Overhead wire fitting abnormality detection device

Abstract

To provide an overhead wire fitting abnormality detection device capable of stably detecting anomalies with high precision.SOLUTION: An overhead wire fitting abnormality detection device includes: a learning processing unit that learns using multiple one-hot expression index image data in normal conditions, which is generated by converting each pixel of the image data generated by photographing the detection target in the normal state; and a data processing unit that generates one-hot expression index image data by converting each pixel of the image data generated by photographing the detection target to perform anomaly determination using the one-hot expression index image data.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] The present invention relates to an overhead line fitting abnormality detection device that detects abnormalities in overhead line fittings using image data.

Background Art

[0002] Generally, visual inspections are performed for the management and maintenance of machinery and equipment, etc., but there is a demand for automation of this inspection. For example, in the field of railway equipment, attempts have been made to detect abnormalities in overhead line fittings using image data obtained by cameras on the roofs of running vehicles. The overhead line fitting refers to a fitting for holding an electric train line for supplying power to a vehicle in a prescribed position, electrically connecting or partitioning between a plurality of wire strands in electric railway equipment.

[0003] Conventionally, various techniques for performing abnormality detection using image data have been proposed. Non-Patent Document 1, which is an example of the prior art, discloses an ALOCC (Adversarially Learned One-Class Classifier) that learns only images of the normal class and performs abnormality detection. Also, Non-Patent Document 2, which is an example of the prior art, discloses a BiGANs (Bidirectional Generative Adversarial Networks) that learns an encoder that maps data to a latent space in an adversarial generative model called GANs (Generative Adversarial Networks). Also, Non-Patent Document 3, which is an example of the prior art, discloses an autoencoder called VAE (Variational Auto Encoder) that assumes a multi-dimensional Gaussian distribution for latent variables. Here, the autoencoder is a network that compresses the dimension of the input into latent variables using an encoder and restores the dimension-compressed latent variables using a decoder. Furthermore, Non-Patent Document 4, an example of prior art, discloses a method for recognizing the type of object at the pixel level. Furthermore, Patent Document 1, which is an example of prior art, discloses a technique that generates images using a generative model such as a neural network, calculates the similarity between the generated image and the captured image, and determines that an anomaly occurs if this similarity is small. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] M. Sabokrou et al., “Adversarially Learned One-Class Classifier for Novelty Detection,” [online], May 24, 2018, [Retrieved February 4, 2021], Internet.<URL: https: / / arxiv.org / pdf / 1802.09088.pdf> [Non-Patent Document 2] J. Donahue et al., “Adversarial Feature Learning,” [online], April 3, 2017, [Retrieved February 4, 2021], Internet.<URL: https: / / arxiv.org / pdf / 1605.09782v7.pdf> [Non-Patent Document 3] DPKingma et al., “Auto-Encoding Variational Bayes,” [online], May 1, 2014, [Retrieved February 4, 2021], Internet<URL: https: / / arxiv.org / pdf / 1312.6114v10.pdf> [Non-Patent Document 4] H. Zhao et al., “Pyramid Scene Parsing Network,” [online], April 27, 2017, [Retrieved February 4, 2021], Internet.<URL: https: / / arxiv.org / pdf / 1612.01105.pdf> [Patent Documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2019-133306 [Overview of the Initiative] [Problems that the invention aims to solve]

[0006] However, each of the above conventional technologies has the following problems. The technology disclosed in Non-Patent Document 1 uses RGB image data for anomaly detection, but it is difficult to converge the learning process and lacks stability.

[0007] The technology disclosed in Non-Patent Document 2 uses three networks—an encoder, a generator, and a discriminator—to perform adversarial learning. However, it is prone to a phenomenon called "Missing Mode," where the generator outputs almost identical data, and lacks stability.

