A method for detecting the access depth of a switch motorized static contact group

Abstract

The present invention relates to a method for detecting the access depth of a switch motor static contact group, comprising the following steps: 1) using a trained MASK-RCNN deep learning network model to identify an image to be detected of a switch motor static contact group in an access state, and generating a mask at the bolt position on the image to be detected; 2) using an interactive iterative algorithm to refine the mask; 3) according to the mask area, using a minimum area coverage method, obtaining an area to be corrected; 4) using a three-dimensional space correction technology to perform distortion correction on the area to be corrected, and obtaining a correction area; 5) according to the correction area, calculating the true distance of the bolts of the dynamic and static contact group; 6) obtaining the access depth of the dynamic and static contact group according to the corresponding relationship between the true distance of the bolts of the dynamic and static contact group and the access depth of the dynamic and static contact group. Compared with the prior art, the present invention has the advantages of fast detection speed, high detection efficiency, high precision, etc.

Description

Technical Field

[0001] The invention relates to the technical field of urban rail transit switch detection technology, in particular to a method for detecting access depth of a switch motor static contact group. Background Art

[0002] With the rapid development of urbanization and urban rail transit, subway passenger flow is increasing, the frequency of operation is becoming more and more frequent, and the speed is getting faster and faster, which puts higher requirements on the reliability, safety and maintainability of subway signal equipment. Switch machines refer to important signal equipment that is directly related to driving safety and is used to reliably switch the position of switches, change the direction of switch opening, lock the switch point rail, and reflect the position of switches. Taking Shanghai Metro as an example, there are 10 failures in the entire network caused by problems related to the dynamic and static contacts inside the switch machines every year, accounting for 9-11% of the switch machine failure rate, which is mainly closely related to the deterioration of the structural parts of the dynamic and static contact group.

[0003] In the fault diagnosis and maintenance of switch machines, there are currently the following methods:

[0004] (1) Manual periodic inspection. Since the internal structure of the switch machine is complex and most of them are outdoors, regular inspections are currently mainly carried out based on the work experience of maintenance personnel. However, manual maintenance work is inefficient and labor costs are high; the maintenance process is uncontrollable, and closed-loop supervision and management of the manual maintenance process is impossible; the accuracy of status detection is poor, and traditional manual maintenance methods have poor accuracy in judging the status of the equipment, and are greatly affected by human factors; paper-based maintenance data brings inconvenience to equipment status analysis and prediction.

[0005] (2) Existing detection technology methods: such as model-based, state-based, expert system-based, Bayesian network, support vector machine and other fault diagnosis methods. These methods all require rich and complete prior knowledge to support traditional fault diagnosis methods. They rely heavily on prior knowledge. Since turnouts are widely distributed along the railway, the external environment and working conditions vary greatly, resulting in a variety of fault modes and causes. It is difficult to comprehensively diagnose all faults.

[0006] (3) The existing monitoring system is a gap detection system, which determines whether the switch machine equipment is operating normally by monitoring the gap offset of the indicator rod. It has a low level of intelligence, and there are cases of misdiagnosis and missed diagnosis. The diagnostic efficiency is low. In actual application, the intervention of on-site personnel is required. The professional quality and experience of on-site personnel have a great influence on the diagnostic results.

[0007] (4) The main research objects of existing detection technologies are: rail fasteners, switch machine voltage, current, power and other electrical signals. Although there are related studies based on deep learning and convolutional neural networks, machine vision technology has not been used to perform fine segmentation of key positions in the image. Neither the algorithms nor the models have been studied based on the key positions after segmentation. Summary of the invention

[0008] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for detecting the access depth of a switch motor-static contact group, which has fast detection speed, high detection efficiency and high accuracy.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] A method for detecting the access depth of a switch motor-static contact group comprises the following steps:

[0011] 1) The trained MASK-RCNN deep learning network model is used to identify the image to be detected of the switch motor and static contact group in the access state, and a mask is generated at the bolt position on the image to be detected;

