Defect detection device for basket-type rotor using AI

Abstract

The purpose of the present invention is to provide a defect detection device for a basket-type rotor, said defect detection device being capable of easily and reliably estimating and reporting the presence or absence of a broken bar or an air hole (defect), which is likely to be present in a conductor rod formed in a locking slot, using an AI without requiring a skilled person. The scale of the defect and the position of the defect can be estimated and reported. This defect detection device for a basket-type rotor is provided with: a fluctuating magnetic field generation means (1) and a detection means (2) that are disposed so as to be able to straddle one conductor rod of a basket-type rotor (R); a rotor model data output means (3) for creating rotor models (M1-Mn) and outputting configuration data (K1-Kn) of the rotor models (M1-Mn); a rotor model analysis means (4) for receiving the structure data (K1-Kn) and AC power source data (AC), performing numerical analysis on the current value induced in the second coil, and generating a large amount of learning data (L1-Ln); and a neural network (5) and a defect information reporting means (6) that are optimized by the learning data (L1-Ln) provided in advance, receive the time-series current value data (C) detected by the detection means (2), perform inverse analysis, and output defect state estimation information indicating the presence or absence of a defect, the scale, and the position of the conductor bar existing in the basket-type rotor (R).

Description

Technical Field

[0001] The present invention relates to a defect detection device for a basket-type rotor using AI that can infer the state (presence or absence of defects, size or location) of pores or broken bars (hereinafter referred to as "defects") that may exist in conductor bars constituting a basket-type rotor and report the inference results. Background Art

[0002] In the induction motor used in the drive unit of electric vehicles and construction machinery, a basket-shaped rotor is used, which is composed of a rotor core, conductor bars formed by flowing molten metal of an electrically conductive metal material such as aluminum and copper into a plurality of holes passing through the outer peripheral surface of the rotor core, and a pair of end rings connecting all the conductor bars at both ends of the rotor core.

[0003] Moreover, defects in the conductor bars and part of the end rings of the basket-type rotor and poor connections in the connecting parts between the conductor bars and the end rings sometimes significantly affect the performance and life of the induction motor, so it is necessary to check these defects and poor connections. The connecting parts between the end rings and the conductor bars and the end rings are relatively easy to check because they are exposed to the outside of the rotor core.

[0004] However, when the molten metal flows into the conductor bar, air is drawn in, and when the conductor bar is solidified as it is, defects may occur inside the conductor bar. If the defect is close to a broken bar, the performance of the induction motor may be affected. Even if it is a pore, depending on the size and shape, it may affect the performance and life of the induction motor. Moreover, pores that shorten the life are difficult to detect in the initial test, so they may cause the production of induction motors of poor quality.

[0005] In particular, when the through holes of the rotor core are closed slots that do not open on the outer peripheral surface, since the conductor bars cannot be visually identified from the outside, operators cannot visually check for defects or measure the resistance of each conductor bar using a tester or the like.

[0006] In order to solve such a problem, Patent Document 1 (Japanese Patent Publication No. 9-37527) describes a defect detection device (40) having an iron core (46) wound with a first coil (42) connected to an AC power source (48) and an iron core (50) wound with a second coil (44) connected to a processing device (52), wherein the iron core (46) and the iron core (50) are respectively provided with a pair of protruding teeth (54) and protruding teeth (56) that can be opposed to the rotor core (18) of the basket rotor (12) (see, in particular, paragraphs 0022 and Figure 4 ).

[0007] In the defect detection device (40), the secondary current induced in the conductor bar (22) of the inspection object by the alternating magnetic flux formed by the first coil (42) forms an induced alternating magnetic flux around the conductor bar (22), and a small alternating electromotive force is generated in the second coil (44) due to the induction of the alternating magnetic flux. The alternating electromotive force is detected by the processing device (52), and regardless of the structure of the peripheral through hole (20), it is possible to correctly determine whether the basket-shaped rotor (12) has electrical defects (especially refer to paragraphs 0023 to 0026).

[0008] However, according to the invention described in Patent Document 1, although it is possible to determine whether there are electrical defects (broken bars) in the conductor rod (22), it is impossible to infer whether there are pores, their size and their location, as described in paragraphs 0025 and Figure 3 As shown, when there are multiple conductor rods (22) to be inspected and one of them has an electrical defect, only the amplitude becomes smaller compared to the detection signal when the conductor rod (22) to be inspected has no defects at all, so skilled personnel are required for the judgment.

[0009] Prior art literature

[0010] Patent Literature

[0011] Patent Document 1: Japanese Patent Application Laid-Open No. 9-37527 (Patent No. 3608846) Summary of the invention

[0012] The present invention solves such a problem. The first subject is to provide a defect detection device for a basket-type rotor. The defect detection device for a basket-type rotor can use AI to easily and reliably infer and report the presence of broken bars or air holes (hereinafter referred to as "defects") that may exist inside the conductor rod formed in the locking slot and cannot be inspected by visual inspection, testing machines, etc., without relying on skilled personnel.