[0008] The technology disclosed in Non-Patent Document 3 uses an autoencoder consisting of an encoder and a decoder to assume a multidimensional Gaussian distribution for the latent variables and to create a meaningful distribution of latent variables as a generative model. This enables anomaly detection using the distribution of latent variables. However, because it uses an autoencoder, it has a larger reconstruction error and lacks stability compared to GANs.

[0009] The technology disclosed in Non-Patent Document 4 performs recognition at the pixel level, but it is not applicable to anomaly detection.

[0010] The technology disclosed in Patent Document 1 uses a generative model to generate an image and the same image as the captured image to detect anomalies, determining that an anomaly is present if the similarity is low. However, because it uses the RGB image data itself for anomaly detection, it is difficult to train and lacks stability.

[0011] Thus, the technologies disclosed in Non-Patent Documents 1-3 and Patent Document 1 make it difficult to reliably perform highly accurate anomaly detection. The present invention has been made in view of the above, and aims to provide a technology that enables stable and highly accurate anomaly detection. [Means for solving the problem]

[0012] One aspect of the present invention that solves the above-mentioned problems and achieves the objective is an overhead line fitting abnormality detection device comprising: a learning processing unit that performs learning using a plurality of normal state one-hot representation index image data generated by transforming each pixel of image data generated by photographing overhead line fittings that are the target of detection in a normal state; and a data processing unit that generates one-hot representation index image data by transforming each pixel of image data generated by photographing overhead line fittings that are the target of detection, and performs abnormality determination using the one-hot representation index image data.

[0013] In the overhead line fitting abnormality detection device with the above configuration, the abnormality determination may be performed by ALOCC.

[0014] In the overhead line fitting abnormality detection device with the above configuration, the abnormality determination may be performed based on the restoration error between the input data, which is the one-hot representation index image data input to the encoder, and the output data output from the decoder, which receives the output from the encoder.

[0015] In the overhead line fitting abnormality detection device configured as described above, the abnormality determination may be performed based on the error between the input data, which is the one-hot representation index image data input to the generator, and the output data output from the generator.

[0016] In the overhead wire fitting abnormality detection device configured as described above, the abnormality determination may be performed based on the distribution and distance of the one-hot representation index image data in the latent variable space.

[0017] In the overhead line fitting abnormality detection device having the above-described configuration, the learning and the abnormality determination may be performed for each target part.

[0018] Alternatively, one aspect of the present invention is a learning processing unit that performs learning for each target part using RGB information of a plurality of one-hot representation index image data generated by converting each pixel of image data generated by photographing an overhead line fitting that is a detection target in a normal state, and a data processing unit that converts each pixel of image data generated by photographing the overhead line fitting that is a detection target to generate RGB information of one-hot representation index image data, and performs an abnormality determination for each target part using the RGB information of the one-hot representation index image data. The overhead line fitting abnormality detection device includes:

Advantages of the Invention

[0019] According to the present invention, highly accurate abnormality detection can be stably performed.

Brief Description of the Drawings

[0020] [Figure 1] FIG. 1 is a diagram showing an installation example of an overhead line fitting abnormality detection device according to Embodiment 1. [Figure 2] FIG. 2 is a block diagram showing the configuration of an overhead line fitting abnormality detection device according to Embodiment 1. [Figure 3] FIG. 3(A) is a diagram showing one-hot representation index image data converted by an image conversion unit in Embodiment 1, and FIG. 3(B) is a diagram showing 256-level 3-channel RGB image data before being converted by the image conversion unit in Embodiment 1. [Figure 4] FIG. 4(A) is a diagram showing the probability (%) of each class of index image data for one pixel in Embodiment 1, and FIG. 4(B) is a diagram showing 256-level 3-channel RGB image data and each channel for one pixel in Embodiment 1. [Figure 5] FIG. 5 is a flowchart showing the learning process of an abnormality detection model. [Figure 6]Figure 6 is a flowchart showing the anomaly detection process using a pre-trained anomaly detection model. [Figure 7] Figure 7 shows the concept of anomaly detection in Embodiment 1. [Figure 8] Figure 8 shows the concept of anomaly detection in Embodiment 2. [Figure 9] Figure 9 shows the concept of anomaly detection in Embodiment 3. [Figure 10] Figure 10 shows the concept of anomaly detection in Embodiment 4. [Figure 11] Figure 11 is a flowchart showing the learning process of the anomaly detection model in Embodiment 5. [Figure 12] Figure 12 is a flowchart showing the anomaly detection process using a trained anomaly detection model in Embodiment 5. [Figure 13] Figure 13 is a conceptual diagram illustrating anomaly detection in Embodiment 5. [Figure 14] Figure 14 is a flowchart showing the learning process of the anomaly detection model in Embodiment 7. [Figure 15] Figure 15 is a flowchart showing the anomaly detection process using a trained anomaly detection model in Embodiment 7. [Figure 16] Figure 16 is a conceptual diagram illustrating anomaly detection in Embodiment 7. [Modes for carrying out the invention]

