Data processing device, data processing method, and program

Abstract

A data processing device according to one embodiment of the present disclosure comprising: a data compression unit that generates compressed data by reducing a dimension of input data; a compact processing unit that maps at least a part of the compressed data on a compact manifold; and a reconstruction unit that uses the compressed data including the data mapped on the compact manifold to generate reconstruction data in which the input data is reconstructed.

Description

Technical Field

[0001] The present disclosure relates to a data processing apparatus, a data processing method, and a program.

Background Art

[0002] Regarding any data, techniques for determining the presence or absence of anomalies in the data have been evolving by using an AI (Artificial Intelligence) model or the like.

[0003] For example, Patent Document 1 discloses a biological signal analysis system that determines arrhythmia based on electrocardiogram information. This system converts electrocardiogram information to generate determination data, generates restored data by restoring the determination data using a learned model, and determines whether the determination data is normal waveform data or arrhythmia waveform data based on the difference between the restored data and the determination data.

[0004] Also, Patent Document 2 discloses a method for image clustering based on an autoencoder. In clustering, the autoencoder projects samples from the original image space to the feature space, and the samples are grouped and distributed for each class on the feature space.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0006] One proposed method involves training a system to generate compressed data by compressing the information of normal data, such as images, and then reconstruct the data based on that compressed data. During training, normal sample data is input into the system, and the system is trained to increase the degree of agreement between the sample data and the reconstructed data.

[0007] In the inference phase, the target data is input into the system, which compresses the data and then reconstructs it based on the compressed data. If the degree of agreement between the target data and the reconstructed data is high, it is assumed that the target data does not contain any data other than that included in the sample data. Therefore, the target data is judged to be normal. However, if the degree of agreement between the two is low, it is assumed that the target data contains data not included in the sample data. Therefore, the target data is judged to be abnormal.

[0008] However, even after training, it is conceivable that the degree of agreement between the system-reconstructed data and the target data may become high when the target data contains data that is not included in the sample data. In this case, the target data, which should be judged as abnormal, may be mistakenly judged as normal. The technologies described in Patent Documents 1 and 2 do not take such system problems into consideration and merely disclose judgment or classification techniques, and therefore cannot solve this problem.

[0009] One of the purposes of this disclosure is to provide a data processing device, a data processing method, and a program that can contribute to the detection of data anomalies. It should be noted that this purpose is only one of several purposes that the embodiments disclosed herein aim to achieve. Other purposes or problems and novel features will be revealed by the description herein or by the accompanying drawings. [Means for solving the problem]

[0010] A data processing device according to one aspect of the present disclosure includes a data compression unit that generates compressed data by reducing the dimension of input data, a compactification unit that maps at least a portion of the compressed data onto a compact manifold, and a reconstruction unit that generates reconstructed data in which the input data is reconstructed using the compressed data, which includes the data mapped onto the compact manifold.

[0011] A data processing method according to one aspect of the present disclosure is performed by a computer. The computer generates compressed data by reducing the dimensionality of the input data, maps at least a portion of the compressed data onto a compact manifold, and generates reconstructed data in which the input data is reconstructed using the compressed data, which includes the data mapped onto the compact manifold.

[0012] A program according to one aspect of the present disclosure causes a computer to generate compressed data by reducing the dimensionality of input data, map at least a portion of the compressed data onto a compact manifold, and generate reconstructed data in which the input data is reconstructed using the compressed data which includes the data mapped onto the compact manifold. [Effects of the Invention]

[0013] This disclosure provides a data processing device, a data processing method, and a program that can contribute to the detection of data anomalies. [Brief explanation of the drawing]

[0014] [Figure 1] Block diagram showing an example of a data processing device related to this disclosure. [Figure 2] This flowchart shows a typical example of processing performed by the data processing device described in this disclosure. [Figure 3] Block diagram showing an example of an anomaly detection device related to this disclosure. [Figure 4] This figure shows an example of converting latent variables to values ​​on a compact manifold. [Figure 5] It is a diagram showing another example of a compact manifold. [Figure 6] It is a diagram showing another example of a compact manifold. [Figure 7A] It is a diagram showing an example of input data and reconstructed data. [Figure 7B] It is a diagram showing another example of input data and reconstructed data. [Figure 8] It is a flowchart showing an example of typical processing of the abnormality determination device according to the present disclosure. [Figure 9A] It is a diagram showing processing for determining an abnormality in input data using related art. [Figure 9B] It is a diagram showing processing for determining an abnormality in input data using the abnormality determination device according to the present disclosure. [Figure 10] It is a block diagram showing an example of the hardware configuration of the device according to the present disclosure.