[0013] In addition, in the defect detection device for a basket-type rotor, a second problem is to be able to estimate and report the scale of the defect, and a third problem is to be able to estimate and report the position of the defect.

[0014] The invention according to claim 1 for solving the above-mentioned problem provides a defect detection device for a basket-type rotor for estimating and reporting defects existing in a conductor bar of the basket-type rotor, characterized in that the defect detection device for a basket-type rotor comprises:

[0015] a variable magnetic field generating unit, comprising a first U-shaped iron core arranged to straddle one conductor bar of the basket-type rotor, a first coil wound around the first iron core, and an AC power source connected to the first coil;

[0016] a detection unit including a U-shaped second core that can be arranged to straddle one conductor bar of the basket-type rotor, a second coil wound around the second core, and a meter for detecting a current value or a voltage value induced in the second coil;

[0017] a rotor model data output unit for producing a large number of rotor models having different defects existing in one conductor bar, with respect to the basket-type rotor, and outputting structural data related to the large number of rotor models;

[0018] a rotor model analysis unit including a numerical analysis unit, the numerical analysis unit receiving the structural data and the AC power supply data outputted from the rotor model data output unit, and numerically analyzing the current value or the voltage value induced in the second coil when the first core and the second core are arranged across one conductor bar of each rotor model and a predetermined AC is applied to the first coil, the rotor model analysis unit outputting, as learning data, identification information for identifying a rotor model to be subjected to numerical analysis and numerical analysis data obtained by the numerical analysis performed on the rotor model;

[0019] a neural network that receives the time-series current value data or voltage value data detected by the detection unit when a predetermined alternating current is applied to the first coil, performs inverse analysis, and outputs defect estimation information; and

[0020] a defect information reporting unit that reports the presence or absence of defects in the conductor bars of the basket-type rotor based on the defect estimation information output from the neural network,

[0021] The neural network is provided in advance with learning data on the large number of rotor models output from the rotor model analyzing unit, and is optimized through autonomous learning.

[0022] The invention according to claim 2 for solving the above-mentioned problem is characterized in that, in the defect detection device according to the invention according to claim 1,

[0023] The neural network outputs defect scale estimation information instead of defect presence estimation information.

[0024] The defect information reporting unit reports the scale of a defect existing in the conductor bar of the basket-type rotor based on the defect scale estimation information.

[0025] The invention according to claim 3 for solving the above-mentioned problem is characterized in that, in the defect detection device of the invention according to claim 1,

[0026] The variable magnetic field generating unit is arranged on one end side of the basket-type rotor.

[0027] The detection units are sequentially arranged at a plurality of locations set between the other end of the basket-type rotor and the location where the variable magnetic field generating unit is arranged.

[0028] In the numerical analysis unit, when the second core is sequentially arranged at a plurality of locations set between the other end of each rotor model and the location where the first core is arranged and a predetermined alternating current is applied to the first coil, a current value or a voltage value induced in the second coil is numerically analyzed,

[0029] The identification information is information that identifies a rotor model that is a target of numerical analysis and in which the second cores are sequentially arranged at the plurality of locations.

[0030] The neural network receives the current value data or voltage value data of the time series and the configuration position data of the detection unit and performs inverse analysis to output defect presence estimation information and defect position estimation information.

[0031] The defect information reporting unit reports the presence or absence of a defect in the conductor bar of the basket-type rotor and the position of the defect based on the defect estimation information and the defect position estimation information output from the neural network.

[0032] The defect detection device for a basket-type rotor involved in claim 1 performs inverse analysis by inputting time-series current value data or voltage value data detected by a detection unit when a predetermined alternating current is applied to a variable magnetic field generating unit into a neural network, outputs defect inference information, and has a defect information reporting unit that reports the presence or absence of defects in the conductor bars of the basket-type rotor based on the output defect inference information, and the neural network is pre-provided with learning data on a large number of rotor models output from a rotor model analysis unit and is optimized through autonomous learning, so that AI can be used to easily and reliably infer and report the presence or absence of defects that cannot be inspected by appearance inspection, testing machines, etc., which are possible defects that exist inside the conductor bars formed in the locking slots.

[0033] Therefore, the performance degradation of the induction motor due to the defect of the conductor bar can be estimated and reported in advance before assembling the induction motor, so the performance degradation of the basket rotor can be prevented, the product quality can be improved, and the occurrence of various failures can be suppressed.

