are crucial for determining the fault level. The calculated real-time entropy features are input into a deep neural network, and a fault classification model is used for forward propagation. The deep neural network learns from the fault features during the training phase, generating a probability distribution for fault types. The output with the highest probability value is selected to initially determine the fault type. To improve the accuracy of the diagnosis, fault direction features are used to verify the initial fault type. These features, through the correspondence between the directionality of the current residual vector and the fault type, confirm the rationality of the classification model's output. If the direction features match the initial classification result, the validity of the classification result is confirmed; otherwise, it needs to be reassessed or labeled as an undetermined fault type. After fault type identification, the severity of the fault is assessed based on the fault amplitude features. The severity level is determined by comparing the residual amplitude with a preset threshold. Different types of faults have specific amplitude distribution characteristics, which are used to refine the fault level classification. For example, in the case of a current overload fault, a larger amplitude residual indicates a severe overload, while a smaller amplitude indicates a minor anomaly. Based on the fault type classification and severity level, combined with fault direction characteristics, the specific location of the fault is determined. The fault location information is determined by mapping the classification results to the equipment layout or operating parameters of the physical system. For example, a specific current direction anomaly points to a fault in a particular phase winding, while a specific amplitude range corresponds to a specific switch fault in the inverter.

[0022] The fault direction characteristics are divided into regions. Using the αβ plane, the complete 360-degree spatial angle is divided into 12 equally spaced sectors, each covering a range of 30 degrees. This division ensures the accuracy and ease of analysis of the direction characteristics. A fault region mapping table is established through this region division to map fault direction characteristics to specific directional ranges within the αβ plane. After completing the region division of the αβ plane, a characteristic direction database is established based on theoretical analysis. By analyzing the electrical characteristics and operating states of different fault types, the specific directional distribution patterns of each fault type within the αβ plane are determined. These directional distribution patterns are represented by angle intervals and stored in the database as directional characteristic templates for each fault type. The establishment of directional characteristic templates is the core step in fault direction verification, providing accurate reference for the initial fault type direction verification by consulting theoretical direction intervals. Based on the initial fault type, the directional characteristic templates are looked up in the table to obtain the theoretical directional range corresponding to that type of fault, i.e., the fault direction verification interval. The verification interval reflects the directional range of the current residual vector appearing in the αβ plane for this type of fault. The actual collected fault direction features are compared with the fault direction verification interval for interval judgment. The direction matching degree is calculated by comparing whether the actual direction falls within the verification interval. The direction matching degree is calculated based on the degree of overlap between the actual direction and the theoretical direction. When the actual direction falls completely within the verification interval, the matching degree is 100%; when the actual direction deviates from the verification interval, the matching degree decreases according to the degree of deviation. After the direction matching degree is calculated, a threshold judgment is performed to determine whether the direction features conform to the theoretical analysis results. By setting a predefined allowable range for direction deviation, a certain degree of direction deviation is tolerated, but if the deviation exceeds the range, the direction verification is considered a failure. Based on this judgment, the direction verification result is generated. When the direction verification passes, the current preliminary fault type is confirmed as the final fault type; however, if the direction verification fails, the preliminary fault type needs to be adjusted. To adjust the preliminary fault type, the fault type corresponding to the second-best probability is selected for re-verification. The selection of the second-best probability is based on the fault type probability distribution output by the fault classification model, selecting a candidate type that is second only to the preliminary fault type. The candidate fault type is re-verified in the direction, and its direction matching degree and verification result are determined by using the same direction feature comparison method. After direction verification, confidence analysis is performed on all candidate fault types. The confidence analysis combines the probability value output by the fault classification model and the direction matching degree, and calculates the comprehensive confidence degree for each fault type through weighted calculation. During the weighted calculation, the probability value and direction matching degree of the fault classification model each have a certain weight to balance the model classification result with the physical meaning of the direction features. By comparing the confidence degrees of all candidate types, the fault type with the highest confidence degree is selected as the final fault type discrimination result.

[0023] In this embodiment of the invention, a state-space model incorporating stator current, rotor flux linkage, and rotational speed is established and analyzed using the Lyapunov energy function. This comprehensively considers the dynamic characteristics of the system and improves the accuracy of fault diagnosis. The Clarke transform is used to convert the three-phase current to the αβ coordinate system, and the state-space model is used for current prediction. By calculating the current residual vector, the abnormal characteristics of the system caused by the fault are effectively captured. A sample entropy calculation method is introduced to analyze the entropy values ​​of the energy function features and the current residual vector, extracting more discriminative fault features. A deep neural network with a multilayer perceptron structure is designed, and the feature learning capability of the model is enhanced through multilayer nonlinear transformation and batch normalization. Fault verification is performed by combining the distribution characteristics of the current residual vector in the αβ plane. Through comprehensive analysis of fault direction and amplitude features, accurate fault type identification and precise location are achieved.

[0024] In one specific embodiment, the process of performing step 101 may specifically include the following steps: A current sensor is used to acquire the current signal of the inverter drive motor system to obtain the raw current data. The raw current data is then subjected to anti-aliasing filtering to obtain the stator three-phase current signal. Voltage sensors are used to sample the DC bus and inverter output of the inverter drive motor system to obtain raw voltage data. Dead-time compensation is performed on the raw voltage data, and the effective voltage value is calculated based on the conduction time of the switching devices to obtain the inverter output voltage signal. A speed sensor is used to collect the rotor speed of the inverter drive motor system, and the mean value is filtered to obtain the speed signal; The stator current dynamic equation is established based on the stator three-phase current signal, the inverter output voltage signal and the speed signal. The stator resistance and stator inductance are used as equation parameters to obtain the stator current dynamic equation. The dynamic equation of stator current is analyzed for flux linkage relationship, and the dynamic equation of rotor flux linkage is established based on rotor inductance and mutual inductance parameters. Based on the dynamic equations of stator current and rotor flux linkage, and combined with the moment of inertia parameter, the dynamic equation of rotational speed is established. The fourth-order Runge-Kutta method is used to solve the state equations to obtain the state-space model.