Embodiments for Carrying Out the Invention

[0015] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In the following description and drawings of the embodiments, for the sake of clarity of explanation, appropriate omissions and simplifications are made. Further, in the present disclosure, unless otherwise specified, when "at least any one of a plurality of items" is defined for a plurality of items, the definition may mean any one item or any plurality of items including all items. Furthermore, unless otherwise specified, the definitions of various terms in the present disclosure use the definitions of terms in mathematics.

[0016] Embodiment 1 FIG. 1 is a block diagram showing an example of a data processing device. The data processing device 10 includes a data compression unit 11, a compacting unit 12, and a reconstruction unit 13. Each part (each means) of the data processing device 10 is controlled by a control unit (controller) not shown. Hereinafter, each part will be described.

[0017] [Description of Configuration] The data compression unit 11 acquires input data and generates compressed data by reducing the dimensionality of the input data. The input data may be image data such as still images or videos, text data, or any other type of data. The input data can be data from various fields, such as security data or medical data.

[0018] Reducing the dimensionality of input data means reducing the dimensionality of a vector when the input data is represented as a vector. For example, if the input data is a still image, it means reducing the number of pixels. Therefore, by reducing the dimensionality of the input data, the amount of information in the input data is compressed.

[0019] The compactification unit 12 maps at least a portion of the compressed data onto a compact manifold. More specifically, when the compressed data is represented as a plurality of points in a certain space, the compactification unit 12 transforms some of the points into points on a compact manifold, which is a subset in Euclidean space. A compact manifold is a manifold that has a finite subcover for any open cover.

[0020] When compressed data is represented as multiple points existing in a certain space, the compactification unit 12 can map even points with extreme numerical values ​​in the space onto a compact manifold. Points with extreme numerical values ​​in the space refer, for example, to data that do not normally appear in the input data, or that rarely appear. Mapping a point onto a compact manifold means, for example, incorporating a point with an extreme numerical value in the space into a certain finite region. An example of a compact manifold will be described later in Embodiment 2.

[0021] The compactification unit 12 may map a portion of the compressed data onto a compact manifold, or it may map all of the compressed data onto a compact manifold.

[0022] The reconstruction unit 13 generates reconstructed data by using compressed data, which includes data mapped onto a compact manifold by the compactification unit 12, to reconstruct the input data. The dimensions of this reconstructed data are the same as the dimensions of the original input data. In other words, the reconstruction unit 13 restores the dimensions that were reduced by the data compression unit 11.

[0023] The reconstructed data generated by the reconstruction unit 13 is obtained by the compactification unit 12, which transforms points with extreme values ​​in space into points on a compact manifold. Therefore, in the reconstructed data, the appearance of anomalous data points that were present in the input data is suppressed.

[0024] Furthermore, at least one of the data compression unit 11, the compactification unit 12, and the reconstruction unit 13 may constitute, for example, an AI (Artificial Intelligence) model, or it may be configured by the control unit reading a predetermined algorithm. Specifically, the data compression unit 11 may be configured as an AI (Artificial Intelligence) model such as an encoder network in an autoencoder. The reconstruction unit 13 may be configured as an AI model such as a decoder network in an autoencoder. Hereinafter, the encoder network and decoder network will also be referred to simply as the encoder and decoder.

[0025] [Explanation of the process] Figure 2 is a flowchart illustrating a typical process of the data processing device 10, and this flowchart explains the processing of the data processing device 10. Details of each process are as described above, so their explanation is omitted here.

[0026] First, the data compression unit 11 of the data processing device 10 generates compressed data by reducing the dimensionality of the input data (Step S11: Compressed data generation step). Next, the compactification unit 12 maps at least a portion of the compressed data onto a compact manifold (Step S12: Mapping step). Then, the reconstruction unit 13 generates reconstructed data in which the input data has been reconstructed using the compressed data (Step S13: Reconstruction step).

[0027] [Explanation of effects] Through the processing described above, the data processing device 10 maps at least a portion of the compressed data used to generate the reconstructed data onto a compact manifold. At this time, even points with extreme values ​​in space are mapped onto the compact manifold, so the appearance of anomalous data points that appeared in the input data is suppressed in the reconstructed data. Therefore, by comparing the reconstructed data with the input data, the anomalous data points in the input data can be clearly identified.

[0028] In compressed data, we assume that even points with extreme values ​​in space are not mapped onto a compact manifold. When the reconstruction unit 13 generates reconstructed data using this compressed data, these anomalous data points may be directly reflected in the reconstructed data. In this case, even by comparing the reconstructed data with the input data, it is not possible to clearly determine the anomalous data points in the input data.

