CN115564769A

CN115564769A - Method for detecting motor rotor doubling defect by using deep learning

Info

Publication number: CN115564769A
Application number: CN202211407596.8A
Authority: CN
Inventors: 宋昭颖; 邓永松; 王君; 孟庆猛
Original assignee: Suzhou Shixin Integration Technology Co ltd
Current assignee: Suzhou Shixin Integration Technology Co ltd
Priority date: 2022-11-10
Filing date: 2022-11-10
Publication date: 2023-01-03

Abstract

The application relates to a method for detecting the doubling defect of a motor rotor by using deep learning, which belongs to the technical field of deep learning and comprises the following steps: the method comprises the steps of acquiring an image acquisition assembly to acquire images of each interval, acquiring at least two target images corresponding to each interval, inputting each target image into an identification model obtained based on deep learning, obtaining a target area where at least one target interval is located in the target image, a target frame where each target interval belongs to and a defect classification result of the target frame, and determining that a motor rotor has a doubling defect under the condition that the defect classification result corresponding to at least one target image indicates that the distance between two enameled wires is smaller than a preset distance in the at least two target images. The problem that the manual work is inaccurate to the distance hold between two strands of enameled wires because the commutator is less, and manual detection efficiency is lower, and the detection effect to rotor doubling defect that consequently leads to is not good can be solved.

Description

Method for detecting motor rotor doubling defect by using deep learning

[ technical field ] A method for producing a semiconductor device

The application relates to a method for detecting the doubling defect of a motor rotor by using deep learning, belonging to the technical field of deep learning.

[ background of the invention ]

The rotor of the motor is an important component of the motor. One end of the motor rotor is provided with a commutator, convex hooks are uniformly distributed on the peripheral side of the commutator, an interval is formed between every two adjacent hooks, an enameled wire of the motor rotor is provided with an insulating coating, and the enameled wire is wound in from the interval on one side of each hook and then is wound out from the interval on the other side of the hook. At this time, two enameled wires are arranged in each groove. The motor during operation, the rotor drives the high-speed rotation of commutator, and if two strands of enameled wires in the slot are too near this moment, the looks mutual friction that can not stop between the enameled wire leads to insulating coating crackle or explode and split even, thereby the enameled wire will leak electricity and lead to the motor short circuit when the insulating layer explodes and splits. Therefore, a doubling defect detection of the motor rotor is required.

In a traditional rotor doubling defect detection method, two strands of enameled wires in a commutator interval are generally identified manually, and whether the defect that the distance between the enameled wires is too short exists is determined.

However, the commutator is small, so that the distance between the enameled wires is not accurately grasped manually, and the manual detection efficiency is low, so that the detection effect on the doubling defect of the rotor is poor.

[ summary of the invention ]

The application provides a method for detecting the doubling defect of a motor rotor by using deep learning, which can solve the problem that the commutator is small, so that the distance between enameled wires is not accurately grasped manually, and the manual detection efficiency is low, thereby causing the problem of poor detection effect on the doubling defect of the rotor. The application provides the following technical scheme:

in a first aspect, a method for detecting a motor rotor doubling defect by using deep learning is provided, wherein a motor rotor to be detected is placed at a detection position of a defect detection table, and an image acquisition assembly positioned on the defect detection table is suitable for acquiring an image of the detection position; one end of the motor rotor is provided with a commutator, at least two raised hooks are uniformly arranged on the periphery of the commutator in a surrounding manner, a space is formed between every two adjacent hooks, and an enameled wire of the motor rotor is wound in from the space on one side of each hook and then is wound out from the space on the other side of the hook; the method comprises the following steps:

acquiring the image acquisition component to acquire images of each interval to obtain at least two target images corresponding to each interval; wherein each target image comprises at least one interval and image data of two enameled wires in the interval;

inputting each target image into an identification model obtained based on deep learning to obtain a target area where at least one target interval is located, a target frame where each target interval belongs and a defect classification result of the target frame; the target interval is an interval which meets the identification standard of the model in the at least one interval, and the position of the target frame is determined based on the positions of two strands of enameled wires in the target interval;

and determining that the motor rotor has a doubling defect under the condition that the defect classification result corresponding to at least one target image in the at least two target images indicates that the distance between the two enameled wires is smaller than the preset distance.

Optionally, the training process of the recognition model includes:

acquiring the image acquisition component to acquire images of each interval to obtain at least two training images corresponding to each interval;

adding a target area label, a target frame label and a defect classification label corresponding to each target frame to each training image by using image processing software to obtain a defect classification training set;

and training a pre-established neural network model by using a defect classification training set to obtain the recognition model, wherein the neural network model is established based on YOLOv 5.

Optionally, the training a pre-created neural network model by using a defect classification training set to obtain the recognition model includes:

in the training process, monitoring the training process of the recognition model by using a deep learning visualization tool to obtain a training value corresponding to the training process, wherein the training value comprises a recall rate, a precision rate and an average precision average value;

and under the condition that the training numerical values are all larger than or equal to the corresponding preset standard numerical values, outputting the identification model obtained by the training.

Optionally, before the step of inputting each target image into the recognition model obtained based on the deep learning to obtain the target region where at least one target interval is located in the target image, the target frame to which each target interval belongs, and the defect classification result of the target frame, the method further includes:

converting the model format into a processing format of an image processing library under the condition that the model format of the identification model does not match the processing format; the image processing library includes: open source computer vision library OpenCV, or HALCON.

Optionally, determining that the motor rotor has a doubling defect when the defect classification result corresponding to at least one of the at least two target images indicates that the distance between the two enameled wires is smaller than a preset distance includes:

acquiring a first confidence coefficient of each target frame in a target area output by the recognition model;

determining whether a first confidence coefficient of the target frame is smaller than a preset confidence coefficient;

deleting the target frame with the first confidence coefficient smaller than the preset confidence coefficient and the defect classification result corresponding to the target frame;

and determining that the motor rotor has a doubling defect under the condition that the defect classification result corresponding to at least one screened target frame indicates that the distance between the two enamelled wires is smaller than the preset distance.