[0025] Specifically, current sensors are used to collect current signals from the inverter-driven motor system to obtain raw current data. This data includes the three-phase current signals of the motor stator. , and Due to high-frequency noise interference during signal acquisition, anti-aliasing filtering is applied to the raw data to eliminate high-frequency components. A low-pass filter is used in the filtering process, and its transfer function is expressed as: ; in It is a variable in the complex frequency domain. This is the cutoff frequency of the filter. The filtered signal is the stator three-phase current signal. These signals will serve as inputs to the motor's electrical dynamic equations. Simultaneously, a voltage sensor will be used to monitor the DC bus voltage of the inverter-driven motor system. and the three-phase voltage at the inverter output Sampling is performed to obtain the raw voltage data. Because the inverter introduces a dead-time effect during switching, causing a deviation between the actual and theoretical output voltage values, dead-time compensation is performed. Assume the conduction time of the switching devices is... The output voltage is corrected using the following formula: ; in This is the actual voltage after compensation. It is the ideal voltage value. It's dead time. This refers to the switching cycle. The compensated output voltage signal serves as the input to the motor voltage dynamic equation. The motor rotor speed signal is acquired via a speed sensor. Since speed signals in mechanical systems are often affected by noise, such as gear meshing noise or fluctuations caused by vibration, the original speed signal is subjected to mean filtering. The mean filtering formula is: ; in It is the size of the filter window. This refers to the sampling interval. The filtered rotational speed signal. As an important variable in the mechanical dynamic equation, the stator current dynamic equation is established based on the collected signals. According to the stator circuit equation of the motor, the following is derived: ; in It is the αβ component of the stator current. These are the αβ components of the stator voltage. (Matrix) and The form is: ; in It is the stator resistance. It is a stator inductor. It is the identity matrix. This equation describes the dynamic behavior of the stator current. A flux linkage analysis is performed on the stator current dynamic equation using the definition of rotor flux linkage. And magnetic flux linkage equations: ; in It is rotor flux linkage. It is the rotor resistance. It is the rotor inductance. This is the mutual inductance coefficient. Based on this, a dynamic equation for the rotor flux linkage is established to describe its time evolution. Combining the above equation with the dynamic equation for the rotational speed of the mechanical part: ; in It is the rotor's mechanical angular velocity. It is the rotational inertia of the system. It is electromagnetic torque. This is the load torque. Combining these equations forms a state-space equation set, and the system's state vector is... To numerically solve the system of state equations, the fourth-order Runge-Kutta method is used, and its basic formula is as follows: ; in: ; ; For time step, This represents the right-hand side function of the state equations. Through the above steps, the Runge-Kutta method is used to calculate the change of the system state over time under given initial conditions, resulting in an accurate state-space model.

[0026] In one specific embodiment, the process of performing step 102 may specifically include the following steps: The state-space model is transformed by coordinate transformation. The Park transformation matrix is ​​constructed based on the stator three-phase current signal and speed signal to obtain the state variables in the dq rotating coordinate system. The state variables in the dq rotating coordinate system are decomposed by separating the stator current component, rotor flux component and mechanical angular velocity component to obtain the state variable set. Based on the state variable set, construct a quadratic function for the stator current error term, add the squares of the d-axis current error and the q-axis current error to obtain the current energy term, and construct a quadratic function for the flux linkage error term based on the state variable set, add the squares of the d-axis flux linkage error and the q-axis flux linkage error to obtain the flux linkage energy term. By weighting the current energy term and the flux energy term, and setting the current weighting coefficient and the flux weighting coefficient to the reciprocal of the motor's rated current and the reciprocal of the rated flux, respectively, the combined energy function is obtained. The derivative expression of the energy function is solved by the state variable set. The energy function is differentiated with respect to time to obtain the energy change rate equation. The coefficient matrix of the energy change rate equation is designed to construct a positive definite matrix so that the energy function change rate is less than zero, thus obtaining the Lyapunov stability criterion. The combined energy function is calculated online based on the Lyapunov stability criterion, and the energy function values ​​under normal operating conditions are recorded to obtain the training energy function characteristics of the inverter drive motor system.