[0012] 2) Refine the mask using an interactive iterative algorithm;

[0013] 3) According to the mask area, the area to be corrected is obtained by the minimum area coverage method;

[0014] 4) Using three-dimensional space correction technology to correct the distortion of the area to be corrected to obtain the correction area;

[0015] 5) Calculate the actual distance between the bolts of the dynamic and static contact groups according to the correction area;

[0016] 6) The access depth of the dynamic and static contact group is obtained according to the corresponding relationship between the actual distance between the bolts of the dynamic and static contact group and the access depth of the dynamic and static contact group.

[0017] Furthermore, the training process of the MASK-RCNN deep learning network model includes:

[0018] Acquire a number of sample images of the switch motor-static contact group in the connected state to form an image data set;

[0019] Label the sample images and divide the image dataset into training set and test set;

[0020] Use the training set and test set to train and test the MASK-RCNN deep learning network model;

[0021] Among them, a group of moving contact fixing bolts and two groups of static contact fixing bolts in the sample image are marked.

[0022] Furthermore, the interactive iterative algorithm is used to refine the mask and obtain six mask coordinate information, which are (X LU0 , Y LU0 )、(X RU0 , Y RU0 )、(X LM0 , Y LM0 )、(X RM0 , Y RM0 )、(X LD0 , Y LD0 ) and (X RD0 , Y RD0 ), the coordinate value of the mask coordinate information is a pixel value;

[0023] The Y coordinate relationship of the six mask coordinate information is:

[0024] |Y LD0 |>|Y LM0 |>|Y LU0 |

[0025] |Y RD0 |>|Y RM0 |>|Y RU0 |

[0026] The X coordinate relationship of the six mask coordinate information is:

[0027] |X LU0 |<|X RU0 |

[0028] |X LM0 |<|X RM0 |

[0029] |X LD0 |<|X RD0 |.

[0030] Furthermore, the correction area is six mask coordinate information, which are (X LU1 , Y LU1 )、(X RU1 , Y RU1 )、(X LM1 ,Y1)、(X RM1 , Y RM1 )、(X LD1 , Y LD1 ) and (X RD1 , Y RD1 ), the coordinate value of the mask coordinate information is a pixel value;

[0031] The Y coordinate relationship of the six mask coordinate information is:

[0032] |Y LD1|>|Y LM1 |>|Y LU1 |

[0033] |Y RD1 |>|Y RM1 |>|Y RU1 |

[0034] The X coordinate relationship of the six mask coordinate information is:

[0035] |X LU1 |<|X RU1 |

[0036] |X LM1 |<|X RM1 |

[0037] |X LD1 |<|X RD1 |.

[0038] Furthermore, the actual distance R between the bolts of the dynamic and static contact groups in the connected state is calculated according to the following formula: DJ :

[0039]

[0040] Among them, D JJ is the pixel value distance of the bolts of the two static contact groups with the largest relative distance, D DJ is the pixel value distance between the bolts of the moving and static contact groups when the moving and static contact groups are connected, R JJ The actual distance between the bolts of the two static joints with the largest relative distance is obtained by using pixel-level length anchoring and proportional scaling between structural parts. JJ , D DJ and R JJ .

[0041] Furthermore, the calculation formula for the distance in pixel values ​​between the two static contact groups with the largest relative distance and the distance in pixel values ​​between the dynamic and static contact groups when the dynamic contact is connected to the static contact group is:

[0042] D JJ =|Y RD1 |-|Y RU1 |

[0043] D DJ =min{|Y RD1 |-|Y RM1 |,|Y RM1 |-|Y RU1 |}

[0044] or:

[0045] D JJ =|YLD1 |-|Y LU1 |

[0046] D DJ =min{|Y LD1 |-|Y LM1 |,|Y LM1 |-|Y LU1 |}

[0047] The actual spacing between the bolts of the two static contact groups with the maximum relative distance is the value actually used in the project.