[0021] Hereinafter, embodiments for carrying out the present invention will be described with reference to the attached drawings. However, the present invention is not limited by the following descriptions of embodiments.

[0022] <Embodiment 1> Figure 1 shows an example of the installation of the overhead line fitting abnormality detection device 13 according to this embodiment. Figure 1 shows a railway vehicle, vehicle 1, the railway track 2 on which vehicle 1 runs, and the overhead wire fittings 3 that are the target of anomaly detection. The vehicle 1 shown in Figure 1 is equipped with a camera 11, a storage device 12, an overhead line fitting abnormality detection device 13, and an interface 14.

[0023] Camera 11 is installed on the roof of vehicle 1, and photographs the overhead line fittings 3, outputting image data. Examples of cameras 11 include area sensor cameras and line sensor cameras. Furthermore, if necessary, a lighting device (not shown) that illuminates the overhead wire fittings 3 for the purpose of taking photographs with camera 11 may be provided.

[0024] The storage device 12 stores the image data output from the camera 11 and the index image data converted by the data processing unit 131.

[0025] Interface 14 includes an input unit that includes at least one of a pointing device and a keyboard, and an output unit that includes a display.

[0026] Figure 2 is a block diagram showing the configuration of the overhead line fitting abnormality detection device 13 according to this embodiment. The overhead line fitting abnormality detection device 13 comprises a data processing unit 131 and a learning processing unit 132, and performs abnormality detection of the overhead line fittings 3 based on image data stored in the storage device 12. The data processing unit 131 comprises a detection target detection and recognition processing unit 1311, an image conversion unit 1312, and an anomaly determination unit 1313, and performs an anomaly determination based on the image data stored in the storage device 12 using an anomaly determination model.

[0027] The detection target detection and recognition processing unit 1311 performs detection and recognition processing for each pixel of the image data stored in the storage device 12. For detecting the overhead wire fittings 3, for example, a fitting detection model can be used. For pixel-level recognition processing, for example, the technology described in Non-Patent Document 4 is used.

[0028] The metal fitting detection model is created by training the object detection model SSD (Single Shot Multibox Detector) using manually created ground truth data on overhead wire images. The correct answer data created here includes two pieces of data: the coordinates of the four vertices of the rectangular area surrounding the metal fittings in the overhead wire image, and the correct answer label. Furthermore, SSD consists of a base network, which is a layer that extracts image features, and a detector, which is a layer that detects objects, allowing for the creation of rectangular regions and class identification with a single network. Furthermore, by combining predictions from feature maps of different scales, the detector can accurately detect objects of various sizes. Furthermore, while this example demonstrates the use of an SSD in the metal fitting detection algorithm, the present invention is not limited to this. Examples of algorithms that can be used in metal fitting detection models include YOLO (You Only Look Once), Faster-RCNN (Regions with Convolutional Neural Networks), FPN-FRCN (Feature Pyramid Network-Faster Regions with Convolutional Neural Networks), and M2Det.