[0027] Specifically, based on the stator three-phase current signal and motor speed signal Construct the Park transformation matrix. Map the current components of the three-phase stationary coordinate system to a dq rotating coordinate system based on the rotor magnetic field. The Park transformation formula is as follows: ; in Indicates the angular position of the rotor magnetic field. and These are the current components in the dq coordinate system. After the coordinate transformation, the state variables in the dq coordinate system are decomposed to obtain the stator current components. Rotor flux component and mechanical angular velocity These variables together constitute the state variable set, used for subsequent energy function construction. Based on the state variable set, a quadratic form function of the stator current error term is constructed. The current error term is defined as the deviation between the actual current and the target current, and its energy expression is: ; in and This is the target current value along the dq axis. and This is the actual current value. This energy term reflects the impact of dynamic current deviation on system performance. A quadratic form function of the flux linkage error term is constructed based on the state variable set. The flux linkage error term is defined as the deviation between the actual flux linkage and the target flux linkage, and its energy expression is: ; in and It is the target flux linkage value along the dq axis. and This is the actual flux linkage value. The flux linkage energy term describes the impact of rotor flux linkage deviation on system stability. A combined energy function is constructed by weighting the current energy term and the flux linkage energy term. The weighting coefficients are set as the reciprocals of the motor's rated current and rated flux linkage, respectively. The introduction of weights ensures that different physical quantities have unified dimensions in the energy function. The expression for the combined energy function is: ; in and Let be the rated current and rated flux linkage of the motor, respectively. To analyze the stability of the system, the derivative of the combined energy function is taken, yielding the energy change rate equation: ; Using the state-space model, the rate of energy change is further expressed as: ; in For the state variable group vector, Let be the coefficient matrix. To ensure system stability, a positive definite matrix is ​​designed. This ensures that the rate of change of the energy function is always less than zero, that is: ; Based on Lyapunov stability theory, the design of the positive definite matrix ensures that the system energy continuously decreases under normal operating conditions, eventually converging to a stable point. In online calculations, the values ​​of the combined energy function are recorded, forming the training energy function characteristics of the inverter drive motor system. Analysis of the energy function characteristics under normal operating conditions provides a benchmark for subsequent fault diagnosis.

[0028] In one specific embodiment, the process of performing step 103 may specifically include the following steps: The stator three-phase current signal is normalized by mapping the three-phase current signal to a unit interval based on the motor's rated current value to obtain the normalized current signal. Based on the normalized current signal, a Clarke transform matrix is ​​established to transform the three-phase stationary coordinate system into a two-phase orthogonal coordinate system, and the current components in the αβ coordinate system are obtained. Voltage equations and flux linkage equations are constructed based on the state-space model. A current prediction model is established based on the voltage balance relationship and the flux linkage coupling relationship to obtain the predicted state equations. The prediction state equation is discretized, and the continuous state equation is converted into a difference equation using the Euler approximation method to obtain a discrete prediction model. The predicted values ​​of the α-axis and β-axis currents for the next sampling period are calculated based on the discrete prediction model. The predicted current values ​​are recorded as prediction vectors to obtain the predicted current values. The current components in the αβ coordinate system are sampled, and the sampled current values ​​are recorded as the measured vector to obtain the measured current values. The difference between the predicted vector and the measured vector is calculated to obtain the α-axis residual component and the β-axis residual component, respectively. The two-axis residual components are then combined to obtain the training current residual vector.

[0029] Specifically, the acquired stator three-phase current signals Normalization is performed to eliminate the influence of the absolute magnitude of the signal amplitude on subsequent calculations, mapping it to a standard unit interval. The normalization formula is: ; in This is the rated current value of the motor. This is the normalized current signal. Normalization effectively processes motor signals with different rated currents, ensuring system versatility. After normalization, a Clarke transform matrix is ​​constructed based on the three-phase current signals to transform the signals from the three-phase stationary coordinate system to a two-phase orthogonal coordinate system (αβ coordinate system). The Clarke transform formula is: ; in They are The current components along the α and β axes reflect the projection of the three-phase signals onto the αβ plane. This transformation simplifies the three-phase system to a two-phase system, reducing computational complexity. Voltage and flux linkage equations are constructed based on a state-space model. These equations reflect the dynamic coupling relationship between voltage, current, and flux linkage. The voltage equation is expressed as: ; in It is the stator voltage. It is the stator resistance. It is a stator inductor. Stator current It is the induced electromotive force. The magnetic flux linkage equation is expressed as: ; in It is rotor flux linkage. This refers to the mutual inductance coefficient. A current prediction model is established using the voltage equation and the flux linkage equation. This is achieved by rewriting the voltage equation in the form of the current derivative: ; A predictive state equation is constructed to calculate the dynamic evolution of the current in the αβ coordinate system. To adapt to the discrete-time characteristics of digital control, the predictive state equation is discretized. Using the Euler approximation method, the continuous state equation is transformed into a difference equation, with the following form: ; in It is the sampling period. This is the current value at the current sampling moment. The predicted current value at the next sampling time is calculated using a discrete prediction model, calculating the predicted current values ​​along the α-axis and β-axis for the next sampling period. These predicted current values ​​are then denoted as a prediction vector. Simultaneously, the current components in the αβ coordinate system are sampled in real time to obtain the measured current values. The difference between the predicted vector and the measured vector is used to obtain the α-axis residual components. and β-axis residual components : ; Combining the two-axis residual components yields the training current residual vector: .

[0030] In one specific embodiment, the process of performing step 104 may specifically include the following steps: A sliding time window is set for the training energy function features and the training current residual vector. The data sequence within each time window is reconstructed into an m-dimensional vector. The data sequence is then segmented according to the embedding dimension m and the time delay τ to obtain the reconstructed vector sequence. Calculate the similarity between vectors in the reconstructed vector sequence, set a similarity tolerance r, determine the degree of similarity between two reconstructed vectors based on Euclidean distance, count the number of reconstructed vectors similar to the template vector, and obtain the similar vector count. The conditional probability is calculated based on the similarity vector count. The number of similar vectors is divided by the total number of reconstructed vectors, and the conditional probabilities for template lengths of m and m+1 are calculated respectively to obtain two sets of conditional probability values. Perform logarithmic operations on the two sets of conditional probability values ​​and calculate the average value. Subtract the average log-conditional probabilities when the template length is m and m+1 to obtain the sample entropy value. The sample entropy values ​​of the training energy function features and the training current residual vector are calculated sequentially according to the sliding time window. The calculated entropy values ​​are arranged in chronological order to obtain the entropy value feature sequence. A deep neural network with a multilayer perceptron structure is constructed. The deep neural network includes an input layer, three hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the entropy feature sequence. The number of neurons in the three hidden layers are 128, 64, and 32, respectively. The number of neurons in the output layer is equal to the number of fault types. The hidden layer is subjected to a nonlinear transformation using the ReLU activation function, the output of each layer is batch normalized, and the output layer is subjected to probability normalization using the Softmax function to obtain the probability distribution of fault types. The entropy feature sequence is input into a deep neural network, the backpropagation algorithm is used to calculate the gradient of the network parameters, and the Adam optimizer is used to optimize and update the network parameters to obtain a fault classification model.