[0029] However, for the reasons stated above, the data processing device 10 can generate reconstructed data that clearly identifies abnormal data locations in the input data. Therefore, the data processing device 10 can contribute to the detection of abnormalities in the input data.

[0030] Furthermore, the processing performed by the data processing device 10 may be distributed and executed by multiple computers. In other words, the processing of the data processing device 10 may be implemented in a distributed system. In a distributed system, part or all of the data processing device 10 may be implemented on the cloud. However, the functions of the data processing device 10 may be installed in a single device.

[0031] Embodiment 2 In Embodiment 2, a specific example of the data processing device 10 described in Embodiment 1 is disclosed below. However, the specific example of the data processing device 10 shown in Embodiment 1 is not limited to the following. Furthermore, the configuration and processing described below are illustrative and not limited thereto.

[0032] [Explanation of the structure] Figure 3 is a block diagram showing an example of an anomaly detection device 20. The anomaly detection device 20 is a device that determines anomalies in input data and comprises a preprocessing unit 21, a data compression unit 22, a compactification unit 23, a reconstruction unit 24, and an anomaly detection unit 25.

[0033] The anomaly detection device 20 is used in both the learning phase and the inference phase. The data compression unit 22, the compactification unit 23, and the reconstruction unit 24 are each blocks that constitute the AI ​​model, and each block consists of one or more layers. The preprocessing unit 21 to the reconstruction unit 24 are blocks used in the learning phase. In particular, the data compression unit 22 to the reconstruction unit 24 are blocks that are trained as an AI model in the learning phase. In the learning phase, the autoencoder method is used.

[0034] After learning is complete, in the inference phase, the preprocessor 21 receives input data that is subject to anomaly detection. In this inference phase, the preprocessor 21 to the anomaly detection unit 25 are used to determine anomalies in the input data. Details of the learning and inference phases described above will be explained later. Below, the details of the processing of each part of the anomaly detection device 20 will be explained.

[0035] The preprocessor unit 21 acquires the input data. The method by which the preprocessor unit 21 acquires the input data is arbitrary. For example, the preprocessor unit 21 may receive the input data from a device other than the anomaly detection device 20, such as a camera, or it may acquire the input data stored in the storage of the anomaly detection device 20.

[0036] In this example, the input data acquired by the preprocessing unit 21 is a still image taken of a pedestrian-only area. However, the input data is not limited to this and may be any data, such as moving images.

[0037] The preprocessing unit 21 performs preprocessing on the acquired input data in order to shorten the processing time required by the data compression unit 22 and beyond, or to enable accurate anomaly detection. For example, the preprocessing unit 21 may convert color image input data to a grayscale image. Alternatively, the preprocessing unit 21 may convert large-sized input data to a smaller size suitable for processing.

[0038] The data compression unit 22 corresponds to the data compression unit 11 according to Embodiment 1. The data compression unit 22 acquires an image of the input data and generates compressed data by reducing the dimensionality of the input data. For example, if the original input data image contains tens of thousands of pixels, the input data is represented in vector form as information of tens of thousands of digit sequences. The data compression unit 22 can compress this digit sequence to, for example, information of several hundred digit sequences. Each digit in this information is defined as a latent variable, and the space to which the latent variables are mapped is defined as the latent space. At this time, the data compression unit 22 performs a process to compress the latent space into a lower-dimensional space.

[0039] The compactification unit 23 corresponds to the compactification unit 12 according to Embodiment 1. The compactification unit 23 transforms points in the latent space of a portion of the compressed data generated by the data compression unit 22 into points on a compact manifold in Euclidean space. The definition of a compact manifold is as described in Embodiment 1.

[0040] In the latent space of the compressed data generated by the data compression unit 22, each value of the latent variable lies on an infinitely extending number line. In other words, the latent variables can take on extremely large or small values ​​(extreme values). Even such latent variables can be transformed into values ​​on a compact manifold by the compactification unit 23.

[0041] Figure 4 shows an example of converting latent variables to values ​​on a compact manifold. In this example, the compact manifold is defined as follows:

number

[0042] Furthermore, the compact manifold that serves as the target of the latent space value transformation is not limited to the above definition, and various other definitions are possible. Examples are explained below.