Optionally, determining that the motor rotor has a doubling defect when a defect classification result corresponding to at least one target image in the at least two target images indicates that the distance between the two enameled wires is smaller than a preset distance includes:

determining a reference target frame with the highest first confidence coefficient from the target area;

calculating the intersection ratio between other target frames in the target area except the reference target frame and the reference target frame;

deleting other target frames with the intersection ratio larger than or equal to the intersection ratio threshold;

and determining that the motor rotor has a doubling defect under the condition that the defect classification result corresponding to at least one deleted target frame indicates that the distance between the two enamelled wires is smaller than the first preset distance.

Optionally, an illumination assembly is further arranged on the defect detection table, and the illumination assembly is suitable for providing auxiliary illumination for the interval;

before the acquiring of the image acquisition assembly for image acquisition of each interval and obtaining of at least two target images corresponding to each interval, the method further comprises:

controlling the lighting assembly to start;

and acquiring an image of each illuminated interval by using the image acquisition assembly.

Optionally, a distance measuring assembly is further arranged on the defect detecting table, and the distance measuring assembly is suitable for measuring the distance between the distance measuring assembly and the commutator;

acquiring a distance value acquired by a distance measurement component;

determining whether the image capture component is opposite a gap using the distance value; a first distance value acquired by a distance measurement component under the condition that the image acquisition component is opposite to the interval is larger than a second distance value acquired by the distance measurement component under the condition that the image acquisition component is opposite to the hook;

and under the condition that the image acquisition assembly is opposite to the interval, performing image acquisition on the interval by using the image acquisition assembly.

Optionally, the defect detecting table includes a rotary driving member, the rotary driving member is adapted to drive the detecting position to rotate, and a rotation axis of the detecting position is an axis of the motor rotor when the motor rotor is placed on the detecting position;

before obtaining the distance value that the range finding subassembly was gathered, still include: controlling the rotary driving piece to drive the detection position to rotate continuously, and controlling the distance measurement assembly to acquire the distance value every preset time so as to trigger and execute the acquisition of the distance value acquired by the distance measurement assembly; determining whether the image capture component is opposite a gap using the distance value; a step of performing image acquisition on the interval by using the image acquisition assembly under the condition that the image acquisition assembly is opposite to the interval;

alternatively, the first and second electrodes may be,

the method further comprises the following steps: under the condition that the image acquisition assembly is opposite to the hook, controlling the rotary driving piece to drive the detection position to rotate by a preset angle; controlling the distance measurement assembly to acquire the distance value so as to trigger and execute the distance value acquired by the distance measurement assembly again; determining whether the image capture component is opposite a gap using the distance value; a step of performing image acquisition on the interval by using the image acquisition assembly under the condition that the image acquisition assembly is opposite to the interval; the preset angle is smaller than an included angle formed by the two adjacent hooks and the image acquisition assembly.

Optionally, after determining that the motor rotor has a doubling defect when the defect classification result corresponding to at least one target image in the at least two target images indicates that the distance between two enameled wires is smaller than a preset distance threshold, the method further includes:

sending a gripping instruction to a gripping device to cause the gripping device to move the motor rotor from the detection position to a target position in response to the gripping instruction, the target position being used for storing the motor rotor with the doubling defect

The beneficial effect of this application includes at least: acquiring an image of each interval by an image acquisition assembly to obtain at least two target images corresponding to each interval, wherein each target image comprises at least one interval and image data of two enameled wires in the interval, inputting each target image into an identification model obtained based on deep learning to obtain a target area where the at least one target interval is located in the target image, a target frame where the each target interval belongs to and a defect classification result of the target frame, the target interval is an interval which meets the identification standard of the model in the at least one interval, the position of the target frame is determined based on the positions of the two enameled wires in the target interval, and the existence of the doubling defect of the motor rotor is determined under the condition that the defect classification results corresponding to the at least one target image indicate that the distance of the enameled wires is smaller than a preset distance in the at least two target images; the problem that the distance between two strands of enameled wires is not accurately grasped manually due to the fact that a commutator is small, and manual detection efficiency is low, and therefore the detection effect on the rotor doubling defect is poor can be solved; the efficiency of identifying the target image by using the identification model obtained through deep learning is higher, and the model can accurately grasp whether the distance between two strands of enameled wires in the target image is smaller than the preset distance, so that the detection effect on the doubling defect of the motor rotor can be improved.

In addition, because only the target area has the target frame, and the target frame only exists in the target area, the target area label, the target frame label and the defect classification label corresponding to each target frame are used for training the neural network model, so that the labels can be mutually verified, and the recognition accuracy obtained after training is improved.

In addition, under the condition that the training values are all larger than or equal to the corresponding preset standard values, the recognition model obtained by the training is output, the performance of the recognition model obtained after the training can be ensured, and therefore the recognition effect of the recognition model is improved.

In addition, the model format is converted into the processing format of the image processing library, so that the recognition model can be loaded by using the image processing library; and the converted recognition model is loaded by using the image processing library, and the target image can be subjected to image processing by using the auxiliary recognition model of the image processing library, so that the processing speed of the recognition model is increased.

In addition, the lower the first confidence coefficient is, the higher the false detection rate is, so that the false detection rate can be reduced by deleting the target frame with the first confidence coefficient smaller than the preset confidence coefficient and the defect classification result corresponding to the target frame, thereby improving the accuracy of the identification model.

In addition, a reference target frame with the highest first confidence coefficient is determined from the target area; calculating the intersection ratio between other target frames except the reference target frame and the reference target frame in the target area; and deleting other target frames with the intersection ratio larger than or equal to the intersection ratio threshold value, so that each target interval is ensured to correspond to one target frame, the target frame which is most matched with the target interval is obtained, and the recognition effect of the recognition model is improved.