[0031] Specifically, a sliding time window is set for the training energy function features and the training current residual vector. The size of the sliding time window is defined as follows: The time series data within the window can be represented as Based on embedding dimension and time delay Data within each time window is processed -Dimensional vector reconstruction. The time series is segmented into a sequence of reconstructed vectors: ; in This represents the starting position of the time series. Represent a -1D Reconstructed Vectors. This maps a one-dimensional time series to a high-dimensional space, thus preserving the temporal characteristics of the data. The similarity between vectors in the reconstructed vector sequence is calculated. To measure the similarity between two reconstructed vectors, Euclidean distance is introduced as a metric. For two... -Dimensional reconstruction vector and Its Euclidean distance is defined as: ; Set similarity tolerance Count all pairs of template vectors The Euclidean distance is less than Number of reconstructed vectors This number reflects the repetition pattern of the current template vector in the time series. The conditional probability is calculated using similar vector counting. The conditional probability is defined as: ; in This is the length of the time series. The embedding dimension is obtained by averaging the conditional probabilities of all template vectors. The overall conditional probability: ; Similarly, for embedding dimension The corresponding conditional probability is calculated. To extract the complexity features of the time series, the logarithmic operation is performed on the two sets of conditional probability values ​​mentioned above, and the average is calculated. The sample entropy is defined as: ; This indicator reflects the dynamic complexity of a time series; the higher the sample entropy value, the greater the randomness and uncertainty of the time series. The sample entropy value is calculated one by one for the training energy function features and the training current residual vector according to a sliding time window. The calculation results are arranged in chronological order to form an entropy feature sequence. The entropy feature sequence is represented as follows: ,in This represents the total number of sliding windows. After extracting entropy features, a deep neural network with a multilayer perceptron structure is constructed to classify fault types. This deep neural network consists of an input layer, three hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the entropy feature sequence, and it is used to receive time-series data. The number of neurons in the three hidden layers are set to 128, 64, and 32, respectively, to progressively extract features and reduce dimensionality. The number of neurons in the output layer is equal to the number of fault types, and it is used to output the probability distribution of fault types. In the hidden layers, the ReLU activation function is used for nonlinear transformation, and its expression is: ; This activation function effectively introduces nonlinearity, enhancing the network's expressive power. Simultaneously, batch normalization is performed on the output of each layer to ensure gradient stability and accelerate convergence. In the output layer, the Softmax activation function is used for probability normalization, its expression being: ; in It is The activation value of each output node. It is The probability of each fault type is calculated. After inputting the entropy feature sequence into the deep neural network, the output is calculated through forward propagation, and the network parameters are updated using the backpropagation algorithm. During backpropagation, the gradient is calculated based on the loss function (such as cross-entropy loss), and the parameters are updated using the Adam optimizer. The update rule of the Adam optimizer is: ; in It is a parameter vector. It's the learning rate. and These are unbiased corrections for the first-order and second-order momentum estimates, respectively. This is a small value used to prevent the denominator from being zero. Through multiple iterations of training, the network gradually adjusts its parameters, ultimately resulting in a deep neural network model capable of accurately classifying fault types.

[0032] In one specific embodiment, the process of performing step 105 may specifically include the following steps: The real-time operating data of the inverter drive motor system is collected and processed, and the energy function value at the current moment is calculated based on the state space model to obtain the real-time energy function characteristics. Sample entropy is calculated for the real-time energy function features, and a sliding window is used with embedding dimension m and similarity tolerance r to obtain the real-time entropy value features; The direction angle of the real-time current residual vector in the αβ coordinate system is calculated. The spatial angle of the current residual vector is calculated based on the ratio of the α-axis component to the β-axis component, and the fault direction characteristics are obtained. The magnitude of the real-time current residual vector is calculated. The residual magnitude is obtained by taking the square root of the sum of the squares of the α-axis and β-axis components. The residual magnitude is then compared with a preset threshold to obtain the fault magnitude characteristics. The real-time entropy value features are input into the fault classification model. The probability distribution of fault types is calculated based on the forward propagation of the deep neural network. The type corresponding to the maximum probability value is selected as the initial fault type. The preliminary fault type is verified based on the fault direction characteristics, and the fault type is confirmed based on the correspondence between the fault type and the current residual vector direction, thus obtaining the fault type discrimination result. Threshold judgment is performed on the fault amplitude characteristics, the severity of the fault is determined according to the amplitude distribution characteristics of different fault types, the fault severity level is obtained, and the fault location information of the fault occurrence is located based on the fault type discrimination result and the fault severity level, combined with the fault direction characteristics.