[0034] In addition to the above-mentioned effects of the invention involved in claim 1, the defect detection device for the basket-type rotor involved in claim 2 has a neural network that outputs defect scale estimation information instead of defect prediction information, and a defect information reporting unit reports the scale of the defect based on the defect scale estimation information, so it is possible to estimate and report the scale of defects in the conductor bars of the basket-type rotor.

[0035] In addition to the above-mentioned effects of the invention involved in claim 1, the defect detection device for the basket-type rotor involved in claim 3 has a neural network that outputs defect inference information and defect location inference information, and a defect information reporting unit reports the presence or absence of defects and their locations based on the defect inference information and defect location inference information, so the locations of defects in the conductor bars of the basket-type rotor can also be inferred and reported. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a block diagram of the basket-type rotor defect detection device involved in Example 1.

[0037] Figure 2 This is a diagram showing an example of a columnar defect simulating a broken bar.

[0038] Figure 3 This is a diagram showing an example of a semi-cylindrical defect simulating a large pore.

[0039] Figure 4 This is a diagram showing an example of a semi-columnar defect simulating a small pore.

[0040] Figure 5 This is a diagram for explaining the principle of generating an induced current in the second coil of the detection unit.

[0041] Figure 6 This is a flowchart showing an algorithm of the basket-type rotor defect detection device according to the first embodiment.

[0042] Figure 7 20 cases of patterns in scale and positions of defects are shown in the table.

[0043] Figure 8 It is shown that Figure 7 FIG. 2 is a graph of numerical analysis data (time-series current value data) of the induced current generated in the second coil in the defective rotor model for 20 cases.

[0044] Fig. 9 This is a table showing the estimated results of the scale and position of the defect obtained by inverse analysis of the verification data.

[0045] Fig.10 This is a block diagram of a basket-type rotor defect detection device according to the second embodiment.

[0046] Fig.11 This is a flowchart showing an algorithm of the basket-type rotor defect detection device according to the second embodiment. DETAILED DESCRIPTION

[0047] Hereinafter, embodiments of the present invention will be described based on examples.

[0048] Example 1

[0049] Figure 1 This is a block diagram of the basket-type rotor defect detection device involved in Example 1.

[0050] like Figure 1 As shown, the defect detection device for the basket-type rotor involved in Example 1 is used to infer and report the presence, scale and location of possible defects inside the conductor bars of the basket-type rotor R of the induction motor (specifically, a structure having conductor bars near the outer circumference of the cylindrical rotor core and all the conductor bars connected by a pair of end rings at both ends of the rotor core), and is composed of the following units.

[0051] (1) A variable magnetic field generating unit 1 includes a U-shaped first core arranged to straddle one conductor bar of a basket-type rotor R, a first coil wound around the first core, and an AC power source connected to the first coil.

[0052] (2) A U-shaped second core that can be arranged to straddle one conductor bar of the basket-type rotor R, a second coil wound around the second core, and a detection unit 2 that detects a current value induced in the second coil.

[0053] In addition, the materials and shapes of the rotor core, conductor bars and end rings, and the first and second cores constituting the basket-type rotor R, as well as the resistivity, wire diameter, number of turns, etc. of the first and second coils are known in advance, and the AC power supply is single-phase 100V, 50Hz.

[0054] (3) Regarding the basket-type rotor R, a rotor model data output means 3 is provided to prepare a large number of rotor models M1 to Mn having different defects in one conductor bar, and to output structural data K1 to Kn related to the rotor models M1 to Mn.

[0055] (4) A numerical analysis unit is provided, which receives the construction data K1~Kn and the AC power supply data AC outputted from the rotor model data output unit 3, and when the variable magnetic field generating unit 1 and the detecting unit 2 are arranged across one conductor bar of each rotor model and a predetermined AC is applied to the first coil, the current value induced in the second coil is numerically analyzed, and with respect to the rotor model M1, the determination information k1 and the numerical analysis data A1 obtained by the numerical analysis are outputted as the learning data L1. Similarly, with respect to the rotor model Mn, the rotor model analysis unit 4 outputs the determination information kn and the numerical analysis data An obtained by the numerical analysis as the learning data Ln.

[0056] (5) When a predetermined alternating current is applied to the variable magnetic field generating unit 1, a neural network 5 receives the time-series current value data C detected by the detecting unit 2 and performs inverse analysis to output defect state estimation information indicating the presence, size and position of defects in the conductor bars of the basket-type rotor R.

[0057] (6) A defect information display unit 6 that displays the presence, size, and position of a defect in the conductor bar of the basket-type rotor R based on the defect state estimation information output from the neural network 5 .

[0058] Hereinafter, each unit will be described in detail.

[0059] (Variation magnetic field generating unit 1)

[0060] The variable magnetic field generating unit 1 is provided at one end of the basket-shaped rotor R ( Figure 1 A U-shaped first iron core is arranged on the right side of the rotor so as to cross a conductor bar, and a predetermined alternating current (100V, 10A, 50Hz) is applied to the first coil from an AC power supply. Through the excitation current flowing through the first coil, an alternating magnetic flux unit that interlinks with a conductor bar is formed between the first iron core and the rotor core.