In addition, the image acquisition assembly is used for acquiring images of each illuminated interval, so that the influence of ambient light on the identification model can be avoided, and the applicability of the identification model is improved.

In addition, whether the image acquisition assembly is opposite to the interval or not is determined by using the distance value, and the image acquisition assembly is used for acquiring images of the interval under the condition that the image acquisition assembly is opposite to the interval, so that the condition that the interval exists at the edge of the target image and the identification model cannot identify the interval can be avoided, and the identification effect of the identification model can be improved.

The foregoing description is only an overview of the technical solutions of the present application, and in order to make the technical solutions of the present application more clear and clear, and to implement the technical solutions according to the content of the description, the following detailed description is made with reference to the preferred embodiments of the present application and the accompanying drawings.

[ description of the drawings ]

FIG. 1 is a schematic diagram of a system for motor rotor doubling defect detection using deep learning according to an embodiment of the present application;

FIG. 2 is a schematic view of a rotor of an electric machine provided in accordance with an embodiment of the present application;

FIG. 3 is a schematic view of two strands of enameled wire in a space provided by one embodiment of the present application;

FIG. 4 is a schematic illustration of a spacing provided by one embodiment of the present application;

FIG. 5 is a flowchart of a method for detecting a motor rotor doubling defect using deep learning according to an embodiment of the present application;

FIG. 6 is a diagram illustrating the addition of a target box label and a defect classification label corresponding to the target box by using labelme software according to an embodiment of the present application;

FIG. 7 is a graph of training values output by the model training process provided by one embodiment of the present application;

FIG. 8 is a diagram illustrating the result of processing and outputting a target image after an OpenCV recognition model is loaded according to an embodiment of the present application;

FIG. 9 is a schematic illustration of a result of a test of a first station provided by one embodiment of the present application;

FIG. 10 is a schematic illustration of a result of a second workstation inspection provided in accordance with an embodiment of the present application.

[ detailed description ] embodiments

The following detailed description of embodiments of the present application will be made with reference to the accompanying drawings and examples. The following examples are intended to illustrate the present application, but are not intended to limit the scope of the present application.

FIG. 1 is a schematic diagram of a system for detecting a motor rotor doubling defect using deep learning according to an embodiment of the present application. The motor is an electromagnetic device for realizing electric energy conversion or transmission according to an electromagnetic induction law. The motor is composed of a motor rotor 2 and a motor stator.

As shown in fig. 1, the system for detecting the doubling defect of the motor rotor 2 by using deep learning at least comprises: a defect inspection station 1 and an image acquisition assembly 12.

The defect detection station 1 is adapted to carry a motor rotor 2.

The motor rotor 2 is a rotating part in a motor, one end of the motor rotor 2 is generally provided with a commutator, at least two raised hooks are uniformly and annularly arranged on the peripheral side of the commutator, and a space is formed between every two adjacent hooks.

The spaces on the commutator are also referred to as hook grooves. The interval between two adjacent hooks may be a concave groove, or may be a part of the circumferential surface of the commutator sandwiched between two hooks, and the embodiment does not limit the implementation manner of the interval.

As shown in fig. 2, one side of the motor rotor 2 is provided with a commutator 20, and one side of the commutator 20 close to the rotor body 21 is uniformly provided with a convex hook 201 in a ring shape.

The enameled wire of the motor rotor 2 winds in from the interval of one side of each hook and then winds out from the interval of the other side of the hook. At this time, two enameled wires are present in each space.

The enameled wire is a winding wire which is wound on the rotor main body of the motor rotor 2 and used for being communicated with current, and the surface of the enameled wire is provided with an insulating coating so as to avoid short circuit of the motor caused by mutual conduction between the electrified enameled wires.

As shown in fig. 3, the enamel wire 31 is wound in from the space on the left side of the hook 32 and then wound out from the space on the right side of the hook 32, and two enamel wires 31 exist in each space.

The defect detection table 1 is also provided with an image acquisition assembly 12. The image capturing assembly 12 is adapted to capture images of the intervals to obtain at least two target images corresponding to each interval. Wherein each target image includes image data of at least one space and two strands of enameled wires within the space.

The side of the commutator having the hooks is close to the rotor body, and as shown in fig. 4, the plane in which the space 41 is located is inclined toward the central axis 42 of the commutator 20, so that the space 41 is not visible from the other side of the commutator 20 in a direction parallel to the central axis 42. Therefore, if it is desired to acquire image data of two enamel wires in all intervals and intervals on the commutator, at least two target images need to be acquired using the image acquisition assembly 12.

The plane of the gap is inclined to the central shaft, and the rotor body can shield the gap. Therefore, the position of the image acquisition assembly 12 needs to be adjusted.

Specifically, the image capturing assembly 12 is installed at a position corresponding to the right side of the central axis on the defect detecting table 1, so that the image capturing assembly 12 is opposite to the plane where the gap is located, and the gap is not blocked by the rotor body.

In the present embodiment, the image capturing component 12 further has a recognition function therein. The image acquisition component 12 identifies the enameled wire in the target image and obtains a defect classification result corresponding to the target image.

And the defect classification result is used for indicating whether the distance between the two enamelled wires in the same interval is smaller than a preset distance.

The gripping device is a device for moving the motor rotor 2 to different positions in a system for performing the doubling defect detection of the motor rotor 2 using deep learning.

The clamping device is in communication connection with the image acquisition assembly 12 to move the motor rotor 2 to different positions according to the defect classification result transmitted by the image acquisition assembly 12.

Because the defect detection table 1 has a large bearing area, and the too large acquisition range of the image acquisition assembly 12 causes the target image acquired by the image acquisition assembly 12 to have a small resolution, the image acquisition assembly 12 cannot accurately identify the space and the enameled wire in the space.

Optionally, the defect inspection table 1 further has a preset inspection position 11 thereon. The motor rotor 2 to be detected is placed on the detection position 11 of the defect detection table 1, and the image acquisition assembly 12 is used for carrying out image acquisition on the detection position 11 so as to obtain a target image.