[0033] Specifically, real-time operating data of the system is collected, including stator current signals, voltage signals, and speed signals. After real-time processing, this data is used to calculate the energy function value at the current moment based on a pre-established state-space model. The formula for calculating the energy function is: ; in It is a state variable vector, including stator current components, voltage components, and rotor flux linkage components. This is a positive definite coefficient matrix, used to ensure the physical meaning of the energy function. Through this calculation, real-time energy function characteristics can be obtained, reflecting the dynamic characteristics of the system. Sample entropy is calculated on the real-time energy function characteristics. Embedding dimension is used. Similarity tolerance The data is segmented and analyzed using a sliding time window. (Time series data) Reconstructed within each window -Dimensional vector: ; in It's a time delay. This is the reconstructed state vector. By calculating the Euclidean distance between each vector and other vectors, it is determined whether the distance is less than the similarity tolerance. And count the number of similar vectors. Conditional probability The calculation formula is: ; Similarly, the embedding dimension is calculated as follows: conditional probability The expression for sample entropy is: ; By calculating the sample entropy values ​​of the energy function one by one through a sliding time window, the real-time entropy characteristics are obtained, reflecting the changes in the dynamic complexity of the system. The direction angle of the real-time current residual vector in the αβ coordinate system is calculated. Assume the real-time current residual vector is... The formula for calculating its direction angle is: ; Direction angle This represents the spatial direction of the residual vector in the αβ plane, serving as a fault direction feature for subsequent fault type verification. The amplitude of the real-time current residual vector is calculated using the following formula: ; By residual amplitude With preset threshold The fault amplitude features are compared and extracted to determine the severity of the fault. Real-time entropy features are input into a pre-trained fault classification model, and the probability distribution of fault types is calculated through forward propagation of a deep neural network. Assume the output of the deep neural network is... ,in Indicates the The probability of a fault type is determined. A preliminary fault type is established by selecting the fault type corresponding to the highest probability. After determining the preliminary fault type, verification is performed using fault direction characteristics. Based on the theoretical analysis results of different fault types, each fault type has a specific distribution range in the residual direction of the αβ plane, called the direction feature template. For example, the direction range of a certain type of fault is... By judgment Whether the fault falls within the directional feature template verifies the rationality of the initial fault type. If the directional feature matches, the current initial fault type is confirmed; otherwise, the fault type corresponding to the suboptimal probability is selected for re-verification. When performing threshold judgment on the fault amplitude features, the severity of the fault is determined based on the amplitude distribution characteristics of different fault types. Assume the amplitude range of a certain type of fault is... By comparing residual amplitudes Whether the fault exceeds the normal range is determined, and the severity level of the fault is classified, such as minor, moderate, or severe. Based on the fault type and severity level, combined with the fault direction characteristics, the location of the fault is determined. For example, if a certain fault type matches the current residual characteristics of a specific phase winding, the fault is located in that phase winding; if the direction angle points to a switching element on the inverter output side, the fault location belongs to that element.

[0034] In one specific embodiment, the process of verifying the preliminary fault type based on the fault direction characteristics, confirming the fault type based on the correspondence between the fault type and the current residual vector direction, and obtaining the fault type discrimination result can specifically include the following steps: The fault direction characteristics are divided into regions. Based on the αβ plane, the 360-degree spatial angle is divided into 12 sectors to obtain a fault region mapping table. A feature direction database is established based on the fault area mapping table. The angle interval corresponding to each type of fault is determined according to the theoretical analysis results of different fault types to obtain the direction feature template. The table is then looked up for the preliminary fault type, and the theoretical direction range of the fault type is obtained according to the direction feature template to obtain the fault direction verification interval. The fault direction characteristics are judged in intervals, and the actual direction of the current residual vector is compared with the fault direction verification interval to obtain the direction matching degree. A threshold judgment is made on the direction matching degree, and the direction characteristics of the fault type are determined according to the preset allowable range of direction deviation to determine whether they conform to the theoretical analysis, thus obtaining the direction verification result; The initial fault type is adjusted based on the direction verification results. When the direction verification passes, the current fault type is confirmed. When the direction verification fails, the fault type corresponding to the suboptimal probability is selected for re-verification to obtain the candidate fault type. A confidence analysis is performed on the candidate fault types. The fault confidence is obtained by weighting the probability value and direction matching degree output by the fault classification model. Based on the fault confidence, the fault type with the highest confidence is selected as the fault type discrimination result.

[0035] Specifically, the αβ plane is divided into 12 equal-angle sectors by a 360-degree spatial angle. Each sector has an angle range of 30 degrees and is numbered sequentially as follows: Within these sectors, the range of each sector is represented as follows: ; This division provides a standardized spatial mapping for subsequent directional feature verification, ensuring that any angle corresponds to a specific sector. Based on this division, a fault region mapping table is constructed and a feature direction database is established. The current residual vectors of different fault types exhibit specific distribution patterns in the αβ plane, which are determined through theoretical analysis and experimental verification. For example, the residual vector of a certain type of stator winding short-circuit fault is mainly distributed in... and Sector, while inverter switch faults are concentrated in to The range. By associating each fault type with its possible directional feature regions, a directional feature template is formed: ; in and It is the first The theoretical scope of fault types. When a certain initial fault type... After the classification model makes a prediction, it looks up the fault area mapping table and extracts the directional feature template of this type of fault from the feature direction database to obtain the theoretical directional range. This refers to the fault direction verification interval. The real-time calculated fault direction characteristics (i.e., the direction angle of the current residual vector) are compared with the fault direction verification interval. (Direction angle) The calculation formula is: ; in and These are the current residual components along the α and β axes, respectively. By judging... Does it fall within the verification range? Inside, calculate the direction matching degree Match: ; To enhance the robustness of orientation verification, a preset allowable range for orientation deviation is introduced. When the direction matching degree falls below a certain threshold, the current fault direction characteristics are determined to be inconsistent with the theoretical direction characteristics, resulting in direction verification failure. In this case, the initial fault type needs to be adjusted. The adjustment method is to select the fault type corresponding to the suboptimal probability. The classification model outputs the type with the second highest probability, and the orientation is re-validated to obtain candidate fault types. After determining the candidate fault types, a confidence analysis is performed. The confidence analysis combines the probability value output by the classification model and the orientation matching degree, and calculates the final fault confidence score through weighted average. The expression for the fault confidence score is: ; in Is the classification model for the first The predicted probability of a fault class, where Match represents the direction matching degree. and These are the weighting factors for probability and direction matching, respectively, satisfying... By calculating the confidence scores of all candidate types, the type with the highest confidence score is selected. This serves as the final result for determining the fault type.