[0061] The U-shaped first iron core may be arranged at any position of the basket-type rotor R as long as it spans one conductor bar, but in the first embodiment, it is arranged at one end side as an example.

[0062] In addition, when the predetermined alternating current becomes a high voltage and a large current, the estimation accuracy can be improved, so the first coil has a thick structure and is preferably wound as much as possible.

[0063] (Detection unit 2)

[0064] The detection unit 2 is on the other end side of the basket-type rotor R ( Figure 1 On the left side of the device, a second U-shaped iron core is arranged so as to cross one conductor rod in the same manner as the first iron core, and an ammeter connected to the second coil is used to detect the current value, and a unit of time-series current value data C induced in the second coil is detected (the principle of induced current in the second coil will be described later).

[0065] In addition, the second core may be arranged at any position of the basket-type rotor R as long as it straddles the one conductor bar, but in the first embodiment, it is fixedly arranged at the other end side for ease of handling.

[0066] In addition, the timing current induced in the second coil is relatively small, so the wire used in the second coil can be thinner than that in the first coil, and it is best to increase the number of turns.

[0067] (Rotor model data output unit 3)

[0068] In the rotor model data output unit 3, rotor models M1 to Mn with different internal structures, i.e., different sizes and positions of defects (including the case of no defects) are prepared by randomly arranging one conductor bar to simulate the following. Figures 2 to 4 The model of a defect of various sizes of semi-cylindrical or cylindrical shape with broken strips or pores is shown.

[0069] Then, as the scale of defects, 500 patterns of the following (1) to (4) were prepared.

[0070] (1) The cross section of the conductor bar with cylindrical defects is almost completely cut (less than 0.1mm 2 ) (shown as a "broken bar" in the table) is called the 31 mode ( Figure 2 is an example).

[0071] (2) The cross-sectional area of ​​the remaining semi-cylindrical defective conductor bar having the same radius as the thickness of the conductor bar is 0.1 mm 2 Above and less than 8mm 2 The situation (shown as "small" in the table) is called the 192 mode ( Figure 3 is an example).

[0072] (3) The cross-sectional area of ​​the remaining semi-cylindrical defective conductor bar with a radius half the thickness of the conductor bar is 8 mm 2 Above and less than 20mm 2 The situation (shown as "large" in the table) is called the 221 mode ( Figure 4 is an example).

[0073] (4) The case where there is no defect at any position of the conductor bar (shown as "no defect" in the table) is called the 56 mode.

[0074] In addition, all 500 patterns of the following (a) to (c) were prepared as defect positions.

[0075] (a) The case where the defect is directly below the detection unit (shown as "directly below" in the table) is called the 65 mode.

[0076] (b) The case where the defect is not directly below the detection unit (shown as "outside" in the table) is called the 379 mode (inside Figures 2 to 4 This mode is used in all (

[0077] (c) The case where there is no defect at any position of the conductor bar (shown as "no defect" in the table) is called the 56 mode.

[0078] In addition, all 500 modes of the rotor model in Example 1 are divided into any modes among the modes of (1) and (a), (1) and (b), (2) and (a), (2) and (b), (3) and (a), (3) and (b), and (4) and (c).

[0079] In addition, structural data K1 to Kn corresponding to each rotor model on a one-to-one basis are output.

[0080] (Rotor model analysis unit 4)

[0081] In the rotor model analysis means 4 of the first embodiment, an analysis means using a finite element method (hereinafter referred to as “FEM electromagnetic analysis means”) is used as a numerical analysis means.

[0082] According to the FEM electromagnetic analysis unit, when the rotor model and the first core and the second core arranged across a conductor bar of the basket-type rotor R are cut into a large number of tetrahedral meshes, and various conditions (the magnetic permeability of the magnetic body part of the rotor model, the magnetic permeability of the surface part, the magnetic permeability and resistivity of the conductor bar, the magnetic permeability of the first core and the second core, and the number of turns of the first coil and the second coil) are set, assuming that a variable magnetic field generating unit 1 and a detecting unit 2 are provided in the rotor model, by applying a predetermined alternating current (100V, 10A, 50Hz) to the first coil, the time series data of the induced current generated in the second coil can be obtained.

[0083] Here, the principle of generating current in the second coil is as follows: Figure 5 As shown in FIG. 1 , when a predetermined alternating current 1p is applied to the first coil 1c, a changing magnetic field is generated around the conductor bar Rc across the first core 1i. A current I(t) is induced in the conductor bar Rc, a changing magnetic field is generated in the second core by the current I(t), and an induced current is generated in the second coil by the changing magnetic field.