[0029] The image conversion unit 1312 converts the image data, which has undergone detection and recognition processing for each pixel of the overhead wire fitting 3, into an appropriate format. The image conversion unit 1312 converts the image data into, for example, one-hot representation index image data.

[0030] The learning processing unit 132 performs training on the anomaly detection model used in the data processing unit 131.

[0031] Furthermore, abnormalities in the overhead line fittings 3 in railway equipment primarily result in changes in shape rather than changes in texture. In this embodiment, an overhead wire fitting 3 in railway equipment is used as an example of application, but the detection target is not limited to overhead wire fittings, as long as it is possible to detect abnormalities due to shape changes from the image.

[0032] In the prior art disclosed in Non-Patent Documents 1-3 and Patent Document 1, RGB image data is used as is, so the data for each pixel is 256 gradations and 3 channels. Therefore, the search space for each pixel is 256. 3 As a result, the amount of information is large, anomaly detection is difficult, and stable operation is challenging.

[0033] Figure 3(A) shows the one-hot representation index image data converted by the image conversion unit 1312 in this embodiment, and Figure 3(B) shows the 256-level 3-channel RGB image data before conversion by the image conversion unit 1312 in this embodiment. As shown in Figure 3(A), the index image data has an index value as quantitative data [0,1] in each of the classes 0 to 9. As shown in Figure 3(B), the RGB image data has quantitative data [0 to 255] in each of the RGB3 channels.

[0034] Figure 4(A) shows the probability (%) of each class of index image data in a single pixel in this embodiment, and Figure 4(B) shows 256-level, 3-channel RGB image data and each channel in a single pixel in this embodiment. As shown in Figure 4(A), in index image data, the probability of each class is assigned, and the data is handled class by class, allowing for more appropriate weight updates. In Figure 4(A), when a value is close to Class 1 or Class 9, 1 is considered 50% and 9 is considered 50%. In Figure 4(B), each RGB channel uses a combination of consecutive numbers from 0 to 255, resulting in a wide range and a large number of possible combinations.

[0035] Figure 5 is a flowchart showing the training process of the anomaly detection model. First, when the learning process begins, the camera 11 takes a picture of the overhead wire fitting 3, which is the target of detection in a normal state, and outputs the image data to the storage device 12 (S1). The storage device 12 stores image data of the overhead wire fitting 3, which is the target of detection of a normal state, obtained by photography (S2). The data processing unit 131 performs detection and recognition processing on each pixel of the image data of the overhead wire fitting 3 in its normal state, which is stored in the storage device 12, and generates index image data (S3). Here, a fitting detection model is used for the detection and recognition process of the overhead wire fitting 3. The storage device 12 stores the generated index image data (S4). Processing steps S1 through S4 will continue until the end of shooting. The learning processing unit 132 uses the index image data of the overhead wire fittings 3 in their normal state, stored in the storage device 12, to perform the training process for the anomaly detection model (S5), and then terminates the process.

[0036] Figure 6 is a flowchart showing the anomaly detection process using a pre-trained anomaly detection model. First, camera 11 photographs the overhead wire fitting 3, which is the object to be detected, and outputs the image data to storage device 12 (S11). The storage device 12 stores the image data of the overhead wire fitting 3, which is the object to be detected, acquired by photography (S12). The data processing unit 131 performs detection and recognition processing on each pixel of the image data of the overhead wire fitting 3 stored in the storage device 12, and generates index image data (S13). Here, a fitting detection model is used for the detection and recognition process of the overhead wire fitting 3. The storage device 12 stores the generated index image data (S14). The abnormality detection unit 1313 uses the index image data of the overhead line fittings 3 stored in the storage device 12 to perform abnormality detection processing using a trained abnormality detection model (S15). If no abnormality is detected, it returns to S11. If an abnormality is detected (S16), it outputs to the interface (S17) and terminates the process. Here, anomaly detection in S15 is performed by ALOCC as shown below.