[0036] The above describes the inverter motor control fault diagnosis method in the embodiments of the present invention. The following describes the inverter motor control fault diagnosis device in the embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the inverter motor control fault diagnosis device of the present invention includes: The acquisition module 201 is used to acquire the stator three-phase current signal, inverter output voltage signal and speed signal of the inverter drive motor system, and to establish a state-space model including the stator current dynamic equation, rotor flux dynamic equation and speed dynamic equation. Module 202 is established to establish quadratic Lyapunov functions for the state variables in the state-space model, and to construct the energy function coefficient matrix based on the derivative terms of the state variables, thereby obtaining the training energy function characteristics of the inverter drive motor system. The calculation module 203 is used to perform Clarke transformation on the stator three-phase current signal to obtain the current components in the αβ coordinate system, and calculate the predicted current value according to the state space model. The difference between the predicted current value and the measured current value is used to obtain the training current residual vector. Training module 204 is used to calculate sample entropy on the training energy function features and training current residual vector to obtain entropy value feature sequence, and train deep neural network based on entropy value feature sequence to obtain fault classification model; The fault identification module 205 is used to input the real-time energy function features and the real-time current residual vector into the fault classification model, and to identify the fault by combining the distribution characteristics of the real-time current residual vector in the αβ plane, thereby obtaining the fault type discrimination result and fault location information.

[0037] Through the collaborative efforts of the aforementioned components, a state-space model incorporating stator current, rotor flux linkage, and rotational speed was established. This model, combined with Lyapunov energy function analysis, comprehensively considered the system's dynamic characteristics, improving the accuracy of fault diagnosis. The Clarke transform was used to convert the three-phase currents to the αβ coordinate system, and the state-space model was used for current prediction. By calculating the current residual vector, the abnormal characteristics of the system caused by the fault were effectively captured. A sample entropy calculation method was introduced to analyze the entropy values ​​of the energy function characteristics and the current residual vector, extracting more discriminative fault features. A deep neural network with a multilayer perceptron structure was designed, and through multilayer nonlinear transformation and batch normalization, the model's feature learning ability was enhanced. Fault verification was performed by combining the distribution characteristics of the current residual vector in the αβ plane. Through comprehensive analysis of fault direction and amplitude characteristics, accurate fault type identification and precise location were achieved.

[0038] above Figure 2 The fault diagnosis device for inverter motor control in this embodiment of the invention is described in detail from the perspective of modular functional entities. The fault diagnosis device for inverter motor control in this embodiment of the invention is described in detail below from the perspective of hardware processing.

[0039] Figure 3This is a schematic diagram of the structure of a frequency converter motor control fault diagnosis device 300 provided in an embodiment of the present invention. The frequency converter motor control fault diagnosis device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the frequency converter motor control fault diagnosis device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute a series of instruction operations in the storage media 330 on the frequency converter motor control fault diagnosis device 300 to implement the steps of the above-described frequency converter motor control fault diagnosis method.

[0040] The inverter motor control fault diagnosis device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The illustrated structure of the inverter motor control fault diagnosis device does not constitute a limitation on the inverter motor control fault diagnosis device provided by the present invention. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0041] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the inverter motor control fault diagnosis method.

[0042] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0043] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0044] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing inverter motor control faults, characterized in that: The method comprises: Collect the stator three-phase current signal, inverter output voltage signal and speed signal of the inverter drive motor system, and establish a state space model including the stator current dynamic equation, rotor flux dynamic equation and speed dynamic equation; Establishing a quadratic Lyapunov function for the state variables in the state space model, and constructing an energy function coefficient matrix according to the derivative terms of the state variables, to obtain training energy function characteristics of the inverter-driven motor system; Performing Clarke transformation on the stator three-phase current signal to obtain current components in the αβ coordinate system, and calculating a predicted current value according to the state space model, and subtracting the predicted current value from the measured current value to obtain a training current residual vector; Performing sample entropy calculation on the training energy function feature and the training current residual vector to obtain an entropy value feature sequence, and training a deep neural network according to the entropy value feature sequence to obtain a fault classification model; The real-time energy function characteristics and the real-time current residual vector are input into the fault classification model, and the fault identification is performed in combination with the distribution characteristics of the real-time current residual vector in the αβ plane to obtain the fault type discrimination result and the fault location information.