Optionally, the defect inspection station 1 further comprises a rotary driving component, the rotary driving component is adapted to drive the inspection position 11 to rotate, and the rotation axis of the inspection position 11 is the axial center of the motor rotor 2 when placed on the inspection position 11.

The axis of the motor rotor 2 is the same as the axis of the central shaft of the commutator, that is, the rotation axis of the detection position 11 is the same as the central shaft of the commutator.

In this embodiment, the rotary drive assembly has two modes of operation:

first, the rotary drive assembly continues to rotate.

Secondly, rotating the rotary driving assembly at preset acquisition angles at preset time intervals; the preset time interval and the preset collection angle may be pre-stored in the device or input by a user, and the implementation manner of the preset time interval and the preset rotation angle is not limited in this embodiment.

Optionally, an illumination assembly 13 is further disposed on the defect inspection table 1. The illumination assembly 13 is adapted to provide supplemental illumination for the gap to reduce the effect of ambient light on the target image captured by the image capture assembly 12.

The rotation driving assembly drives the detection position 11 to rotate, when the image acquisition assembly performs image acquisition on the interval, the condition that the hook is opposite to the image acquisition assembly 12 can occur, and in the acquired target image, the image data of the enameled wire in the interval and the interval can not accord with the identification standard, so that the enameled wire cannot be identified.

The condition that the image data of the two enamelled wires in the interval have noise which influences identification, image distortion or the interval area is not completely acquired is indicated when the identification standard is not met.

Optionally, a distance measuring assembly 120 is further disposed on the defect detecting table 1. The ranging assembly 120 is adapted to measure the distance between the ranging assembly 120 and the commutator.

Ranging assembly 120 includes, but is not limited to: fiber optic sensors, radar sensors, etc.

The distance measuring assembly 120 is located on the same plane as the image capturing assembly 12, so that the distance between the distance measuring assembly 120 and the commutator can be similar to the distance between the image capturing assembly 12 and the commutator.

Since there is a height difference between the space on the commutator and the hook, and the distance between the hook and the ranging assembly 120 is smaller than the distance between the space and the ranging assembly 120, it can be determined whether the space is opposite to the image capturing assembly 12 by using the ranging assembly 120 to capture the distance value between the ranging assembly 120 and the commutator.

When the motor started working, electric motor rotor 2 high-speed rotatory, if the distance between two strands of enameled wires in the interval is less than preset distance, promptly, electric motor rotor 2 has the doubling defect, then two strands of enameled wires can rub each other always during electric motor rotor 2 high-speed rotatory, lead to insulating layer on the enameled wire to produce crackle or even explode and split for thereby the enameled wire electric leakage leads to the motor short circuit.

In actual implementation, the preset distance can be set according to different requirements, so that the preset distances used by different users are the same or different.

In a traditional rotor doubling defect detection method, two strands of enameled wires in a commutator interval are generally identified manually, and whether a defect that the enameled wires are too close to each other exists is determined.

In view of the above problem, in the present embodiment, the image capturing assembly 12 is configured to: acquiring at least two target images corresponding to each interval by acquiring the images of each interval by the image acquisition component 12; inputting each target image into an identification model obtained based on deep learning to obtain a target area where at least one target interval is located, a target frame where each target interval belongs and a defect classification result of the target frame; and determining that the doubling defect exists in the motor rotor 2 under the condition that the defect classification result corresponding to at least one target image in the at least two target images indicates that the distance between the two enameled wires is smaller than the preset distance.

The target interval is an interval which meets the recognition standard of the model in at least one interval, and the position of the target frame is determined based on the positions of two enameled wires in the target interval.

In the embodiment, image acquisition is performed on each interval through an image acquisition component, at least two target images corresponding to each interval are obtained, wherein each target image comprises at least one interval and image data of two enameled wires in the interval, each target image is input into an identification model obtained based on deep learning, a target area where the at least one target interval is located in the target image, a target frame where the each target interval belongs to and a defect classification result of the target frame are obtained, the target interval is an interval which meets the identification standard of the model in the at least one interval, the position of the target frame is determined based on the positions of the two enameled wires in the target interval, and when the defect classification result corresponding to the at least one target image indicates that the distance between the two enameled wires is smaller than a preset distance, it is determined that a doubling defect exists in a motor rotor; the problem that the distance between two strands of enameled wires cannot be accurately grasped manually due to the fact that a commutator is small and manual detection efficiency is low, and therefore the detection effect on the rotor doubling defect is poor can be solved; the efficiency of identifying the target image by using the identification model obtained through deep learning is higher, and the model can accurately grasp whether the distance between two strands of enameled wires in the target image is smaller than the preset distance, so that the detection effect on the doubling defect of the motor rotor can be improved.

The method for detecting the doubling defect of the motor rotor by using deep learning provided by the present application is described in detail below. The following embodiments are described by taking as an example that the method is used in an electronic device communicatively connected to the image capturing assembly shown in fig. 1, and specifically used in a processor in the electronic device, or other devices communicatively connected to the image capturing assembly, such as: for user terminals, or servers, etc., wherein the user terminals include but are not limited to: a cell phone, a tablet computer, a wearable device, or an industrial control computer, etc.

The communication connection mode may be wired communication or wireless communication, and the wireless communication mode may be short-range communication or wireless communication, and the present embodiment does not limit the communication mode between the mobile device and other devices.

Specifically, the communication connection mode includes a Modbus protocol, a Controller Area Network bus (CAN) protocol, or the like.

In practical implementation, the method may also be applied to an image capturing component, and the implementation manner of other devices and the implementation manner of the user terminal are not limited in this embodiment.

Fig. 5 is a flowchart of a method for detecting a motor rotor doubling defect using deep learning according to an embodiment of the present application. The method at least comprises the following steps:

step 501, acquiring an image acquisition assembly to acquire images of each interval, and acquiring at least two target images corresponding to each interval.