[0084] Moreover, by utilizing the rotor model data output unit 3, rotor models M1 to Mn having different internal structures (scale and location of defects) are produced. When the structural data K1 to Kn and the AC power supply data AC related to each rotor model are output, by configuring the variable magnetic field generating unit 1 and the detection unit 2 in a conductor bar of each rotor model and applying a predetermined AC to the first coil, numerical analysis data of the induced current generated in the second coil can be obtained. Therefore, a large amount of learning data L1 to Ln consisting of the determination information k1 to kn for determining each rotor model and the numerical analysis data A1 to An obtained about each rotor model can be prepared in a short time.

[0085] (Neural Network 5)

[0086] Neural Network 5 Figure 1 As shown, it includes an input layer 5I, an output layer 5O, a learning data storage unit 5L, a learning data input unit 5T that inputs a large amount of learning data L1~Ln to the learning data storage unit 5L, and multiple intermediate layers not shown.

[0087] In addition, various layers studied, such as a convolutional layer, a pooling layer, and a fully connected layer, are combined as needed to form an intermediate layer.

[0088] Furthermore, the neural network 5 accumulates a large amount of learning data L1 to Ln in the learning data accumulation unit 5L and is optimized by autonomous learning. That is, by using parameters optimized using a sufficient amount of learning data (10,000 to several hundred thousand depending on the size and position of the defect to be estimated), an estimation accuracy of 95% or more can be obtained.

[0089] On this basis, when the time-series current value data C actually output from the detection unit 2 is input to the input layer 5I, the neural network 5 performs inverse analysis and outputs defect state estimation information indicating the presence, scale and position of defects in the conductor bars of the basket-type rotor R from the output layer 5O.

[0090] (Defect information display unit 6)

[0091] The defect information display unit 6 displays the presence, scale and position of defects in the conductor bars of the basket-type rotor R based on the defect state inference information output from the neural network 5, and numbers the line segments simulating each conductor bar in a usual display manner, and displays the scale and position of the defects at the corresponding position of each line segment.

[0092] Furthermore, it is also possible to select a method of displaying numbers identifying which conductor rods and characters, symbols, or numbers indicating the size and position of defects as a table.

[0093] Figure 6 This is a flow chart showing the algorithm of the defect detection device for the basket-type rotor involved in Example 1. In the following processes (T1) to (T4), a large amount of learning data L1 to Ln are generated and accumulated in the neural network 5 for autonomous learning. On this basis, in the processes (1) to (6), it is estimated whether there are defects, scale and position of the conductor bars of the basket-type rotor R, and the estimated presence or absence of defects, scale and position are displayed.

[0094] (T1) A large number of rotor models (M1 to Mn) are produced based on the materials and shapes of the rotor core, conductor bars, end rings, etc. of the basket-type rotor R known in advance.

[0095] (T2) Create structural data K1 to Kn of each rotor model.

[0096] (T3) Using the FEM electromagnetic analysis unit, based on each structural data K1~Kn and the AC power supply data AC, when a predetermined AC is applied to the first coil in each rotor model, numerical analysis data A1~An representing time series data of the current value induced in the second coil is output.

[0097] (T4) Learning data L1 to Ln consisting of identification information k1 to kn for identifying each rotor model and the numerical analysis data A1 to An outputted at T3 are generated and stored in the neural network 5.

[0098] (1) A predetermined alternating current is applied to the variable magnetic field generating unit 1 disposed on one end side of the basket-type rotor R.

[0099] (2) When a predetermined alternating current is applied to the first coil of the varying magnetic field generating means 1 , the time-series current value data C induced in the second coil of the detecting means 2 is detected.

[0100] (3) The detected time-series current value data C is sent to the neural network 5.

[0101] (4) The neural network 5 performs inverse analysis based on the input time-series current value data C.

[0102] (5) Defect state estimation information indicating the presence, size, and position of defects in the conductor bars of the basket-type rotor R is output by inverse analysis and sent to the defect information display unit 6 .

[0103] (6) The defect information display unit 6 displays the presence, size, and position of a defect in the conductor bar of the basket-type rotor R based on the defect state estimation information.

[0104] If there are conductor rods that have not been estimated and displayed, the variable magnetic field generating unit 1 and the detecting unit 2 are moved across the conductor rods to execute the above-mentioned processes (1) to (6), and the estimated defect presence, size and position are displayed for all the conductor rods before the process ends.

[0105] Figure 7 is a table showing 20 cases related to the pattern of defect size and the pattern of location, Figure 8 About Figure 7 The rotor models of Cases 1 to 20 shown are graphs of numerical analysis data (time-series current value data) of the induced current generated in the second coil.