2. The inverter motor control fault diagnosis method according to claim 1, characterized in that: The stator three-phase current signal, inverter output voltage signal and speed signal of the inverter drive motor system are collected, and a state space model including a stator current dynamic equation, a rotor flux dynamic equation and a speed dynamic equation is established, including: A current sensor is used to collect current signals from a frequency converter driven motor system to obtain raw current data, and anti-aliasing filtering is performed on the raw current data to obtain a stator three-phase current signal; A voltage sensor is used to sample the DC bus and the inverter output terminal of the inverter-driven motor system to obtain original voltage data, and dead zone effect compensation is performed on the original voltage data, and an effective voltage value is calculated according to the conduction time of the switch device to obtain an inverter output voltage signal; A speed sensor is used to collect the rotor speed of the inverter-driven motor system, and mean filtering is performed to obtain a speed signal; Establishing a stator current dynamic equation according to the stator three-phase current signal, the inverter output voltage signal and the speed signal, taking the stator resistance and the stator inductance as equation parameters, to obtain the stator current dynamic equation; Performing a flux relationship analysis on the stator current dynamic equation, and establishing a rotor flux dynamic equation based on rotor inductance and mutual inductance parameters; According to the stator current dynamic equation and the rotor flux dynamic equation, a speed dynamic equation is established in combination with the moment of inertia parameter, and the state equation group is solved by the fourth-order Runge-Kutta method to obtain a state space model.

3. The inverter motor control fault diagnosis method according to claim 2, characterized in that: The quadratic Lyapunov function is established for the state variables in the state space model, and an energy function coefficient matrix is ​​constructed according to the derivative terms of the state variables to obtain the training energy function characteristics of the inverter driven motor system, including: Performing coordinate transformation on the state space model, constructing a Park transformation matrix according to the stator three-phase current signal and the speed signal, and obtaining state variables in a dq rotating coordinate system; Decomposing the state variables in the dq rotating coordinate system, separating the stator current component, the rotor flux component and the mechanical angular velocity component, and obtaining a state variable group; A quadratic function of a stator current error term is constructed according to the state variable group, and the square terms of the d-axis current error and the q-axis current error are added to obtain a current energy term; and a quadratic function of a flux linkage error term is constructed based on the state variable group, and the square terms of the d-axis flux linkage error and the q-axis flux linkage error are added to obtain a flux linkage energy term; Performing a weighted combination on the current energy term and the flux linkage energy term, setting the current weight coefficient and the flux linkage weight coefficient to the inverse of the rated current and the inverse of the rated flux linkage of the motor, respectively, to obtain a combined energy function; Solving the energy function derivative expression according to the state variable group, differentiating the combined energy function with respect to time to obtain an energy change rate equation, performing coefficient matrix design on the energy change rate equation, constructing a positive definite matrix to make the energy function change rate less than zero, and obtaining a Lyapunov stability criterion; The combined energy function is calculated online according to the Lyapunov stability criterion, the energy function value under normal operating conditions is recorded, and the training energy function characteristics of the inverter-driven motor system are obtained.

4. The inverter motor control fault diagnosis method according to claim 3, characterized in that: The Clarke transformation is performed on the stator three-phase current signal to obtain the current component in the αβ coordinate system, and the predicted current value is calculated according to the state space model, and the predicted current value is subtracted from the measured current value to obtain the training current residual vector, including: Normalizing the stator three-phase current signal, mapping the three-phase current signal to a unit interval according to the rated current value of the motor, and obtaining a normalized current signal; A Clarke transformation matrix is ​​established according to the normalized current signal, and a three-phase stationary coordinate system is transformed into a two-phase orthogonal coordinate system to obtain current components in an αβ coordinate system; Based on the state space model, a voltage equation and a flux equation are constructed, and a current prediction model is established according to a voltage balance relationship and a flux coupling relationship to obtain a prediction state equation; Discretizing the prediction state equation, converting the continuous state equation into a difference equation using the Euler approximation method, and obtaining a discrete prediction model; Calculate the predicted current values ​​of the α-axis and β-axis of the next sampling period according to the discrete prediction model, record the predicted current value as a prediction vector, and obtain the predicted current value; Sampling the current component in the αβ coordinate system, recording the sampled current value as a measured vector, and obtaining a measured current value; The difference is calculated based on the predicted vector and the measured vector to obtain an α-axis residual component and a β-axis residual component respectively, and the two-axis residual components are combined to obtain a training current residual vector.

5. The inverter motor control fault diagnosis method according to claim 4, characterized in that: The sample entropy calculation is performed on the training energy function feature and the training current residual vector to obtain an entropy value feature sequence, and a deep neural network is trained according to the entropy value feature sequence to obtain a fault classification model, including: Setting a sliding time window for the training energy function feature and the training current residual vector, reconstructing the data sequence in each time window by an m-dimensional vector, segmenting the data sequence according to the embedding dimension m and the time delay τ, and obtaining a reconstructed vector sequence; Calculating the similarity between the reconstructed vector sequence, setting a similarity tolerance r, judging the similarity between two reconstructed vectors according to the Euclidean distance, and counting the number of reconstructed vectors similar to the template vector to obtain a similar vector count; Calculate the conditional probability according to the similar vector count, divide the number of similar vectors by the total number of reconstructed vectors, calculate the conditional probability when the template length is m and m+1 respectively, and obtain two sets of conditional probability values; Performing logarithmic operations on the two sets of conditional probability values ​​and calculating the average value, subtracting the average logarithmic conditional probabilities when the template length is m and m+1, to obtain the sample entropy value; Calculating the sample entropy values ​​of the training energy function characteristics and the training current residual vector in sequence according to the sliding time window, arranging the calculated entropy values ​​in chronological order, and obtaining an entropy value characteristic sequence; Constructing a deep neural network with a multi-layer perceptron structure, the deep neural network includes an input layer, three hidden layers and an output layer, wherein the number of neurons in the input layer is equal to the dimension of the entropy value feature sequence, the number of neurons in the three hidden layers is 128, 64 and 32 respectively, and the number of neurons in the output layer is equal to the number of fault types; The hidden layer is subjected to nonlinear transformation using a ReLU activation function, the output of each layer is subjected to batch normalization, and the output layer is subjected to probability normalization using a Softmax function to obtain a probability distribution of the fault type; The entropy value feature sequence is input into the deep neural network, the network parameter gradient is calculated using the back propagation algorithm, and the network parameters are optimized and updated using the Adam optimizer to obtain a fault classification model.