[0048] Furthermore, the corresponding relationship between the actual distance between the bolts of the dynamic and static contact groups and the access depth of the dynamic and static contact groups is fitted by a power function, and the expression of the corresponding relationship is:

[0049]

[0050] Among them, a i is the coefficient, and L is the standard insertion depth of the dynamic and static contacts.

[0051] Furthermore, the three access depth data of the switch motor static contact group in a row are L from left to right. L , L M and L R ,

[0052] Furthermore, the specific process of the minimum area coverage method includes:

[0053] The center coordinates and radius of the circle with the smallest area are obtained in the mask area, and the abstract points are further obtained to generate the area to be corrected.

[0054] Furthermore, the specific process of the three-dimensional space correction technology includes:

[0055] According to the four-corner mapping principle, the projection path from the deformed surface to the corrected plane is established;

[0056] Generate a perspective transformation affine matrix based on the projection mapping path;

[0057] The distorted image pixel correction is completed through the affine matrix.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] (1) The present invention uses a convolutional neural network to identify and locate the key positions of the dynamic and static contact group and generates a mask. It uses an interactive iterative algorithm to finely segment the bolt area of ​​the dynamic and static contact group. It adopts a minimum area coverage method to abstract the key positions and obtain the area to be corrected. It uses a three-dimensional space correction algorithm to reduce or even eliminate spatial distortion. It uses pixel-level length anchoring and geometric scaling of structural member features to obtain accurate data such as the shape and displacement of the structural member. It applies image processing technology based on machine learning to the railway system, solves the problems of low manual efficiency and poor accuracy judgment in the monitoring of the access depth of the dynamic and static contact group in the prior art, and realizes intelligent and efficient detection of the access depth of the dynamic and static contact group of the switch, with fast detection speed, high detection efficiency and high accuracy.

[0060] (2) The present invention uses an interactive iterative algorithm to refine the mask key to obtain more complete and accurate bolt coordinate position information, thereby improving the detection accuracy;

[0061] (3) The present invention obtains the area to be corrected by the minimum area coverage method based on the mask area, finds the circle with the smallest area on the mask coverage surface generated after the secondary segmentation, finds the coordinates and radius of the center of the covering mask, further obtains the abstract point, generates the area to be corrected, and can more accurately obtain the coordinates and radius of the center point of the bolt, reduce the jagged and scattered interference caused by the secondary segmentation, and improve the detection accuracy;

[0062] (4) Use three-dimensional space correction technology to correct the distortion of the area to be corrected, obtain the correction area, reduce or even eliminate spatial distortion, reduce the data error caused by image distortion, and improve detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION

[0064] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0065] A method for detecting the access depth of a switch motor static contact group, such as Figure 1 , including the following steps:

[0066] 1) The trained MASK-RCNN deep learning network model is used to identify the image to be detected of the switch motor and static contact group in the access state, and a mask is generated at the bolt position on the image to be detected;

[0067] 2) Refine the mask using an interactive iterative algorithm;

[0068] 3) According to the mask area, the area to be corrected is obtained by the minimum area coverage method;

[0069] 4) Using three-dimensional space correction technology to correct the distortion of the area to be corrected to obtain the correction area;

[0070] 5) Calculate the actual distance between the bolts of the dynamic and static contact groups according to the correction area;

[0071] 6) The access depth of the dynamic and static contact group is obtained according to the corresponding relationship between the actual distance between the bolts of the dynamic and static contact group and the access depth of the dynamic and static contact group.

