Method for Training a Single-Phase Leg Diagnosis Model for Diagnosing a Multilevel Inverter and Method for Diagnosing the Condition of a Multilevel Inverter

Abstract

The present disclosure relates to a method for learning a single-phase leg diagnostic model for diagnosing a multi-level inverter and a method for diagnosing the state of a multi-level inverter. More specifically, the method for learning a single-phase leg diagnostic model for diagnosing a multi-level inverter comprising a plurality of legs including a plurality of switches, which is performed by a processor, wherein the processor comprises: (a1) a step of measuring the pole voltage of a leg to be learned, which is one of the plurality of legs; (a2) a step of obtaining the half-bridge voltage of the leg to be learned; and (a3) ​​a step of learning a single-phase leg diagnostic model that diagnoses the switch state of the leg to be learned based on the pole voltage, the half-bridge voltage, and the state information of the switch of the leg to be learned.

Description

Technology Field

[0001] The present disclosure relates to a method for learning a single-phase leg diagnostic model for diagnosing a multi-level inverter and a method for diagnosing the state of a multi-level inverter. More specifically, it relates to a method for learning a single-phase leg diagnostic model that learns switch failures of a single-phase leg of a multi-level inverter with a small amount of computation, and a method for diagnosing the state of a multi-level inverter that can rapidly diagnose switch failures for multiple legs using a single-phase leg diagnostic model. Background Technology

[0003] Multilevel inverters (MLIs) use multiple switches to generate outputs at multiple voltage levels and have the advantages of low Total Harmonic Distortion (THD) and high power conversion efficiency.

[0004] However, when the voltage level increases, the number of switches in the multi-level inverter also increases, which increases the likelihood of failure and can reduce the reliability of the system.

[0005] Multi-level inverters utilize Insulated Gate Bipolar Transistors (IGBTs), Metal-Oxide-Semiconductor Field-Effect Transistors (MOSFETs), Thyristors, and silicon carbide (SiC) and gallium nitride (GaN)-based switching devices. Among these, IGBTs are widely used in medium-to-high voltage inverters due to their ability to handle high voltages and currents.

[0006] The major failure types of IGBTs include short-circuit faults and open-circuit faults.

[0007] A short-circuit fault is a phenomenon in which the insulation between the gate and collector or emitter of an IGBT is broken, causing the switch to conduct continuously. In this case, an overcurrent flows, which can cause serious damage to the inverter and load within a short period of time.

[0008] An open fault is a phenomenon where the connection between the gate and collector is severed, preventing the switch from conducting normally. Open faults do not cause immediate damage. However, they induce voltage imbalance and distort the output voltage waveform, causing the system to operate abnormally.

[0009] To diagnose such IGBT switch failures, model-based, signal-based, and data-based diagnostic methods have conventionally been used. Model-based diagnostic methods construct a mathematical model of the inverter system and determine whether a switch failure exists by comparing values ​​measured during real-time operation with values ​​predicted by the model. Signal-based diagnostic methods detect failures by analyzing signals such as voltage, current, temperature, and frequency generated during inverter operation in real time. Data-based diagnostic methods diagnose failures by utilizing machine learning (ML) or artificial intelligence (AI) technologies.

[0010] Conventional model-based and signal-based diagnostic methods require complex mathematical models. When the number of voltage levels in a multi-level inverter increases, the number of switches to be diagnosed increases, which in turn increases the complexity of the mathematical model and the algorithms required for signal processing, making reliable fault detection difficult.

[0011] Conventional data-based diagnostic methods require a large number of samples to train a diagnostic model in advance for reliable fault diagnosis. However, as the number of samples increases, the amount of data throughput and computation increases exponentially, making it difficult to perform real-time fault diagnosis. In particular, for multi-stage inverters such as 4-level Active Neutral-Point-Clamped (ANPC) inverters, the computational burden is further increased as the number of combinations of multiple switch failure situations increases significantly, as open faults in two or more switches must be considered.

[0012] Therefore, technology is required that can maintain high diagnostic accuracy while reducing the number of samples for diagnosing multi-level inverters. Prior art literature

[0013] Published Patent Application No. 10-2018-0110540 (October 10, 2018) The problem to be solved

[0014] The present disclosure is devised to solve the problems described above. The purpose of the single-phase leg diagnostic model learning method and the multi-level inverter state diagnosis method according to the present disclosure is to provide a single-phase leg diagnostic model learning method and a multi-level inverter state diagnosis method that enable a reduction in the number of samples required for learning and a saving of computational resources compared to conventional methods, by training a single-phase leg diagnostic model using information on a single leg and using this to perform switch fault diagnosis for each of a plurality of legs one by one. means of solving the problem

[0016] A method for learning a single-phase leg diagnostic model for diagnosing a multi-level inverter according to the present disclosure for solving the problems described above is performed more specifically by a processor, and in a method for learning a single-phase leg diagnostic model for diagnosing a multi-level inverter comprising a plurality of legs including a plurality of switches, the processor comprises: (a1) a step of measuring the pole voltage of a leg to be learned, which is one of the plurality of legs; (a2) a step of obtaining the half-bridge voltage of the leg to be learned; and (a3) ​​a step of learning a single-phase leg diagnostic model that diagnoses the switch state of the leg to be learned based on the pole voltage, the half-bridge voltage, and the state information of the switch of the leg to be learned.