6. The inverter motor control fault diagnosis method according to claim 5, characterized in that: The real-time energy function feature and the real-time current residual vector are input into the fault classification model, and the fault identification is performed in combination with the distribution characteristics of the real-time current residual vector in the αβ plane to obtain the fault type discrimination result and the fault location information, including: Collecting and processing the real-time operation data of the inverter-driven motor system, calculating the energy function value at the current moment according to the state space model, and obtaining the real-time energy function characteristics; Performing sample entropy calculation on the real-time energy function feature, performing sliding window processing using the embedding dimension m and the similarity tolerance r, and obtaining a real-time entropy value feature; The direction angle of the real-time current residual vector in the αβ coordinate system is calculated, and the spatial angle of the current residual vector is calculated according to the ratio of the α-axis component to the β-axis component to obtain the fault direction feature; Performing amplitude calculation on the real-time current residual vector, obtaining a residual amplitude according to the square root of the sum of the squares of the α-axis component and the β-axis component, and comparing the residual amplitude with a preset threshold to obtain a fault amplitude feature; Input the real-time entropy value feature into the fault classification model, calculate the fault type probability distribution according to the forward propagation of the deep neural network, and select the type corresponding to the maximum probability as the preliminary fault type; Verifying the preliminary fault type according to the fault direction feature, confirming the fault type based on the corresponding relationship between the fault type and the current residual vector direction, and obtaining a fault type discrimination result; A threshold judgment is performed on the fault amplitude feature, the fault severity is determined according to the amplitude distribution characteristics of different fault types, and the fault degree level is obtained. The fault location information of the fault occurrence is located based on the fault type judgment result and the fault degree level in combination with the fault direction feature.

7. The inverter motor control fault diagnosis method according to claim 6, characterized in that: The verifying the preliminary fault type according to the fault direction feature, confirming the fault type based on the corresponding relationship between the fault type and the current residual vector direction, and obtaining the fault type determination result includes: Performing regional division on the fault direction feature, dividing the 360-degree spatial angle into 12 sectors according to the αβ plane, and obtaining a fault area mapping table; Establish a characteristic direction database based on the fault area mapping table, determine the angle interval corresponding to each type of fault according to the theoretical analysis results of different fault types, obtain a direction characteristic template, and look up the table for the preliminary fault type, obtain the theoretical direction range of this type of fault according to the direction characteristic template, and obtain the fault direction verification interval; Performing interval judgment on the fault direction feature, comparing the actual direction of the current residual vector with the fault direction verification interval to obtain a direction matching degree; A threshold value is judged on the direction matching degree, and whether the direction characteristics of the fault type conform to the theoretical analysis is determined according to a preset direction deviation allowable range, so as to obtain a direction verification result; The preliminary fault type is adjusted according to the direction verification result, and the current fault type is confirmed when the direction verification passes, and the fault type corresponding to the suboptimal probability is selected for re-verification when the direction verification fails to obtain a candidate fault type; A confidence analysis is performed on the candidate fault types, and a fault confidence is obtained based on a weighted calculation of the probability value and the direction matching degree output by the fault classification model. According to the fault confidence, the fault type with the highest confidence is selected as the fault type discrimination result.

8. A frequency converter motor control fault diagnosis device, characterized in that: The device is used to execute the inverter motor control fault diagnosis method according to any one of claims 1 to 7, comprising: The acquisition module is used to acquire the stator three-phase current signal, inverter output voltage signal and speed signal of the inverter drive motor system, and establish a state space model including the stator current dynamic equation, the rotor flux dynamic equation and the speed dynamic equation; An establishment module is used to establish a quadratic Lyapunov function for the state variables in the state space model, and to construct an energy function coefficient matrix according to the derivative terms of the state variables to obtain the training energy function characteristics of the inverter driven motor system; A calculation module, used for performing Clarke transformation on the stator three-phase current signal to obtain current components in the αβ coordinate system, and calculating a predicted current value according to the state space model, and subtracting the predicted current value from the measured current value to obtain a training current residual vector; A training module, used to perform sample entropy calculation on the training energy function characteristics and the training current residual vector to obtain an entropy value feature sequence, and train a deep neural network according to the entropy value feature sequence to obtain a fault classification model; The fault identification module is used to input the real-time energy function characteristics and the real-time current residual vector into the fault classification model, and perform fault identification in combination with the distribution characteristics of the real-time current residual vector in the αβ plane to obtain the fault type discrimination result and fault location information.

9. A frequency converter motor control fault diagnosis device, characterized in that: The inverter motor control fault diagnosis device comprises: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instructions in the memory to enable the inverter motor control fault diagnosis device to execute the inverter motor control fault diagnosis method according to any one of claims 1 to 7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the inverter motor control fault diagnosis method according to any one of claims 1 to 7 is implemented.

Images