[0072] The MASK-RCNN convolutional neural network is a deep learning network that can identify and locate the key position (bolts) with high quality and generate bolt masks. In step 1), the training process of the MASK-RCNN deep learning network model includes:

[0073] Select and classify the image datasets required for the project, and label the training and test sets;

[0074] A deep network learning model based on MASK-RCNN is constructed, which is mainly used to identify and locate bolts that fix the relative positions of switch machine structural parts;

[0075] Initialize the parameters, image path, image address, etc. of the deep network learning model;

[0076] Using the classified training and test sets, the MASK-RCNN deep learning network model is trained and tested until the performance reaches the predicted value. The abnormal image information set is also trained.

[0077] Among them, a group of moving contact fixing bolts and two groups of static contact fixing bolts in the sample image are marked. The switch machine accounts for more than 75% of the image. The image is clear and the light is bright and appropriate. The image information is judged. For abnormal image information samples, they are not involved in the calculation and the information is acquired again. In addition, an abnormal image information sample set should be constructed, and a separate model training should be performed on the abnormal image set to increase the model recognition.

[0078] In step 1), the bolt masks generated in this embodiment are of two types, six in total, namely, two moving contact bolt masks and four stationary contact bolt masks.

[0079] In step 2), an interactive iterative algorithm is used to perform secondary segmentation on the two types of six masks generated in step 5, and the mask key is refined to obtain more complete and accurate bolt coordinate position information.

[0080] The main function of the secondary segmentation function used is to generate a rectangular frame with the center of the bolt as the center and a circumscribed circle with a radius of 1.5 times the circle based on the mask generated by MASK-RCNN. The bolt area is secondary segmented with the rectangular frame and superimposed on the mask position generated by MASK-RCNN. The more times the rectangular frame and MASK-RCNN mask are superimposed, the more accurate the generated coordinate position will be.

[0081] After refining the mask using the interactive iterative algorithm, six mask coordinate information are obtained, namely, the upper left (X LU0 , Y LU0 ), Upper Right (X RU0 , Y RU0 ), Left Middle (X LM0 , Y LM0 )、Right Middle(X RM0 , Y RM0 ), Lower Left (X LD0 , Y LD0 ) and lower right (X RD0 , Y RD0 ), the coordinate value of the mask coordinate information is the pixel value;

[0082] The Y coordinate relationship of the six mask coordinate information is:

[0083] |Y LD0 |>|Y LM0 |>|Y LU0 |

[0084] |Y RD0 |>|Y RM0 |>|Y RU0 |

[0085] The X coordinate relationship of the six mask coordinate information is:

[0086] |X LU0 |<|X RU0 |

[0087] |X LM0 |<|X RM0 |

[0088] |X LD0 |<|X RD0 |.

[0089] In step 3), the specific process of the minimum area coverage method includes:

[0090] Find the circle with the smallest area on the mask coverage surface generated after the secondary segmentation, find the coordinates and radius of the center of the covering mask, and more accurately obtain the coordinates and radius of the center point of the bolt, reducing the jagged and scattered interference caused by the secondary segmentation; further obtain abstract points and generate the area to be corrected.

[0091] In step 4), the specific process of the three-dimensional space correction technology includes:

[0092] According to the four-corner mapping principle, the projection path from the deformed surface to the corrected plane is established;

[0093] Generate a perspective transformation affine matrix based on the projection mapping path;

[0094] The distorted image pixel correction is completed through the affine matrix.

[0095] Due to the non-horizontal shooting angle and non-standard operation of the staff when the image is acquired on site, the image is folded and distorted during the image size compression transformation and format conversion process of the image data set. In this embodiment, in order to reduce the data error caused by image distortion, three-dimensional space correction technology is used to reduce or even eliminate spatial distortion.