[0017] Additionally, in step (a2), the processor calculates the half-bridge voltage from the pole voltage based on the switch On / Off information of the leg to be learned.

[0018] In addition, the single-phase leg diagnostic model is an adaptive model in which the size or characteristics of the filter are dynamically adjusted.

[0019] In addition, the single-phase leg diagnostic model is optimized using a Reinforced Attention Optimizer (RAO).

[0020] In addition, the single-phase leg diagnosis model includes an Adaptive CNN, but is optimized using a Reinforced Attention Optimizer (RAO).

[0021] A method for diagnosing the state of a multi-level inverter according to the present disclosure is performed more specifically by a processor and, in a method for diagnosing the state of a multi-level inverter including a plurality of legs using a single-phase leg diagnosis model, the processor comprises: (b1) receiving a pole voltage of each of the plurality of legs; (b2) obtaining a half-bridge voltage of each of the plurality of legs; and (b3) diagnosing the state of the multi-level inverter based on the pole voltage of each of the plurality of legs and the half-bridge voltage using a plurality of pre-learned single-phase leg diagnosis models, wherein in step (b3), the plurality of single-phase leg diagnosis models are matched one by one for each of the plurality of legs and diagnose the state of the multi-level inverter in parallel.

[0022] Additionally, in step (b2), the processor calculates the half-bridge voltage from the pole voltage based on the switch On / Off information of the leg to be learned.

[0023] Additionally, in step (b3), the processor receives the pole voltage and half-bridge voltage of the matched leg for each of the plurality of single-phase leg diagnostic models and diagnoses the fault status of each switch included in the leg matched to each of the plurality of single-phase leg diagnostic models. Effects of the invention

[0025] According to the single-phase leg diagnostic model learning method for diagnosing a multi-level inverter according to the present disclosure as described above, since the single-phase leg diagnostic model performs learning only on the switch failure state of a single leg, the number of samples required for learning is significantly reduced compared to the prior art, and the amount of computation required to process it is reduced, thereby providing the effect of saving computing resources.

[0026] In addition, when the half-bridge voltage is calculated based on the pole voltage and switch status information, it is not necessary to have a circuit for sensing the half-bridge voltage, so the structure of the multi-level inverter can be simplified and manufacturing costs can be saved.

[0027] Furthermore, according to the state diagnosis method of a multi-level inverter according to the present disclosure, when diagnosing multiple legs of a multi-level inverter, a single-phase leg diagnosis model is independently matched to each leg to diagnose switch faults. As a result, the fault diagnosis algorithm is simplified, and since each leg operates independently, diagnosis based on the location of the fault is performed quickly and accurately, and the complexity of the system is reduced.

[0028] Through this, it can be applied to the diagnosis of switch failures in multi-level inverters of 4 levels or more, thus possessing high versatility that allows for wide application to various types of inverters. Brief explanation of the drawing

[0029] FIG. 1 is a circuit diagram of a multi-level inverter to which the present disclosure is applied. FIGS. 2a through 2f are voltage graphs at the time of individual switch failure of exemplary multi-level inverter legs. Figures 3a and 3b are voltage graphs at multiple switch failures of an exemplary multi-level inverter leg. FIG. 4 is a flowchart of a single-phase leg diagnostic model learning method according to an embodiment of the present disclosure. FIGS. 5a and 5b are block diagrams of a single-phase leg diagnostic model according to one embodiment. FIG. 6 is a flowchart of a state diagnosis method for a multi-level inverter according to an embodiment of the present disclosure. FIG. 7 is a flowchart of a single-phase leg diagnostic model learning method and a multi-level inverter status diagnostic method according to another embodiment. Figures 8a through d are graphs of exemplary pole voltages and half-bridge voltages. Figures 9a through 9f are graphs of pole voltage, half-bridge voltage, and switch failure occurrence flag according to switch failure. Specific details for implementing the invention