Illustratively, the number of target images is related to the number of target intervals per acquisition on the target image and the intervals on the motor rotor.

For example, the motor rotor has 26 intervals, and the interval of the targets collected in each target image is 2, and in this case, 13 target images exist.

Wherein the target interval is an interval that meets the recognition criteria of the model in at least one interval.

Because image acquisition subassembly is when carrying out image acquisition to electric motor rotor, ambient light can lead to the target image to have uncontrollable noise, influences the recognition effect of identification model.

Optionally, the acquiring the image acquisition assembly performs image acquisition on each interval, and before obtaining at least two target images corresponding to each interval, the method further includes: controlling the lighting assembly to start; and using the image acquisition assembly to acquire images of each interval after illumination.

The illumination assembly is used for assisting illumination, so that the ambient light in the target image is ensured to be consistent, the interference of the ambient light to the recognition model is eliminated, and the recognition effect of the recognition model is improved.

Because the acquisition range of the image acquisition assembly is limited, and the image acquisition assembly acquires the peripheral side of the motor rotor, when the hook is opposite to the image acquisition assembly, the interval can be at the edge of a target image, so that the image data in the interval and the interval have the conditions of noise, distortion, incomplete acquisition and the like; and because the image acquisition assembly continues to acquire the next target image after a certain preset acquisition duration, the subsequently acquired target images may be dislocated with the interval under the condition that the first target image is dislocated with the interval, so that the motor rotor corresponding to the target image cannot be subjected to doubling defect detection. Therefore, it is necessary to determine whether the image capturing assembly is opposite the gap before the image capturing assembly captures the target image.

Optionally, the acquiring the image acquisition assembly performs image acquisition on each interval, and before obtaining at least two target images corresponding to each interval, the method further includes: acquiring a distance value acquired by a distance measurement component; determining whether the image capture assembly is opposite the gap using the distance value; the first distance value acquired by the distance measurement component under the condition that the image acquisition component is opposite to the interval is larger than the second distance value acquired by the distance measurement component under the condition that the image acquisition component is opposite to the hook; and under the condition that the image acquisition assembly is opposite to the interval, the image acquisition assembly is used for carrying out image acquisition on the interval.

Because the distance measuring assembly and the image acquisition assembly are positioned on the same plane, the distance between the distance measuring assembly and the commutator can be similar to the distance between the image acquisition assembly and the commutator.

Because the height difference exists between the interval and the hook on the commutator, and the second distance value between the hook and the distance measuring component is smaller than the first distance value between the interval and the distance measuring component, the distance measuring component is used for collecting the distance value between the distance measuring component and the commutator, and whether the image collecting component is opposite to the interval or not can be determined.

Because the rotational drive assemblies operate in different ways, the use of the range finding assembly to determine the image capture assembly relative to the gap to capture the target image also includes, but is not limited to, the following two ways.

The first method, before obtaining the distance value collected by the ranging component, further comprises: controlling a rotary driving piece to drive the detection position to rotate continuously, and controlling a distance measurement component to acquire distance values every preset time so as to trigger execution of acquiring the distance values acquired by the distance measurement component; determining whether the image capture assembly is opposite the gap using the distance value; and under the condition that the image acquisition assembly is opposite to the interval, carrying out image acquisition on the interval by using the image acquisition assembly.

Secondly, before obtaining the distance value collected by the ranging component, the method further comprises: under the condition that the image acquisition assembly is opposite to the hook, controlling the rotary driving piece to drive the detection position to rotate by a preset angle; controlling the distance measurement assembly to acquire the distance value so as to trigger and execute the distance value acquired by the distance measurement assembly again; determining whether the image capture assembly is opposite the gap using the distance value; and under the condition that the image acquisition assembly is opposite to the interval, carrying out image acquisition on the interval by using the image acquisition assembly.

The preset angle is smaller than an included angle formed by the two adjacent hooks and the image acquisition assembly, so that the condition that other hooks are opposite to the image acquisition assembly after the motor rotor rotates can be avoided.

Under the condition that the image acquisition assembly is determined to be opposite to the interval, the target image is acquired, the interval can be determined to be within the acquisition range of the image acquisition assembly, the target interval is prevented from appearing at the edge of the target image, and therefore the identification effect of the identification model can be improved.

Step 502, inputting each target image into an identification model obtained based on deep learning, and obtaining a target area where at least one target interval is located in the target image, a target frame where each target interval belongs, and a defect classification result of the target frame.

The target area is an area containing all target intervals in the target image, and the positions of the target frame are determined based on the positions of two strands of enameled wires in the target intervals.

Before the recognition model is used, deep learning training needs to be performed on a pre-created neural network model to obtain the recognition model.

Optionally, the training process of the recognition model includes: acquiring an image acquisition component to acquire images of each interval to obtain at least two training images corresponding to each interval; adding a target area label, a target frame label and a defect classification label corresponding to each target frame to each training image by using image processing software to obtain a defect classification training set; and training a pre-established neural network model by using a defect classification training set to obtain an identification model, wherein the neural network model is established based on YOLOv 5.

Image processing software includes, but is not limited to: labelme, labellimg, and the like.

For example, as shown in fig. 6, a target frame label 61 and a defect classification label 61 corresponding to the target frame 60 are added to the training image using labelme.

At this time, a defect classification training set obtained by labeling the training image by using labelme is in a json format, and a neural network model created based on YOLOv5 can only process data in a txt format; therefore, a Python script is also needed to convert the format of the defect classification training set into a defect classification training set in a txt format, so as to train the neural network model created based on YOLOv5 through the converted defect classification training set.

Wherein, the defect classification label 61 ng' means that the distance between two strands of enameled wires in the target frame 60 is smaller than a preset distance; and the defect classification label 63 ok means that the distance between the two strands of enamel wire within the target frame 62 is greater than a preset distance.

Because the target frame exists only in the target area and only exists in the target area, the target area label, the target frame label and the defect classification label corresponding to the target frame are used for training the model at the same time, so that the labels can be verified mutually, and the accuracy of the recognition model obtained after training is improved.