[0037] Figure 7 shows the concept of anomaly detection in this embodiment. In the ALOCC shown in Figure 7, the encoder, decoder, and discriminator undergo training, and anomaly detection is performed using the trained discriminator. For anomaly detection, image data restored by an encoder and decoder may be used, or image data that has not been restored may be used.

[0038] In this embodiment, by generating an index image that represents the image data in a one-hot format, the search space is kept small, making the learning process and anomaly detection process easier and thus enabling stable processing. Therefore, according to this embodiment, highly accurate anomaly detection can be performed stably.

[0039] <Embodiment 2> In Embodiment 1, ALOCC was used for anomaly detection, but the present invention is not limited thereto. For example, instead of using ALOCC for anomaly detection, one could take the difference between the input image and the output image and perform anomaly detection based on that difference, which is the reconstruction error. Since the method for detecting anomalies is the same as in Embodiment 1, the explanation will be omitted, and Embodiment 1 will be used as a reference.

[0040] Figure 8 shows the concept of anomaly detection in this embodiment. Figure 8 shows the encoder and decoder. As shown in Figure 8, input data is fed into the encoder, the output of the encoder is fed into the decoder, and output data is output from the decoder. If the difference between the input data and the output data, i.e., the reconstruction error, is greater than or equal to a predetermined value, it is judged as abnormal; if the reconstruction error is less than the predetermined value, it is judged as normal. The training process for the encoder and decoder may use a general autoencoder or adversarial learning approach.

[0041] In this embodiment as well, by generating an index image that represents the image data in one-hot mode, the search space is kept small, making learning and anomaly detection processes easier and thus enabling stable processing. Furthermore, according to this embodiment, it becomes possible to detect anomalies using the restoration error, which is a common method. Therefore, according to this embodiment, highly accurate anomaly detection can be performed stably.

[0042] <Embodiment 3> Embodiment 1 describes a configuration using ALOCC for anomaly detection, and Embodiment 2 describes a configuration using the reconstruction error, which is the difference between the input image and the output image. However, the present invention is not limited to these. For example, anomaly detection could be performed by taking the difference between the input image and the image generated by a generator that outputs the image closest to the input image, and then using that difference as the basis for anomaly detection. Since the method for detecting anomalies is the same as in Embodiment 1, the explanation will be omitted, and Embodiment 1 will be used as a reference.

[0043] Figure 9 shows the concept of anomaly detection in this embodiment. Figure 9 shows the generator. The generator shown in Figure 9 searches for image generation parameters and outputs the image closest to the input image by exploratoryally changing latent variables. Here, the search can be performed using a gradient-based method, but it is not limited to this.

[0044] In this embodiment as well, by generating an index image that represents the image data in one-hot mode, the search space is kept small, making learning and anomaly detection processes easier and thus enabling stable processing. Furthermore, according to this embodiment, it becomes possible to detect anomalies using a generator capable of high-precision restoration. Therefore, according to this embodiment, highly accurate anomaly detection can be performed stably.

[0045] <Embodiment 4> The methods of Embodiments 1 to 3 involve anomaly detection using an image feature space based on index image data that represents image data in a one-hot scan; however, the present invention is not limited thereto. Anomaly detection may be performed using a latent variable space obtained by converting the image feature space into latent variables, instead of the image feature space.

[0046] Figure 10 shows the concept of anomaly detection in this embodiment. In Figure 10, the input data is fed into the encoder to learn a normal multidimensional Gaussian distribution. The input data is then fed into a pre-trained encoder, and anomalies are detected based on the distance from the center of the Gaussian distribution. Here, the VAE disclosed in Non-Patent Document 3 was used to obtain the latent variables, but the present invention is not limited thereto.

[0047] This embodiment also reduces the search space by generating an index image that represents the image data in one-hot format, thereby facilitating learning and anomaly detection processes. Furthermore, according to this embodiment, anomaly detection based on probability distribution becomes possible. Therefore, according to this embodiment, highly accurate anomaly detection can be performed stably, and anomaly detection can be performed based on a probability distribution.