[0030] The purpose, features, and advantages of the above-described purpose of the present disclosure will become more apparent through the following embodiments in conjunction with the accompanying drawings. The following specific structural or functional descriptions are merely illustrative for the purpose of explaining embodiments according to the concept of the present disclosure, and embodiments according to the concept of the present disclosure may be implemented in various forms and should not be interpreted as being limited to the embodiments described in this specification or application. Since embodiments according to the concept of the present disclosure may be subject to various modifications and may take various forms, specific embodiments are illustrated in the drawings and described in detail in this specification or application. However, this is not intended to limit embodiments according to the concept of the present disclosure to specific disclosed forms, and should be understood to include all modifications, equivalents, and substitutions that fall within the spirit and scope of the present disclosure. Terms such as "first" and / or "second" may be used to describe various components, but the components are not limited to these terms. Terms may be used solely for the purpose of distinguishing one component from other components, for example, without departing from the scope of rights according to the concept of the present disclosure, such that the first component may be named the second component, and similarly, the second component may be named the first component. Where it is stated that a component is connected to or coupled with another component, it should be understood that it may be directly connected to or coupled with that other component, or that there may be other components in between. On the other hand, where it is stated that a component is directly connected to or directly coupled with another component, it should be understood that there are no other components in between. Other expressions used to describe the relationship between components, such as between, directly between, adjacent to, and directly adjacent to, should be interpreted in the same way.The terms used in this specification are used merely to describe specific embodiments and are not intended to limit the disclosure. Singular expressions include plural expressions unless the context clearly indicates otherwise. Terms such as "include" or "have" in this specification are intended to specify the existence of the described features, numbers, steps, actions, components, parts, or combinations thereof, and should not be understood as precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof. Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which this disclosure pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this specification. Hereinafter, the disclosure will be described in detail by describing preferred embodiments of the disclosure with reference to the accompanying drawings. Identical reference numerals in each drawing indicate identical components.

[0031] Preferred embodiments of a single-phase leg diagnostic model learning method for diagnosing a multi-level inverter and a state diagnostic method for a multi-level inverter according to the present disclosure will be described in detail below with reference to the attached drawings.

[0032] In the specification and drawings of the present disclosure, Vhb represents the half-bridge voltage of any one leg, Vp represents the pole voltage of any one leg, Vhba, Vhbb, and Vhbc each represent the half-bridge voltages of the first to third legs, and Vpa, Vpb, and Vpc each represent the pole voltages of the first to third legs.

[0033] FIG. 1 is a circuit diagram of a multi-level inverter (100) to which the present disclosure is applied.

[0034] The multi-level inverter (100) to which the present disclosure applies can drive a motor (200) by receiving power through an input terminal. Through this, rotational force can be provided to a load (300) connected through a shaft. In one embodiment, the multi-level inverter (100) may be a 4-Level Active Neutral-Point-Clamped inverter capable of outputting 4 levels of voltage.

[0035] A multi-level inverter (100) includes a plurality of legs (110) and a series capacitor circuit (120). Specifically, any one of the plurality of legs (110) includes a first to third totem pole circuit.

[0036] The first totem pole circuit includes first and second switches (S1, S2) connected in series with each other. Additionally, one end of the first totem pole circuit is connected to the positive terminal (P) of the input terminal of the multi-level inverter (100).

[0037] The second totem pole circuit includes third and fourth switches (S3, S4) connected in series with each other. Additionally, one end of the second totem pole circuit is connected to the common node of the first and second switches (S1, S2), and the other end is connected to the common node of the fifth and sixth switches (S5, S6). Additionally, the common node of the third and fourth switches (S3, S4) is connected to a motor (200) and delivers four levels of voltage (+Vdc, +Vdc / 3, -Vdc / 3, -Vdc) and current to drive the motor (200).

[0038] The third totem pole circuit includes fifth and sixth switches (S5, S6) connected in series with each other. Additionally, the other end of the third totem pole circuit is connected to the negative terminal (N) of the input terminal of the multi-level inverter (100).

[0039] The series capacitor circuit (120) includes first to third capacitors connected in series sequentially. Specifically, the first capacitor (C1) is connected in parallel with the first totem pole circuit. Additionally, one end of the second capacitor (C2) is connected to the other end of the first totem pole circuit, and the other end is connected to one end of the second totem pole circuit. Additionally, the third capacitor (C3) is connected in parallel with the second totem pole circuit. When a voltage Vdc is applied to the input terminal, a voltage of Vdc / 3 is applied to each of the first to third capacitors (C1 to 3).

[0040] In one embodiment, each of the plurality of legs (110) may include first to third totem pole circuits, and the legs (110) may be provided in constants of the motor (200). For example, referring to FIG. 1, if the motor (200) is a three-phase motor (200) driven in phases a, b, and c, the plurality of legs (110) include first to third legs (110-1 to 3) capable of supplying current to each of the phases a to c of the motor (200). Specifically, each of the first to third legs (110-1 to 3) can supply a phase a current (ia), a phase b current (ib), and a phase c current (ic) to the stator windings included in the motor (200) to drive the motor (200).