[0096] The correction area is six mask coordinate information, namely the upper left (X LU1 , Y LU1 ), Upper Right (X RU1 , Y RU1 ), Left Middle (X LM1 , Y1), middle right (X RM1 , Y RM1 ), Lower Left (X LD1 , Y LD1 ) and lower right (X RD1 , Y RD1 ), the coordinate value of the mask coordinate information is the pixel value;

[0097] The Y coordinate relationship of the six mask coordinate information is:

[0098] |Y LD1 |>|Y LM1 |>|Y LU1 |

[0099] |Y RD1 |>|Y RM1 |>|Y RU1 |

[0100] The X coordinate relationship of the six mask coordinate information is:

[0101] |X LU1 |<|X RU1 |

[0102] |X LM1 |<|X RM1 |

[0103] |X LD1 |<|X RD1 |.

[0104] Calculate the actual distance R between the bolts of the dynamic and static contact groups in the connected state according to the following formula DJ :

[0105]

[0106] Among them, D JJ is the pixel value distance of the bolts of the two static contact groups with the largest relative distance, D DJ is the pixel value distance between the bolts of the moving and static contact groups when the moving and static contact groups are connected, R JJ The actual distance between the bolts of the two static joints with the largest relative distance is obtained by using pixel-level length anchoring and proportional scaling between structural parts. JJ , D DJ and R JJ .

[0107] The calculation formula for the pixel value distance between the bolts of the two static contact groups with the largest relative distance and the pixel value distance between the bolts of the dynamic and static contact groups when the dynamic contact is connected to the static contact group is:

[0108] D JJ =|Y RD1 |-|Y RU1 | (2)

[0109] D DJ =min{|Y RD1 |-|Y RM1 |,|Y RM1 |-|Y RU1 |} (3)

[0110] or:

[0111] D JJ =|Y LD1 |-|Y LU1 | (4)

[0112] D DJ =min{|Y LD1 |-|Y LM1 |,|Y LM1 |-|Y LU1 |} (5)

[0113] D JJ Choose to use formula (2) to calculate, then D DJ Choose to use formula (3) to calculate, D JJ Choose to use formula (4) to calculate, then D DJ Choose to use formula (5) to calculate. DJ It is the value to be measured in this embodiment.

[0114] The actual spacing between the bolts of the two static contact groups with the largest relative distance is the value actually used in the project.

[0115] The corresponding relationship between the actual distance between the bolts of the dynamic and static joint group and the access depth of the dynamic and static joint group is fitted by the power function. The expression of the corresponding relationship is:

[0116]

[0117] Among them, a i are the coefficients of the fitting function, and L is the standard insertion depth of the dynamic and static contacts.

[0118] The corresponding relationship of this embodiment is:

[0119] L=-3.0023+2.8735*R DJ -0.5648*R DJ *R DJ

[0120] The three access depth data of the switch motor static contact group in a row are L from left to right. L , L M and L R ,like Then R DJ Select formula (1), formula (2) and formula (3) to calculate, if Then R DJ Using formula (1), formula (4) and formula (5) to calculate,

[0121] At this point, through machine learning methods, the access depth data L of the switch motor static contact group is obtained intelligently and accurately. L , L M and L R .

[0122] This embodiment proposes a method for detecting the access depth of a dynamic and static contact group of a switch. The method uses a convolutional neural network to identify and locate the key positions of the dynamic and static contact groups and generate masks. The interactive iterative algorithm is used to finely segment the bolt area of ​​the dynamic and static contact groups. The minimum area coverage method is used to abstract the key positions and obtain the area to be corrected. The three-dimensional space correction algorithm is used to reduce or even eliminate spatial distortion. The pixel-level length anchoring is used to obtain accurate data such as the shape and displacement of the structural parts through geometric scaling of the structural parts features. The method solves the problems of low manual efficiency and poor accuracy judgment in the prior art when monitoring the access depth of the dynamic and static contact groups. The method has the characteristics of fast detection speed, high detection efficiency and high precision.