[0048] <Embodiment 5> While highly accurate anomaly detection can be reliably performed according to embodiments 1 to 4, it is difficult to pinpoint the location of the anomaly because anomaly detection is performed from the entire image. This embodiment describes an overhead line fitting abnormality detection device that can also identify the location where the abnormality occurred. In cases where the same reference numerals are used as in Embodiments 1 to 4, their descriptions are omitted, and Embodiments 1 to 4 are used by reference.

[0049] Figure 11 is a flowchart showing the learning process of the anomaly detection model in this embodiment. The processing in S1 to S4 is the same as in Figure 5. The data processing unit 131 extracts the target parts of the overhead wire fitting 3 by extracting an index for each target part (S6). The learning processing unit 132 uses the index image data of the target parts of the overhead wire fittings 3 in their normal state, stored in the storage device 12, to perform the learning process for the anomaly detection model (S5a), and then terminates the process.

[0050] Figure 12 is a flowchart showing the anomaly detection process using the trained anomaly detection model in this embodiment. The processes in S11 to S14 are the same as in Figure 6. The data processing unit 131 determines whether or not there is index image data for the target part (S21). If there is no index image data (S21:N), it detects a missing part anomaly for the target part (S22) and proceeds to S16. If index image data exists (S21:Y), the anomaly detection unit 1313 uses the index image data of the target part stored in the storage device 12 to perform anomaly detection processing using a trained anomaly detection model (S15), takes the difference between the input image and the output image, and determines whether the restoration error is greater than or equal to a threshold (S23). If the restoration error is greater than or equal to the threshold (S23:Y), the abnormality determination unit 1313 detects an abnormality (S24) and proceeds to S16. The abnormality detection unit 1313 returns to S11 if it does not detect an abnormality (S16), and outputs to the interface (S17) and terminates processing if it detects an abnormality (S16). Here, the anomaly detection process may be performed by ALOCC as described in Embodiment 1, instead of the anomaly detection process using the restoration error shown in S23.

[0051] Figure 13 is a conceptual diagram illustrating anomaly detection in this embodiment. As shown in Figure 13, in this embodiment, the original image is converted to one-hot representation index image data by one-hot encoding, and anomalies are detected using an anomaly detection model corresponding to each index.

[0052] As described above in this embodiment, by performing a learning process for each target part and performing anomaly detection for each target part using the learned anomaly detection model, it is possible to identify the location where the anomaly occurred.

[0053] <Embodiment 6> The method of Embodiment 5 is an anomaly detection method that uses an image feature space with index image data that represents image data in a one-hot format. However, even when performing anomaly detection for each target part, anomaly detection may be performed using a latent variable space converted to latent variables, similar to Embodiment 4.

[0054] According to this embodiment, when performing anomaly detection for each target part, the search space can be kept small, and the learning process and anomaly detection process can be made easier. Furthermore, according to this embodiment, it becomes possible to detect anomalies in each target part based on a probability distribution. Therefore, according to this embodiment, highly accurate anomaly detection can be stably performed for each target part, and anomaly detection can be performed based on a probability distribution.

[0055] <Embodiment 7> In embodiments 1 to 6, anomaly detection is performed based on shape, making it difficult to detect anomalies that appear as changes in texture, such as rust. This embodiment describes an overhead line fitting abnormality detection device that enables the detection of rust and other abnormalities by detecting texture abnormalities. In cases where the same reference numerals are used as in Embodiments 1 to 6, their descriptions are omitted, and Embodiments 1 to 6 are used by reference.

[0056] Figure 14 is a flowchart showing the learning process of the anomaly detection model in this embodiment. The processing in S1 to S4 is the same as in Figure 5. The data processing unit 131 extracts the target parts of the overhead line fitting 3 by extracting an index for each target part, similar to the embodiment 4 (S6a). The data processing unit 131 creates extracted image data (RGB information) for the target parts by performing a masking process on the original image using the index image data of the target parts extracted in S6a (S7), and then creates an anomaly detection model for each part. The learning processing unit 132 uses the index image data of the target parts of the overhead wire fittings 3 in their normal state, stored in the storage device 12, to perform the learning process for the anomaly detection model (S5a), and then terminates the process.