[0106] Figure 8Among the graphs shown, the dashed line graph with the smallest amplitude is the numerical analysis data of the rotor model having the defect of situation 7 (a broken bar that is not directly below the detection unit 2), and the double-dotted line graph with the second smallest amplitude is the numerical analysis data of the rotor model having the defect of situation 8 (a broken bar that is not directly below the detection unit 2).

[0107] Furthermore, the reason why the graphs are different despite the same defect of a broken bar not being located directly below the detection unit 2 is that the degree of the broken bar is different.

[0108] The double-dashed line graph with the third smallest amplitude is for a defect with case 20 (a cross-sectional area of ​​0.1 to 8 mm not directly below the detection unit 2). 2 ) of the rotor model. The fourth smallest dashed line in the graph is about the defect with case 9 (the cross-sectional area is 0.1 to 8 mm not directly below the detection unit 2). 2 )’s rotor model.

[0109] Regarding these two graphs, although the cross-sectional area not directly below the detection unit 2 is the same, it is 0.1 to 8 mm 2 The reason why the graphs are different is that the cross-sectional areas remaining in the conductor rods are different.

[0110] Furthermore, the solid line graph with the largest amplitude is the numerical analysis data of the defective (undamaged) rotor model as Case 1, and the single-point dashed line graph of the defective (undamaged) rotor model as Case 4 completely overlaps with the graph of Case 1.

[0111] As for the other graphs, the remaining cross-sectional area of ​​the conductor bar is relatively large, so the difference in amplitude change is small. Figure 8 It is difficult to distinguish in the graph, but it actually becomes a different graph.

[0112] In Example 1, a total of 500 rotor models with various defect sizes and locations (including intact) are produced, the numerical analysis data of the rotor models for 450 patterns are used for learning, and the numerical analysis data of the rotor models for the remaining 50 patterns are used for verification.

[0113] Fig. 9 This is a table showing defect state estimation information (probability of defect scale and location) obtained by inputting 25 of the numerical analysis data used for verification (25 sets of time-series current value data) into the neural network 5 and inversely analyzing them, as well as the results obtained by comparing the estimated state with the actual state.

[0114] In addition, the records of "Broken Bar", "Small", "Large", "Uninjured", "Outside", "Directly Below" and "Uninjured" in the speculation column indicate that the probability of each is greater than 75%. Regarding "Small Break", "Small and Large", "Large Break", "Big and Small", "No Big", "Outside and Below" and "No Outside", when there is no example with a probability of more than 75%, the example with the highest probability is recorded on the left and the example with the second highest probability is recorded on the right.

[0115] Among them, the "duan" in "xiaoduan" and "daduan" means "broken strips", the "xia" in "waixia" means "directly below", and the "wu" in "wuda" and "wuwai" means "uninjured".

[0116] Moreover, when comparing the inferred defect scale (area of ​​the remaining part) with the actual defect scale, 14 were completely consistent "◎", 9 were consistent "○" (although the probability was not greater than 75%, the part with the maximum probability was consistent), and 2 were partially consistent "△" (the maximum probability was less than 75% and the part with the second highest probability was consistent). Even if the secondary consistency was removed, the accuracy rate was still 92%.

[0117] Furthermore, since all parts with a probability of 75% or more are completely consistent and not removed at all, the extent of damage can be confirmed, and even with only 450 learning data (450 sets of time-series current value data), it can be estimated with high accuracy.

[0118] Next, when trying to compare the inferred defect locations with the actual defect locations, 13 were completely consistent "◎", 8 were consistent "○", 1 was partially consistent "△", and 3 were removed with "×" (the parts with a probability of more than 75% were either inconsistent or partially consistent). Even after removing the partially consistent parts, the accuracy rate was still 84%.

[0119] Example 2

[0120] Fig.10 FIG. 2 is a block diagram of a defect detection device for a basket-type rotor according to Embodiment 2. In addition, the basic structure is substantially the same as that of Embodiment 1, so the same Figure 1 The same numbers and symbols are used for explanation.

[0121] In Example 2, based on the results of the three verification data that verified that the actual defect position was "directly below" in Example 1, two were removed, and a large amount of numerical analysis data on the rotor model having a defect directly below the detection unit 2 was obtained.

[0122] That is, the detection unit 2 fixedly arranged on the other end side in Example 1 for easy handling is sequentially arranged at a plurality of positions (m positions, where m≥2) set between the other end of the rotor model and the position where the variable magnetic field generating unit 1 is arranged, and a plurality of numerical analysis data about the configuration position of the detection unit 2 changed for one rotor model are obtained, and a large amount of learning data L11~L1m, L21~L2m...Ln1~Lnm (hereinafter abbreviated as "L11~Lnm") consisting of the configuration position of the detection unit 2 and the determination information k11~k1m, k21~k2m...kn1~knm (hereinafter abbreviated as "k11~knm") for determining the rotor model and the numerical analysis data A11~A1m, A21~A2m...An1~Anm (hereinafter abbreviated as "A11~Anm") obtained for each rotor model configured with the detection unit 2 are prepared.