[0123] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for detecting the access depth of a switch motor-static contact group, characterized in that: The following steps are involved: 1) Use the trained MASK-RCNN deep learning network model to identify the image to be detected of the switch motor and static contact group in the access state, and generate a mask at the bolt position on the image to be detected; 2) Refine the mask using an interactive iterative algorithm; 3) According to the mask area, the area to be corrected is obtained by the minimum area coverage method; 4) Use three-dimensional space correction technology to correct the distortion of the area to be corrected to obtain the correction area; 5) Calculate the actual distance between the bolts of the dynamic and static contact groups according to the correction area; 6) The access depth of the dynamic and static contact group is obtained according to the corresponding relationship between the actual distance between the dynamic and static contact group bolts and the access depth of the dynamic and static contact group. After refining the mask using the interactive iterative algorithm, six mask coordinate information are obtained, which are (X LU0 , Y LU0 ), (X RU0 , Y RU0 )、(X LM0 , Y LM0 )、(X RM0 , Y RM0 )、(X LD0 , Y LD0 ) and (X RD0 , Y RD0 ), the coordinate value of the mask coordinate information is the pixel value; The Y coordinate relationship of the six mask coordinate information is: ， ， The X coordinate relationship of the six mask coordinate information is: ， ， ， The specific process of the minimum area coverage method includes: Obtain the center coordinates and radius of the circle with the smallest area in the mask area, further obtain the abstract point, and generate the area to be corrected. The specific process of the three-dimensional space correction technology includes: According to the four-corner mapping principle, the projection path from the deformed surface to the corrected plane is established; Generate a perspective transformation affine matrix based on the projection mapping path; The distorted image pixel correction is completed through the affine matrix.

2. A method for detecting the access depth of a switch motor-static contact group according to claim 1, characterized in that: The training process of the MASK-RCNN deep learning network model includes: Acquire a number of sample images of the switch motor-static contact group in the connected state to form an image data set; Label the sample images and divide the image dataset into training set and test set; Use the training set and test set to train and test the MASK-RCNN deep learning network model; Among them, a group of moving contact fixing bolts and two groups of static contact fixing bolts in the sample image are marked.

3. A method for detecting the access depth of a switch motor-operated static contact group according to claim 1, characterized in that: The correction area is six mask coordinate information, which are (X LU1 , Y LU1 ), (X RU1 , Y RU1 )、(X LM1 , Y1), (X RM1 , Y RM1 )、(X LD1 , Y LD1 ) and (X RD1 , Y RD1 ), the coordinate value of the mask coordinate information is the pixel value; The Y coordinate relationship of the six mask coordinate information is: ， ， The X coordinate relationship of the six mask coordinate information is: ， ， 。 4. A method for detecting the access depth of a switch motor-static contact group according to claim 3, characterized in that: Calculate the actual distance between the bolts of the dynamic and static contact groups in the connected state according to the following formula : ， in, is the pixel value distance of the bolts of the two static contact groups with the largest relative distance, is the pixel value distance between the bolts of the moving and static contact groups when the moving and static contact groups are connected. The actual spacing of the bolts of the two static joints with the largest relative distance is obtained by using pixel-level length anchoring and proportional scaling between structural parts. , as well as .

5. A method for detecting the access depth of a switch motor-static contact group according to claim 4, characterized in that: The calculation formula for the pixel value distance between the two static contact groups with the largest relative distance and the pixel value distance between the dynamic and static contact groups when the dynamic contact is connected to the static contact group is: ， ， or: ， ， The actual spacing between the bolts of the two static contact groups with the maximum relative distance is the value actually used in the project.

6. A method for detecting the access depth of a switch motor-static contact group according to claim 4, characterized in that: The corresponding relationship between the actual distance between the bolts of the dynamic and static contact groups and the access depth of the dynamic and static contact groups is fitted by a power function. The expression of the corresponding relationship is: ， in, is the coefficient, This is the standard insertion depth for dynamic and static contacts.

7. A method for detecting the access depth of a switch motor-static contact group according to claim 1, characterized in that: The three access depth data of the switch motor static contact group are as follows from left to right: , as well as , .

Images