The pre-created Neural Network model may be a Region-based Convolutional Neural Network (R-CNN), spatial Pyramid Pooling (SPPNet), or real-time fast object detection (Youynloy look: unefielded, real-time object detection, yoLO) series.

Wherein the YOLO series include: YOLOv3, YOLOv4, and YOLOv5, and the like.

Compared with other neural network models, YOLOv5 can perform mosaic enhancement on the training image, so that the detection effect of the small target object is improved.

The mosaic enhancement means that four images are spliced in a random scaling, random cutting and random distribution mode. Random scaling can improve the detection effect of small target objects.

And the interval of the motor rotor and two enameled wires in the interval belong to small target objects, so that the detection effect on the target interval can be improved by using YOLOv 5.

When the neural network model is trained, the user only knows the input defect classification training set and the output defect classification training result, but cannot know the training process. When the output result of the model does not meet the expectation of the user, the user needs to spend a lot of time to find the reason for the deviation of the model from the expectation, thereby affecting the efficiency of model training.

Optionally, training a neural network model created in advance by using a defect classification training set to obtain a recognition model, including: in the training process, monitoring the training process of the recognition model by using a deep learning visualization tool to obtain training values corresponding to the training process, wherein the training values comprise recall rate, precision rate and average precision mean value; and under the condition that the training numerical values are all larger than or equal to the corresponding preset standard numerical values, outputting the identification model obtained by the training.

Wherein, the deep learning visualization tool includes but is not limited to: weight & Biases (Weights & Biases, wandb) visualization tools, or TensorBoard tools, etc.

For example, wandb can record the change condition of each index and the setting of the super parameter in the model training process, and can also visually compare the output result, so that a user is helped to analyze the problems existing in the model training process, and the model training efficiency is improved.

Illustratively, monitoring a training process of the recognition model by using a deep learning visualization tool to obtain a training value corresponding to the training process, including: and obtaining a confusion matrix of the model by using a deep learning visualization tool, and obtaining a training numerical value corresponding to the training process through the confusion matrix.

The confusion matrix is an analysis table for summarizing model prediction results in data analysis and machine learning, and records in a training data set are summarized in a matrix form according to two standards of real categories and classification judgment made by a classification model.

Taking the confusion matrix corresponding to the binary model as an example, as shown in table 1:

table 1:

wherein TP refers to the number of samples predicted as positive samples in the actually positive samples, i.e. the number of correctly predicted samples in the positive samples; FN refers to the number of samples predicted as negative samples among actually positive samples, i.e., the number of samples in which a prediction is erroneous among positive samples; TN means the number of samples predicted as negative samples in the actually negative samples, i.e. the number of correctly predicted samples in the negative samples; FP refers to the number of samples that are predicted to be positive samples among the actually negative samples, i.e., the number of samples that are mispredicted among the negative samples.

Recall (Recall), also called Recall, refers to the probability of being predicted as a positive sample among the actual positive samples. The formula is as follows:

precision ratio (P), also called Precision ratio, refers to the probability of all samples being predicted as positive, i.e. the prediction accuracy probability in the result of predicting as positive, and the formula is as follows:

in practical application, the recall rate and the precision rate are in one-to-one correspondence and are inversely related.

The mean average precision (mAP) is the average of the average precision of all classes.

The average precision Average (AP) is obtained by averaging precision rates corresponding to the recall rates in each category.

For example, a defect classification training set is input into a neural network model for training, a graph obtained based on a visualization tool wanbd is shown in fig. 7, the graph is established by taking call as a horizontal axis and precision as a vertical axis, and a positive sample is a target box corresponding to a defect classification label ok; the negative sample is a target frame corresponding to the defect classification label 'ng'; and when the Intersection over Union (iou) value is greater than 0.5 and the defect classification labels are the same, judging that the identified defect classification result is correct.

Wherein iou refers to the ratio of the intersection of the two target frames to the union of the two target frames. In other embodiments, the iou value may also be 0.7, and the implementation manner of the iou is not limited in this embodiment.

As shown in fig. 7, the category "ok" refers to a sample identified as a positive sample among the actually positive samples, and the corresponding AP is 0.988; the category "ng" refers to the sample identified as a negative sample among the actually negative samples, and the corresponding AP is 0.883; the mAP for all classes was 0.910.

The recognition model needs to recognize the target image, and the typical image processing library can assist the recognition model in processing the target image, so that the processing speed of the recognition model on the target image is improved. The recognition model may thus be loaded using the image processing library before processing the image using the recognition model.

Optionally, before inputting each target image into the recognition model obtained based on the deep learning to obtain the target region where at least one target interval is located in the target image, the target frame to which each target interval belongs, and the defect classification result of the target frame, the method further includes: under the condition that the model format of the recognition model is not matched with the processing format of the image processing library, converting the model format into the processing format; the image processing library includes: open source computer vision library OpenCV, or HALCON.

Specifically, the recognition model is loaded by using OpenCV, when the recognition model is obtained by YOLOv5 training, the format of the recognition model is "py", and OpenCV cannot load the "py" model, and then the recognition model needs to be converted into the processing format of OpenCV by using export. The Onnx format. At this time, openCV may load the converted recognition model, and accelerate a data processing process of the converted recognition model by using a CUDA module carried by OpenCV, so as to improve a data processing speed of the converted recognition model.

For example, the recognition result obtained by loading the converted recognition model with OpenCV is shown in fig. 8, and a total of two enamel wires within 155 target intervals and intervals are detected, wherein 67 defect classification results indicate that the distance between the two enamel wires is smaller than the preset distance. At this time, two target frames 82 exist in the target area 81 of the target image loaded in OpenCV.

Step 503, determining that a doubling defect exists in the motor rotor under the condition that the defect classification result corresponding to at least one target image in the at least two target images indicates that the distance between the two enameled wires is smaller than a preset distance.