[0041] In each leg (110), the half-bridge voltage (Vhb) represents the potential difference between the common node of the first and second switches (S1, S2) and the common node of the fifth and sixth switches (S5, S6). Additionally, the pole voltage (Vp) represents the voltage of the common node of the third and fourth switches (S3, S4).

[0042] FIGS. 2a through f are voltage graphs at the time of individual switch failure of an exemplary multi-level inverter (100) leg (110).

[0043] Specifically, FIGS. 2a to 2f each illustrate a pole voltage graph when an open fault occurs in each of the first to sixth switches (S1 to 6). Referring to FIGS. 2a to 2f, when a fault occurs in an individual switch, a different pole voltage waveform is generated depending on the switch in which the fault occurred.

[0044] FIGS. 3a and 3b are voltage graphs at the time of multiple switch failures of an exemplary multi-level inverter (100) leg (110).

[0045] Specifically, FIG. 3a shows a pole voltage graph when an open fault occurs in the third switch (S3) of the first leg (110-1) and the third switch (S3) of the second leg (110-2). FIG. 3b shows a pole voltage graph when an open fault occurs in the third switch (S3) of each of the first to third legs (110-1 to 3).

[0046] Referring to FIGS. 3a and 3b, when an open fault occurs in a switch included in each of the multiple legs (110), a distorted pole voltage waveform is generated for each leg (110). This means that each leg (110) generates a pole voltage waveform independently of each other.

[0047] Accordingly, the single-phase leg diagnostic model learning method and the state diagnostic method of the multi-level inverter (100) of the present disclosure, which utilize the characteristics of such pole voltage waveforms, will be described.

[0048] [Single-phase Leg Diagnosis Model Training Method]

[0049] FIG. 4 is a flowchart of a single-phase leg diagnostic model learning method according to an embodiment of the present disclosure, and

[0050] The single-phase leg diagnostic model learning method of the present disclosure is performed by a processor and can diagnose a multi-level inverter (100) comprising a plurality of legs (110) including a plurality of switches.

[0051] A single-phase leg diagnostic model learning method according to one embodiment of the present disclosure may include steps (a1) to (a3).

[0052] In step (a1) of the present disclosure, the processor measures the pole voltage of a learning target leg (110) which is one of a plurality of legs (110). For example, referring to FIGS. 1 and 4, if the learning target leg (110) is a first leg (110-1), the processor can measure the pole voltage of the first leg (110-1) using a separately provided voltage sensor or the like.

[0053] In step (a2) of the present disclosure, the processor obtains the half-bridge voltage of the leg to be learned (110). To detect multiple switch failures, the half-bridge voltage as well as the pole voltage is required. Therefore, the processor obtains the half-bridge voltage of the leg to be learned (110) and can subsequently use it as training data for a single-phase leg diagnosis model.

[0054] In one embodiment, in step (a2), the processor may measure the half-bridge voltage of the leg to be studied (110). For example, referring to FIGS. 1 and 4, if the leg to be studied (110) is the first leg (110-1), the processor may measure the half-bridge voltage of the first leg (110-1) using a separately provided voltage sensor or the like.

[0055] However, the processor must measure not only the pole voltage but also the half-bridge voltage using a separate voltage sensor. In this case, since a separate sensing circuit is required to measure the half-bridge voltage, there is a problem in that the structure becomes complex due to the sensing circuit and manufacturing costs increase.

[0056] Accordingly, to solve the above-mentioned problem, in one embodiment, at step (a2), the processor can calculate a half-bridge voltage from the pole voltage based on the switch On / Off information of the leg (110) to be learned. Here, the half-bridge voltage is training data for learning a single-phase leg diagnostic model together with the pole voltage.

[0057] Therefore, to train a single-phase leg diagnostic model, it is necessary to calculate the half-bridge voltage in a fault state where a switch has failed and the half-bridge voltage in a normal state where no switch has failed.

[0058] In one embodiment, in step (a2), the processor can calculate the half-bridge voltage when any one of the multiple switches included in the learning target leg (110) is in a faulty state from the pole voltage based on the switch On / Off information of the learning target leg (110) using a previously stored lookup table.

[0059] Specifically, the lookup table may include On / Off information of the switch according to the switch fault state, the pole voltage, and the corresponding half-bridge voltage as shown in Table 1. The processor can use the lookup table to reconstruct and calculate the half-bridge voltage in the switch fault state without directly measuring the half-bridge voltage.