[0123] Furthermore, after the learning data L11 to Lnm are accumulated in the neural network 5 and optimized through autonomous learning, Fig.10 As shown, the detection units 2 are sequentially arranged at a plurality of positions set between the other end of the basket-type rotor R and the position where the variable magnetic field generating unit 1 is arranged, a predetermined alternating current is applied to the first coil, and the current value data C1~Cm of the detection unit 2 detection timing are input to the input layer 5I of the neural network 5 together with the configuration position data P1~Pm of the detection unit 2.

[0124] Then, the neural network 5 performs inverse analysis, and outputs defect state estimation information indicating the presence, size, and position of a defect in the conductor bar of the basket-type rotor R from the output layer 50.

[0125] Furthermore, when performing inverse analysis, the detection unit 2 is sequentially arranged at a plurality of set locations to detect the current value data C1 to Cm of the timing. It is best to set the plurality of locations at equal intervals. When the intervals are small, defects directly below the detection unit 2 can be investigated with high precision.

[0126] Fig.11 It is a flow chart showing the algorithm of the defect detection device for the basket-type rotor involved in Example 2. A large amount of learning data L11~Lnm is generated through the following processes (T1')~(T4'), which are accumulated in the neural network 5 and autonomously learned. Then, the presence, scale and position of defects in the conductor bars of the basket-type rotor R are inferred through processes (1')~(6'), and the inferred presence, scale and position of defects are displayed.

[0127] (T1') A large number of rotor models (M1 to Mn) are produced based on the materials and shapes of the rotor core, conductor bars, end rings, etc. of the basket-type rotor R known in advance.

[0128] (T2') Create structural data K1 to Kn of each rotor model.

[0129] (T3') Using the FEM electromagnetic analysis unit, according to each structural data K1~Kn and the AC power supply data AC, the detection unit 2 is sequentially arranged at multiple set positions in each rotor model and a predetermined AC is applied to the first coil. Then, numerical analysis data A11~Anm representing the timing data of the current value induced in the second coil is output.

[0130] (T4′) Learning data L11 to Lnm consisting of identification information k11 to knm for identifying each rotor model in which the detection units 2 are sequentially arranged at a plurality of set locations and the numerical analysis data A11 to Anm output at T3′ are generated and stored in the neural network 5.

[0131] (1′) The detection units 2 are sequentially arranged at a plurality of locations set between the other end of the basket-type rotor R and the location where the variable magnetic field generating unit 1 is arranged, and a predetermined alternating current is applied to the variable magnetic field generating unit 1 arranged at one end of the basket-type rotor R.

[0132] (2′) When a predetermined alternating current is applied to the first coil of the varying magnetic field generating means 1 , time-series current value data C1 to Cm induced in the second coils of the sequentially arranged detecting means 2 are detected.

[0133] (3′) The detected time-series current value data C1 to Cm and the arrangement position data P1 to Pm of the sequentially arranged detection units 2 are sent to the neural network 5 .

[0134] (4′) The neural network 5 performs inverse analysis based on the input time-series current value data C1 to Cm and arrangement position data P1 to Pm.

[0135] (5′) Defect state estimation information indicating the presence, size, and position of defects in the conductor bars of the basket-type rotor R is output by inverse analysis and sent to the defect information display unit 6 .

[0136] (6′) The defect information display unit 6 displays the presence, size, and position of a defect in the conductor bar of the basket-type rotor R based on the defect state estimation information.

[0137] If there are conductor bars that have not been estimated and displayed, the variable magnetic field generating unit 1 and the detecting unit 2 are moved across the conductor bars to execute the above-mentioned processes (1') to (6'), and the estimated defect presence, size and position are displayed for all conductor bars and then the process ends.

[0138] Modifications of the embodiment will be listed.

[0139] (1) The detection unit 2 according to the first and second embodiments detects the current value induced in the second coil, but may also detect a voltage value.

[0140] (2) In the rotor model data output unit 3 of Examples 1 and 2, a semi-cylindrical or cylindrical defect of various sizes is randomly arranged in a conductor rod, but the shape of the defect is not limited to a semi-cylindrical or cylindrical shape, and can also be set to a shape close to the pore actually formed.

[0141] Alternatively, one or more defects may be randomly arranged.

[0142] (3) The neural network 5 of Examples 1 and 2 outputs defect state inference information indicating the presence, size and position of defects in the conductor bars of the basket-type rotor R, but may also output defect inference information indicating only the presence or absence of defects, defect scale inference information indicating the presence or absence of defects and their size, or defect presence or absence inference information indicating the presence or absence of defects and defect position inference information indicating the position of the defects.