Each motor rotor corresponds to at least two target images, and when the defect classification result corresponding to one target image indicates that the distance between the two enamelled wires is smaller than the preset distance, the motor rotor has a doubling defect. At this time, the position where the motor rotor has the doubling defect can be recorded.

Optionally, in the at least two target images, determining that a doubling defect exists in the motor rotor when a defect classification result corresponding to the at least one target image indicates that the distance between the two enameled wires is smaller than a preset distance, includes: numbering at least two target images according to the acquisition time; and under the condition that the defect classification result corresponding to at least one target image indicates that the distance between the two enamelled wires is smaller than the preset distance, determining that the motor rotor has a doubling defect, and outputting the number of the target image and the defect coordinate of the target frame corresponding to the defect classification result.

The defect coordinates can be coordinates of a central point of the target frame, the width of the target frame and the height of the target frame; the defect coordinates may also be coordinates of the four vertices of the target box. The embodiment does not limit the implementation manner of the defect coordinates.

Since the target image is used for image acquisition of the interval positioned on the peripheral side of the motor rotor, an interval with a poor shooting angle may exist at the edge of the target image, but the interval conforms to the recognition standard of the model, so the recognition model still recognizes the interval, and at this time, the false detection rate of the model is high.

Optionally, in the at least two target images, determining that a doubling defect exists in the motor rotor when a defect classification result corresponding to the at least one target image indicates that the distance between the two enameled wires is smaller than a preset distance, includes: acquiring a first confidence coefficient of each target frame in a target area output by a recognition model; determining whether the first confidence of the target frame is smaller than a preset confidence; deleting the target frame with the first confidence coefficient smaller than the preset confidence coefficient and the defect classification result corresponding to the target frame; and determining that the motor rotor has a doubling defect under the condition that the defect classification result corresponding to at least one screened target frame indicates that the distance between the two enameled wires is smaller than the preset distance.

When the recognition model outputs the target frame, the first confidence degree corresponding to the target frame is also output. The higher the first confidence of the target frame, the lower the false detection rate, while the first confidence of the target frame located at the edge of the target image is generally lower. Therefore, the target frames are screened through the first confidence coefficient, the false detection rate can be reduced, and the effect of model identification is improved.

In the same target image, the same target interval may correspond to a plurality of mutually overlapped target frames, and one target interval only needs to obtain one target frame, so that the target frames can be screened.

Optionally, determining that a doubling defect exists in the motor rotor when the defect classification result corresponding to at least one target image in the at least two target images indicates that the distance between the two enameled wires is smaller than a preset distance, including: acquiring a first confidence coefficient of each target frame in a target area output by a recognition model; determining a reference target frame with the highest first confidence coefficient from the target area; calculating the intersection ratio between other target frames except the reference target frame in the target area and the reference target frame; deleting other target frames with the intersection ratio larger than or equal to the intersection ratio threshold; and determining that the motor rotor has a doubling defect under the condition that the defect classification result corresponding to at least one deleted target frame indicates that the distance between the two enamelled wires is smaller than the preset distance.

Specifically, other target frames that intersect the reference target frame with the highest first confidence level and have a coincidence ratio less than or equal to the coincidence ratio threshold may be deleted by using a Non-Maximum Suppression (NMS) algorithm.

The intersection ratio threshold may be pre-stored in the image capturing component, or may be input by a user, and the obtaining manner of the intersection ratio threshold is not limited in this embodiment.

Because the target intervals cannot be overlapped with each other, the target frames corresponding to the target intervals cannot have a larger overlapping area, and the target frames which are most matched with the target intervals can be obtained by screening through the first confidence coefficient and the intersection ratio, so that the recognition effect of the recognition model is improved.

After the doubling defect detection is performed on the motor rotor, the motor rotor with the doubling defect needs to be separated, so that the motor rotor with the doubling defect can be subjected to defect cause analysis or repeated inspection in the following process.

Optionally, in the at least two target images, when the defect classification result corresponding to the at least one target image indicates that the distance between the two enameled wires is smaller than a preset distance threshold, after it is determined that the motor rotor has a doubling defect, the method further includes: and sending a gripping instruction to the gripping device so that the gripping device moves the motor rotor from the detection position to the target position in response to the gripping instruction.

The target position is used for storing the motor rotor with the doubling defect.

Illustratively, the method for detecting the doubling defect of the motor rotor by using deep learning can also be applied to a visual defect detection process of the motor rotor, wherein the visual defect detection process comprises three different motor rotor defect detection stations and a classified placement station.

The first station is used for detecting dynamic balance, poor hook (flat hook, askew hook and fried hook), iron core (dynamic balance, length) and poor length of slot wedge of the motor rotor.

For example, the detection result of the first station is shown in fig. 9, and a total of 152 motor rotors are detected, wherein three unqualified products which should be rejected exist, and the corresponding mark on the software is "NG".

The second station is used for realizing the detection of the surface damage of the commutator, the finish turning of the length, the paint feeding in the groove and the burr detection in the groove. When the second station detection is carried out, the motor rotor needs to be electrified.

For example, the detection result of the second station is shown in fig. 10, and a total of 152 motor rotors are detected, wherein 67 unqualified products which should be rejected exist and are correspondingly marked as "NG"; the corresponding identification is 'RW' after 11 motor rotors which can be qualified after repair; and 9 defective hook products exist in the target image, and the defective hook rejection is closed at the moment, namely only the motor rotor with the defective hook is detected, but the motor rotor with the defective hook is not rejected.

And the third station is used for detecting the doubling defects of the motor rotor.

After the motor rotors sequentially pass through the three stations, the motor rotors enter the last classified placing station, the clamping device moves the motor rotors without defects to a storage area corresponding to qualified products, seals are printed on the surfaces of the rotor main bodies of the motor rotors, and the qualified motor rotors are placed; and moving the motor rotor with the defect to a storage area corresponding to the disqualification.