[0060]

[0061] The information included in the lookup table can be calculated in advance through simulation and loaded into the processor's storage unit.

[0062] In addition, in one embodiment, at step (a2), the processor can calculate the half-bridge voltage in a normal state based on the switch On / Off information of the learning target leg (110) using the following mathematical formula 1.

[0063] [Mathematical Formula 1]

[0064] Here, Vhb is the half-bridge voltage, S1 is the On / Off information of the first switch (S1), S5 is the On / Off information of the fifth switch (S5), and Vdc is the DC link voltage of the input terminal.

[0065] That is, the normal half-bridge voltage, in which no switch-opening fault occurs in the learning target leg (110), can be calculated by the state information of the first and fifth switches and the DC link voltage applied to the input terminal.

[0066] Through this, since a separate sensor for measuring half-bridge voltage is not provided, the circuit is simplified, and there is an effect of reducing the amount of training data for the single-phase leg (110) diagnostic learning model compared to the conventional technology.

[0067] In step (a3) ​​of the present disclosure, the processor trains a single-phase leg diagnostic model that diagnoses the switch status of the leg to be studied (110) based on the pole voltage, half-bridge voltage, and state information of the switch of the leg to be studied (110). Specifically, the processor can train the single-phase leg diagnostic model based on the pole voltage measured in step (a1), the half-bridge voltage calculated in step (a2), and switch status information such as the fault / normal state of the switch of the leg to be studied (110) accordingly.

[0068] Next, a single-phase leg diagnostic model according to one embodiment of the present disclosure will be described in detail.

[0069] A single-phase leg diagnostic model according to one embodiment of the present disclosure can receive learning data, such as the pole voltage and half-bridge voltage of a leg (110) to be learned and the state information of a switch of a leg (110) to be learned, and perform learning using a neural network.

[0070] Figures 5a and 5b are block diagrams of a single-phase leg diagnostic model according to one embodiment.

[0071] In step (a3) ​​of the present disclosure, the processor may input the pole voltage and half-bridge voltage and the state information of the switch of the learning target leg (110) corresponding to the correct answer for each case into the neural network of the single-phase leg diagnosis model.

[0072] In one embodiment, the single-phase leg diagnostic model may be an adaptive model in which the size or characteristics of the filter are dynamically adjusted. Specifically, referring to FIG. 5a, the single-phase leg diagnostic model includes a neural network, and the processor may adjust the size or characteristics of the filter provided in the neural network of the single-phase leg diagnostic model differently according to the pole voltage and half-bridge voltage, which are the training data.

[0073] Here, the neural networks include CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), ANN (Artificial Neural Network), DNN (Deep Neural Network), MLP (Multi-Layer Perceptron), Transformer, GAN (Generative Adversarial Network), DBN (Deep Belief Network), and RBM (Restricted). It can be any one of Boltzmann Machine), SOM (Self-Organizing Map), Autoencoder, Capsule Network, and Spiking Neural Network (SNN).

[0074] For example, the neural network illustrated in FIG. 5a is an Adaptive-CNN and may include multiple lightweight convolutional blocks combined with DSConv (Depthwise Separable Convolution) layers, ReLU layers, and Maxpooling layers. The processor may pass the pole voltage and half-bridge voltage of the training target leg (110) through multiple lightweight convolutional blocks to convert them into multiple feature maps.

[0075] Through this, high performance can be maintained while using fewer filters than existing neural networks, and fault detection performance can be maximized by automatically generating optimal filters for each fault type.

[0076] In addition, as one embodiment, the single-phase leg diagnostic model can be optimized using a Reinforced Attention Optimizer (RAO).

[0077] Specifically, referring to Fig. 5b, the single-phase leg diagnosis model may include a reinforcement attention optimizer. The reinforcement attention optimizer is a technique that optimizes the attention mechanism based on reinforcement learning (RL). In other words, it is a method in which the single-phase leg diagnosis model learns which features are important for switch fault diagnosis during the learning process and assigns weights accordingly. Conventional attention mechanisms assign weights to all input data in the same way. However, the reinforcement attention optimizer can form an optimal learning structure by continuously adjusting the weights for high-importance features as the neural network learns.

[0078] For example, the processor can perform Reinforced Attention Optimization (RAO) on each of the multiple feature maps generated in Adaptive-CNN, removing low-importance feature maps by lowering their weights, and emphasizing high-importance feature maps.

[0079] The processor can output state information of the switch of the predicted target leg (110) using the Reinforced Attention Optimization (RAO) technique of the single-phase leg diagnosis model.