[0057] Figure 15 is a flowchart showing the anomaly detection process using the trained anomaly detection model in this embodiment. The processes in S11-S14 and S21-S22 are the same as in Figure 12. If index image data is available (S21:Y), the data processing unit 131 extracts the target part by masking (S31), and the anomaly detection unit 1313 performs anomaly detection processing using a trained anomaly detection model with the index image data of the target part stored in the storage device 12 (S15), and determines whether the distance of the latent variable from the center of the Gaussian distribution is greater than or equal to a threshold (S23a). If the distance of the latent variable from the center of the Gaussian distribution is greater than or equal to a threshold (S23a:Y), the anomaly detection unit 1313 detects an anomaly (S24) and proceeds to S16. The abnormality detection unit 1313 returns to S11 if it does not detect an abnormality (S16), and outputs to the interface (S17) and terminates processing if it detects an abnormality (S16).

[0058] Figure 16 is a conceptual diagram illustrating anomaly detection in this embodiment. As shown in Figure 16, in this embodiment, the original image is converted to one-hot representation index image data by one-hot encoding, and a region extraction image (RGB information) for each index is generated by masking each index. Anomalies are then detected using an anomaly detection model corresponding to the generated region extraction image (RGB information) for each index.

[0059] As described above in this embodiment, by performing a learning process for each extracted image data (RGB information) of the target part, it becomes possible to perform texture anomaly detection for each target part using the trained anomaly detection model. Therefore, it is possible to identify areas where abnormalities such as rust have occurred, regardless of their shape.

[0060] Furthermore, the present invention is not limited to the embodiments described above, and includes various modifications in which components are added, deleted, or replaced to the above-described configuration. [Explanation of Symbols]

[0061] 1 vehicle 11 Cameras 12 Storage device 13. Overhead wire fitting abnormality detection device 131 Data Processing Unit 1311 Detection and Recognition Processing Unit for Detecting Targets 1312 Image Conversion Unit 1313 Abnormality determination section 132 Learning Processing Unit 14 Interfaces 2 orbits 3. Overhead wire fittings

Claims

1. A learning processing unit that trains an anomaly detection model using a first one-hot representation index image data of multiple normal states, which is generated by transforming each pixel of the image data generated from photographing overhead line fittings that are the target of detection in a normal state, and The system includes a data processing unit that generates a second one-hot representation index image data by converting each pixel of the image data generated by photographing the overhead wire fittings that are the target of detection, and performs an anomaly determination using the anomaly detection model and the second one-hot representation index image data, The overhead wire fitting abnormality detection device is a multi-channel image in which the first one-hot representation index image data and the second one-hot representation index image data are separated into channels for each class of object to be detected.

2. The overhead line fitting abnormality detection device according to claim 1, characterized in that the abnormality determination is performed by ALOCC.

3. The aforementioned abnormality determination is, The input data, which is the second one-hot representation index image data input to the encoder, The overhead line fitting abnormality detection device according to claim 1, characterized in that the process is performed based on the restoration error between the output from the encoder and the output data output from the decoder that receives the input.

4. The aforementioned abnormality determination is, The overhead wire fitting abnormality detection device according to claim 1, characterized in that it is performed based on the distribution and distance of the second one-hot representation index image data in the latent variable space.

5. The overhead line fitting abnormality detection device according to any one of claims 1 to 4, characterized in that the learning and abnormality determination are performed for each target part.

6. The learning for each part is performed using the RGB information of the first one-hot representation index image data. The overhead line fitting abnormality detection device according to claim 5, characterized in that the abnormality determination for each part is performed using the RGB information of the second one-hot representation index image data.

Images