In practical implementation, after the detection of each station is completed, the unqualified motor rotor can be moved to the storage area corresponding to the unqualified product through the clamping device, and the implementation mode of storing the unqualified motor rotor is not limited in the embodiment.

In summary, in the method for detecting a doubling defect of a motor rotor by using deep learning provided by this embodiment, image acquisition is performed on each interval by obtaining an image acquisition component, so as to obtain at least two target images corresponding to each interval, where each target image includes image data of two enameled wires in at least one interval and the interval, each target image is input into an identification model obtained based on the deep learning, so as to obtain a target region where at least one target interval is located in the target image, a target frame to which each target interval belongs, and a defect classification result of the target frame, the target interval is an interval in at least one interval that meets an identification standard of the model, positions of the target frame are determined based on positions of the two enameled wires in the target interval, and when the defect classification result corresponding to at least one target image indicates that a distance between the two enameled wires is smaller than a preset distance, it is determined that the motor rotor has the doubling defect; the problem that the distance between two strands of enameled wires cannot be accurately grasped manually due to the fact that a commutator is small and manual detection efficiency is low, and therefore the detection effect on the rotor doubling defect is poor can be solved; the efficiency of identifying the target image by using the identification model obtained through deep learning is higher, and the model can accurately grasp whether the distance between two strands of enameled wires in the target image is smaller than the preset distance, so that the detection effect on the doubling defect of the motor rotor can be improved.

In addition, under the condition that the training numerical values are all larger than or equal to the corresponding preset standard numerical values, the recognition model obtained by the training is output, the performance of the recognition model obtained after the training can be ensured, and therefore the recognition effect of the recognition model is improved.

In addition, a reference target frame with the highest first confidence coefficient is determined from the target area; calculating the intersection ratio between other target frames except the reference target frame in the target area and the reference target frame; and deleting other target frames with the intersection ratio larger than or equal to the intersection ratio threshold value, so that each target interval is ensured to correspond to one target frame, the target frame which is most matched with the target interval is obtained, and the recognition effect of the recognition model is improved.

Optionally, the present application further provides a computer readable storage medium, in which a program is stored, and the program is loaded and executed by a processor to implement the method for detecting the motor rotor doubling defect using deep learning of the above method embodiment.

Optionally, the present application further provides a computer product, which includes a computer-readable storage medium, in which a program is stored, where the program is loaded and executed by a processor to implement the method for detecting a parallel connection defect of a motor rotor using deep learning in the above method embodiment.

The technical features of the embodiments described above may be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the embodiments described above are not described, but should be considered as being within the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims

1. A method for detecting the doubling defect of a motor rotor by using deep learning is characterized in that the motor rotor to be detected is placed at the detection position of a defect detection table, and an image acquisition assembly positioned at the defect detection table is suitable for acquiring images of the detection position; one end of the motor rotor is provided with a commutator, at least two raised hooks are uniformly arranged on the periphery of the commutator in a surrounding manner, a space is formed between every two adjacent hooks, and an enameled wire of the motor rotor is wound in from the space on one side of each hook and then is wound out from the space on the other side of the hook; the method comprises the following steps:

acquiring at least two target images corresponding to each interval by acquiring the images of each interval by the image acquisition assembly; wherein each target image comprises at least one interval and image data of two enamelled wires in the interval;

2. The method of claim 1, wherein the training process of the recognition model comprises:

3. The method of claim 2, wherein the training a pre-created neural network model using a defect classification training set to obtain the recognition model comprises:

4. The method according to claim 1, wherein before inputting each target image into the recognition model obtained based on deep learning to obtain a target region where at least one target interval is located in the target image, a target frame to which each target interval belongs, and a defect classification result of the target frame, the method further comprises:

5. The method according to claim 1, wherein the determining that the motor rotor has a doubling defect when the defect classification result corresponding to at least one of the at least two target images indicates that the distance between two enameled wires is smaller than a preset distance comprises:

6. The method according to claim 1, wherein when the defect classification result corresponding to at least one target image in the at least two target images indicates that the distance between two enameled wires is smaller than a preset distance, it is determined that the motor rotor has a doubling defect, and the method comprises:

calculating intersection ratios between other target frames except the reference target frame in the target area and the reference target frame;

and determining that the motor rotor has a doubling defect under the condition that the defect classification result corresponding to at least one deleted target frame indicates that the distance between the two enamelled wires is smaller than the preset distance.

7. The method of claim 1, wherein an illumination assembly is further disposed on the defect inspection station, the illumination assembly adapted to provide supplemental illumination for the gap;

controlling the lighting assembly to start;

8. The method of claim 1, wherein a distance measuring assembly is further disposed on the defect detection station, the distance measuring assembly being adapted to measure a distance between the distance measuring assembly and the commutator;

obtaining a distance value collected by a distance measurement component;

9. The method of claim 8, wherein the defect inspection station comprises a rotary drive adapted to rotate the inspection location, and the axis of rotation of the inspection location is the axis of the motor rotor when placed on the inspection location;

alternatively, the first and second electrodes may be,

the method further comprises the following steps: under the condition that the image acquisition assembly is opposite to the hook, controlling the rotary driving piece to drive the detection position to rotate by a preset angle; controlling the distance measurement assembly to acquire the distance value so as to trigger and execute the distance value acquired by the distance measurement assembly again; determining whether the image capture assembly is opposite a space using the distance value; a step of performing image acquisition on the interval by using the image acquisition assembly under the condition that the image acquisition assembly is opposite to the interval; the preset angle is smaller than an included angle formed by the two adjacent hooks and the image acquisition assembly.

10. The method according to claim 1, wherein when the defect classification result corresponding to at least one of the at least two target images indicates that the distance between two enameled wires is smaller than a preset distance threshold, after it is determined that the doubling defect exists in the motor rotor, the method further comprises:

sending a gripping instruction to a gripping device to cause the gripping device to move the motor rotor from the detection position to a target position in response to the gripping instruction, the target position being used for storing the motor rotor having the doubling defect.