[0080] The processor can calculate the difference between the state information of the switch of the predicted learning target leg (110) and the state information of the switch corresponding to the correct answer using the loss function of the single-phase leg diagnosis model. The processor can repeatedly perform learning by inputting the difference between the state information of the switch of the predicted learning target leg (110) calculated through the loss function and the state information of the switch corresponding to the correct answer back into the single-phase leg diagnosis model.

[0081] Through this, unnecessary fault signals can be effectively eliminated while emphasizing the most important fault signals via a reinforcement learning-based optimization process. Additionally, high accuracy can be maintained even with a small number of data samples, and computational load is reduced.

[0082] In another embodiment, the single-phase leg diagnosis model may include an Adaptive CNN and be optimized using a Reinforced Attention Optimizer. That is, the single-phase leg diagnosis model can learn the pole voltage and half-bridge voltage of the leg (110) to be learned by simultaneously using the aforementioned Adaptive CNN and the Reinforced Attention Optimizer (RAO). Since this embodiment has been described above, a detailed description is omitted.

[0083] [Method for diagnosing the condition of a multi-level inverter (100)]

[0084] FIG. 6 is a flowchart of a state diagnosis method for a multi-level inverter (100) according to an embodiment of the present disclosure.

[0085] A method for diagnosing the state of a multi-level inverter (100) according to one embodiment of the present disclosure is performed by a processor, and the state of a multi-level inverter (100) including a plurality of legs (110) can be diagnosed using a single-phase leg diagnosis model according to one embodiment of the present disclosure described above.

[0086] A method for diagnosing the state of a multi-level inverter (100) according to one embodiment of the present disclosure may include steps (b1) to (b3).

[0087] In step (b1) of the present disclosure, the processor receives the pole voltage of each of the plurality of legs (110). For example, referring to FIGS. 1 and 6, the processor can measure the pole voltage of each of the plurality of legs (110) using a separately provided voltage sensor or the like.

[0088] In step (b2) of the present disclosure, the processor obtains the half-bridge voltage of each of the plurality of legs (110). In one embodiment, in step (a2), the processor may measure the half-bridge voltage of each of the plurality of legs (110). For example, referring to FIGS. 1 and 6, the processor may measure the half-bridge voltage of the first to third legs (110-1 to 3) using a separately provided voltage sensor or the like.

[0089] However, in this case, since a separate sensing circuit is required to measure the half-bridge voltage, there is a problem in that the structure becomes complex due to the sensing circuit and manufacturing costs increase.

[0090] Accordingly, to solve the above problem, in one embodiment, at step (b2), the processor can calculate a half-bridge voltage from the pole voltage based on the switch On / Off information of each of the plurality of legs (110).

[0091] In one embodiment, in step (b2), the processor can calculate the half-bridge voltage of each of the plurality of legs (110) based on the pole voltage of each of the plurality of legs (110) and the switch On / Off information of each of the plurality of legs (110) using a previously stored lookup table such as Table 1.

[0092] Through this, the half-bridge voltage can be calculated and reconstructed using only the pole voltage and switch On / Off information via a pre-prepared lookup table without directly measuring the half-bridge voltage. This simplifies the circuit by eliminating the need for a separate voltage sensing circuit, and reduces the amount of computation by not requiring a computational process for processing analog signals. The information included in such a lookup table can be calculated in advance through simulation and stored in the processor's storage unit.

[0093] In step (b3) of the present disclosure, the processor diagnoses the state of the multi-level inverter (100) based on the pole voltage and half-bridge voltage of each of the plurality of legs (110) using a plurality of pre-learned single-phase leg diagnostic models. Additionally, in step (b3) of the present disclosure, the plurality of single-phase leg diagnostic models are matched one by one for each of the plurality of legs (110) to diagnose the state of the multi-level inverter (100) in parallel.

[0094] For example, referring to FIG. 1, one first to third single-phase leg diagnostic model can be matched to each of the first to third legs (110-1 to 3). For example, the first single-phase leg diagnostic model can diagnose the switch status of the first leg (110-1). These first to third single-phase leg diagnostic models operate in parallel, and multiple single-phase leg diagnostic models can simultaneously diagnose the switching status of each leg (110).

[0095] In one embodiment, in step (b3), the processor can receive the pole voltage and half-bridge voltage of the leg (110) matched to each of the plurality of learning models and diagnose the fault status of each switch included in the leg (110) matched to each of the plurality of learning models.

[0096] The first leg (110-1) and the first single-phase leg diagnostic model are described as examples. The processor inputs the pole voltage of the first leg (110-1) and the half-bridge voltage of the first leg (110-1) into the first single-phase leg diagnostic model. Accordingly, the processor obtains the predicted switch failure state of the first leg from the first single-phase leg diagnostic model. Through this, the failure state of the switch of the first leg (110-1) can be diagnosed.