[0143] (4) The defect information display unit 6 involved in Examples 1 and 2 displays the presence, size and position of defects in the conductor bars of the basket-type rotor R, but it can also use sound reporting instead of or in addition to the display, and can also send information related to the presence, size and position of defects to other processing devices.

[0144] Therefore, the expression "defect information reporting unit" is used in the claims.

[0145] (5) In the defect detection device for the basket-type rotor involved in Examples 1 and 2, it is determined after process (6) whether there are conductor bars that have not been estimated and displayed, but the determination may also be made after process (5). After the estimation of all the conductor bars is completed, as process (6), the defect information display unit 6 is used to collectively display the presence, size and position of defects in all the conductor bars of the basket-type rotor R.

[0146] (6) In the defect detection device for the basket-type rotor involved in Examples 1 and 2, a FEM electromagnetic analysis unit is used to generate learning data and verification data, but it is not limited to the FEM electromagnetic analysis unit, and a numerical analysis unit based on the differential method, boundary element method, etc. can also be used.

[0147] (Explanation of symbols)

[0148] 1: variable magnetic field generating unit; 1c: 1st coil; 1i: 1st core; 1p: alternating current; 2: detection unit; 3: rotor model data output unit; 4: rotor model analysis unit; 5: neural network; 5I: input layer; 5L: learning data accumulation unit; 5O: output layer; 5T: learning data input unit; 6: defect information display unit; A1~An: numerical analysis data; AC: alternating current power supply data; C: time series current value data; C1~Cm: time series current value data; I(t): current; K1~Kn: structural data; k1~kn: determination information; k11~knm: determination information; L1~Ln: learning data; L11~Lnm: learning data; M1~Mn: rotor model; R: basket rotor; Rc: conductor rod; Changing magnetic field.

Claims

1. A defect detection device for a basket-type rotor, used for inferring and reporting defects existing in a conductor bar of a basket-type rotor, characterized in that: The defect detection device of the basket-type rotor comprises: a variable magnetic field generating unit, comprising a first U-shaped iron core arranged to straddle one conductor bar of the basket-type rotor, a first coil wound around the first iron core, and an AC power source connected to the first coil; a detection unit including a U-shaped second core that can be arranged to straddle one conductor bar of the basket-type rotor, a second coil wound around the second core, and a meter for detecting a current value or a voltage value induced in the second coil; a rotor model data output unit for producing a large number of rotor models having different defects existing in one conductor bar, with respect to the basket-type rotor, and outputting structural data related to the large number of rotor models; a rotor model analysis unit including a numerical analysis unit, the numerical analysis unit receiving the structural data and the AC power supply data outputted from the rotor model data output unit, and numerically analyzing the current value or the voltage value induced in the second coil when the first core and the second core are arranged across one conductor bar of each rotor model and a predetermined AC is applied to the first coil, the rotor model analysis unit outputting, as learning data, identification information for identifying a rotor model to be subjected to numerical analysis and numerical analysis data obtained by the numerical analysis performed on the rotor model; A neural network receives the time-series current value data or voltage value data detected by the detection unit when a predetermined alternating current is applied to the first coil, performs inverse analysis, and outputs defect estimation information; as well as a defect information reporting unit that reports the presence or absence of defects in the conductor bars of the basket-type rotor based on the defect estimation information output from the neural network, The neural network is provided in advance with learning data on the large number of rotor models output from the rotor model analyzing unit, and is optimized through autonomous learning.

2. The defect detection device for a basket-type rotor according to claim 1, characterized in that: The neural network outputs defect scale estimation information instead of defect presence estimation information. The defect information reporting unit reports the scale of a defect existing in the conductor bar of the basket-type rotor based on the defect scale estimation information.

3. The defect detection device for a basket-type rotor according to claim 1, characterized in that: The variable magnetic field generating unit is arranged on one end side of the basket-type rotor. The detection units are sequentially arranged at a plurality of locations set between the other end of the basket-type rotor and the location where the variable magnetic field generating unit is arranged. In the numerical analysis unit, when the second core is sequentially arranged at a plurality of locations set between the other end of each rotor model and the location where the first core is arranged and a predetermined alternating current is applied to the first coil, a current value or a voltage value induced in the second coil is numerically analyzed, The identification information is information that identifies a rotor model that is a target of numerical analysis and in which the second cores are sequentially arranged at the plurality of locations. The neural network receives the current value data or voltage value data of the time series and the configuration position data of the detection unit and performs inverse analysis to output defect presence estimation information and defect position estimation information. The defect information reporting unit reports the presence or absence of a defect in the conductor bar of the basket-type rotor and the position of the defect based on the defect estimation information and the defect position estimation information output from the neural network.

Images