[0097] In one embodiment, the single-phase leg diagnostic model can diagnose and output the switch failure status of one leg (110) by labeling individual switch failures and combinations of multiple switch failures that may occur in one leg (110) as shown in Table 2.

[0098]

[0099] FIG. 7 is a flowchart of a single-phase leg diagnostic model learning method and a state diagnostic method of a multi-level inverter (100) according to another embodiment.

[0100] The Offline Training Model on the left side of FIG. 7 is a single-phase leg diagnosis model learning method according to one embodiment of the present disclosure, and the Online Fault Diagnosis Process on the right side of FIG. 7 is a flowchart illustrating a state diagnosis method of a multi-level inverter (100) according to one embodiment of the present disclosure.

[0101] In one embodiment, after step (a2), the processor may normalize and perform window slicing on the pole voltage and half-bridge voltage of the leg (110) to be learned. Also, in one embodiment, after step (b2), the processor may normalize and perform window slicing on the pole voltage and half-bridge voltage of each of the plurality of legs (110).

[0102] Referring to the Offline Training Model and Online Fault Diagnosis Process in Fig. 7, normalization is essential when a neural network is trained to maintain a constant data distribution, improve training speed, and aid in stable optimization. Additionally, window slicing is a method of analyzing continuous signal data by dividing it into small time intervals (windows). This is done to configure data segments so that changes over time can be effectively learned.

[0103] [Effects of the present disclosure]

[0104] Referring to Table 3, conventionally, the failures of each switch were combined to train the learning model. Therefore, in the case of multi-level inverters with 4 or more levels, the number of switch failure combinations increased significantly, requiring training with nearly 1,000 samples. Consequently, there was a problem where the amount of computation required for training data increased significantly depending on the number of switches, resulting in a large consumption of computational resources.

[0105] On the other hand, in the case of the present disclosure, samples of switch failure combinations for a single-phase learning target leg (110) are produced, so a single-phase leg diagnosis model can be trained with only about 4.1% of the samples compared to the number of conventional samples.

[0106]

[0107] Figures 8a to 8d are graphs of exemplary pole voltages and half-bridge voltages.

[0108] Specifically, FIGS. 8a to 8d are voltage graphs comparing the reconstructed half-bridge voltage, which is the half-bridge voltage calculated from the pole voltage based on switch On / Off information, and the actual measured half-bridge voltage in steps (a2) and (b2) of the present disclosure.

[0109] To explain each figure in more detail, Figure 8a is a graph of the reconfigured half-bridge voltage calculated in a normal state where no fault occurs in the switch and the measured half-bridge voltage. In addition, Figures 8b, c, and d are graphs of the reconfigured half-bridge voltage calculated in a fault state where a fault occurs in the first to third switches, respectively and the measured half-bridge voltage.

[0110] Referring to Figures 8a through 8d, the voltage steps of the waveforms of the reconstructed half-bridge voltage and the measured half-bridge voltage in the normal state and fault state are repeated consistently. Therefore, it can be confirmed that the reconstructed half-bridge voltage and the measured half-bridge voltage waveforms are similar to each other.

[0111] Figures 9a to f are graphs of the pole voltage, half-bridge voltage, and switch failure occurrence flag according to the switch failure.

[0112] FIG. 9a illustrates the pole voltage, half-bridge voltage, and switch failure occurrence flag of each leg (110) when an open fault occurs in the third switch (S3) of the first leg (110-1).

[0113] Referring to Fig. 9a, when an open fault occurs in the third switch (S3) of the first leg (110-1), a fault flag is generated within approximately 75 ms, which has the effect of enabling rapid fault detection with high accuracy.

[0114] FIG. 9b illustrates the pole voltage, half-bridge voltage, and switch failure occurrence flag of each leg (110) when an open fault occurs at the third switch (S3) of the first leg (110-1) and the second switch (S2) of the second leg (110-2).

[0115] Referring to FIG. 9b, when an open fault occurs simultaneously in the third switch (S3) of the first leg (110-1) and the second switch (S2) of the second leg (110-2), a fault flag for the fault of the two switches is generated within approximately 75 ms, so that the fault can be detected quickly with high accuracy even in the case of two open switch faults.

[0116] FIG. 9c illustrates the pole voltage, half-bridge voltage, and switch failure occurrence flag of each leg (110) when an open fault occurs simultaneously at the third switch (S3) of the first leg (110-1), the second switch (S2) of the second leg (110-2), and the first switch (S1) of the third leg (110-3).

[0117] Referring to FIG. 9c, when an open fault occurs simultaneously in the third switch (S3) of the first leg (110-1), the second switch (S2) of the second leg (110-2), and the first switch (S1) of the third leg (110-3), a fault flag for the fault of the three switches is generated within approximately 75 ms, thereby enabling rapid fault detection with high accuracy even in the case of three open switch faults.

[0118] FIG. 9d illustrates the pole voltage, half-bridge voltage, and switch failure flag of each leg (110) when an open fault occurs in the first and third switches of the first leg (110-1).

[0119] That is, unlike FIGS. 9a to c, FIG. 9d is a case where multiple switch open failures occur in a single leg (110). In this case, a fault flag for the failure of two switches is generated within approximately 100 ms from the occurrence of the failure. Therefore, there is an effect of being able to quickly detect the failure with high accuracy even in the case of multiple switch open failures included in a single leg (110).

[0120] FIG. 9e illustrates the pole voltage, half-bridge voltage, and switch failure flag of each leg (110) when the speed of the motor (200) is variable, and FIG. 9f illustrates the pole voltage, half-bridge voltage, and switch failure flag of each leg (110) when the load (300) connected to the motor (200) is variable.

[0121]

[0122] Referring to FIG. 9e and Table 4, when the speed of the motor (200) is varied to 300, 800, and 1200 [rpm], a high switch fault diagnosis accuracy of 99.30 [%] is obtained with an appropriate number of samples. Also, referring to FIG. 9f and Table 4, when the load (300) is varied to 0, 3, and 5 [N·m], a high switch fault diagnosis accuracy of 99.46 [%] is obtained. Additionally, when the speed of the motor (200) and the load (300) are varied simultaneously, a high switch fault diagnosis accuracy of 99.58 [%] is obtained.

[0123] In conclusion, the present disclosure utilizes a single-phase leg diagnostic model, which not only enables rapid switch fault diagnosis with a smaller number of samples compared to conventional methods but also has the effect of improving the accuracy of switch fault diagnosis.

[0124] A processor refers to a central processing unit (CPU) responsible for computational tasks and data processing. The processor interprets and executes instructions, and processes program tasks by performing various calculations and logic operations. The processor may be a microcontroller and can perform tasks according to algorithms of a program stored on a storage medium (not shown). The processor can output calculated values ​​through user interface devices such as display devices.

[0125] The technical concept of this disclosure should not be interpreted as being limited to the embodiments described above. Not only is the scope of application diverse, but various modifications are possible at the level of a person skilled in the art without departing from the essence of this disclosure as claimed in the claims. Accordingly, such improvements and modifications fall within the scope of protection of this disclosure insofar as they are obvious to a person skilled in the art. Explanation of the symbols

[0127] 100: Multi-level inverter 110 : Leg 110-1~110-3 : 1st to 3rd legs S1~S6: 1st to 6th switches 120: Series capacitor circuit C1~C3: 1st to 3rd capacitors 200 : Motor 300 : Load

Claims

Claim 1 A method for learning a single-phase leg diagnostic model for diagnosing a multi-level inverter comprising a plurality of legs including a plurality of switches, performed by a processor, wherein the processor comprises: (a1) a step of measuring a pole voltage of a leg to be learned, which is one of the plurality of legs; (a2) a step of obtaining a half-bridge voltage of the leg to be learned; and (a3) ​​a step of learning a single-phase leg diagnostic model that diagnoses the switch state of the leg to be learned based on the pole voltage, the half-bridge voltage, and state information of the switch of the leg to be learned; wherein, in step (a2), the processor calculates a half-bridge voltage from the pole voltage based on the switch On / Off information of the leg to be learned, and the single-phase leg diagnostic model comprises an Adaptive CNN in which the size or characteristics of the filter are dynamically adjusted, and is optimized using a Reinforced Attention Optimizer (RAO). Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 delete Claim 6 A method for diagnosing the state of a multi-level inverter comprising a plurality of legs using a single-phase leg diagnostic model of claim 1, wherein the processor comprises: (b1) receiving a pole voltage of each of the plurality of legs; (b2) obtaining a half-bridge voltage of each of the plurality of legs; and (b3) diagnosing the state of the multi-level inverter based on the pole voltage and half-bridge voltage of each of the plurality of legs using a plurality of pre-learned single-phase leg diagnostic models; wherein in step (b3), the plurality of single-phase leg diagnostic models are matched one by one for each of the plurality of legs to diagnose the state of the multi-level inverter in parallel; and in step (b2), the processor calculates the half-bridge voltage from the pole voltage based on the switch On / Off information of the leg to be learned. Claim 7 delete Claim 8 A method for diagnosing the state of a multi-level inverter according to claim 6, wherein in step (b3), the processor receives the pole voltage and half-bridge voltage of the matched leg for each of the plurality of single-phase leg diagnostic models and diagnoses the fault state of each switch included in the leg matched to each of the plurality of single-phase leg diagnostic models.

Images