[0043] Figure 5 shows another example of a compact manifold. In this example, a compact manifold is defined as follows:

number

[0044] Figure 6 shows another example of a compact manifold. In this example, a compact manifold is defined as follows:

number

[0045] Furthermore, the compactification unit 23 may or may not satisfy the injective condition for the mapping transformation using the compact manifold. However, in order to reliably load the values ​​of the latent space into a finite region through the transformation, it is preferable that the variables of the transformed object may be duplicated. Therefore, it is preferable that the mapping transformation using the compact manifold does not satisfy the injective condition.

[0046] The reconstruction unit 24 corresponds to the reconstruction unit 13 according to Embodiment 1. The reconstruction unit 24 generates reconstructed data in which the input data is reconstructed using compressed data which includes data mapped onto a compact manifold by the compactification unit 23. For example, if the input data image is represented as information of tens of thousands of digits, and this digit sequence is compressed to information of several hundred digits by the data compression unit 22, the reconstruction unit 24 generates reconstructed data of the image which is represented as information of tens of thousands of digits. In this way, the reconstructed data reproduces the same amount of information as the preprocessed input data and is data that can be compared with the input data.

[0047] The data compression unit 22 and the reconstruction unit 24 constitute the encoder and decoder, respectively, in the autoencoder.

[0048] (Learning Phase) In the learning phase, sample input data for learning is input to the preprocessing unit 21, and each AI model from the data compression unit 22 to the reconstruction unit 24 is simultaneously trained, optimizing the parameters that make up each AI model. Here, the reconstructed data output from the reconstruction unit 24 is trained to be closer to the sample input data that has been preprocessed in the preprocessing unit 21.

[0049] In other words, in this learning phase, as unsupervised learning, the edge weights in the layers of the data compression unit 22 and the reconstruction unit 24 are set so that the preprocessed sample input data and the reconstructed data match as closely as possible. In this learning phase, the data compression unit 22 is trained to extract the information necessary for image reconstruction from the input data. The reconstruction unit 24 is then trained to reconstruct an image from the compressed data that is similar to the input data.

[0050] In this example, the sample input data is a still image taken in a pedestrian-only area. The image in this sample input data includes objects that are not unusual to be in a pedestrian-only area, such as pedestrians and buildings within the area, but does not include objects that are unusual to be in a pedestrian-only area, such as cars and bicycles.

[0051] Through training using this sample input data, the trained AI models of the data compression unit 22 to reconstruction unit 24 become capable of accurately reconstructing pedestrians and other objects captured in the input data in the reconstructed data. However, vehicles are not included in the training. Therefore, it is assumed that even if the trained AI models of the data compression unit 22 to reconstruction unit 24 acquire input data in which a vehicle is captured in the image, they will not be able to accurately reconstruct that vehicle in the reconstructed data.

[0052] (Inference Phase) Next, in the inference phase, the image of the input data to be anomaly detection is input to the preprocessing unit 21. Each AI model from the preprocessing unit 21 to the reconstruction unit 24 performs the processing described above on this input data. As a result of the processing, the reconstruction unit 24 generates reconstructed data corresponding to the input data.

[0053] Here, if no abnormal objects are captured in the input image, the reconstruction unit 24 can generate reconstruction data that approximates the input data. Specifically, let's assume a case where the difference between the pixel value of a certain pixel in the input image and the pixel value of the corresponding pixel in the reconstruction image is calculated. When this difference is added for all or some pixels of the input (or reconstruction) image, if no abnormal objects are captured in the input image, the sum of the calculated differences will be less than a predetermined threshold. This sum of differences is also called the reconstruction error.

[0054] On the other hand, if an abnormal object is captured in the input image, the reconstruction unit 24 is expected to be unable to accurately reconstruct the abnormal object in the reconstructed data. In this case, the reconstruction error between the input data and the reconstructed data will exceed a predetermined threshold.

[0055] Figures 7A and 7B show examples of input data and reconstructed data. In Figure 7A, when a person P is visible in the image of input data I1, person P is accurately reconstructed in the reconstructed data I2. In contrast, in Figure 7B, when both person P and vehicle V are visible in the image of input data I1, person P is accurately reconstructed in the reconstructed data I2, but vehicle V is not accurately reconstructed. Therefore, the reconstruction error is small in the example in Figure 7A, while it is large in the example in Figure 7B.

[0056] The anomaly detection unit 25 calculates the reconstruction error shown above for the input data and the reconstructed data, and compares the calculated reconstruction error with a predetermined threshold. The details of the comparison calculation of the relationship between the reconstruction error and the predetermined threshold performed by the anomaly detection unit 25 are as described above. If this reconstruction error is less than the predetermined threshold, the anomaly detection unit 25 determines that the image of the input data does not contain any abnormal objects, just like the sample input data. In other words, the anomaly detection unit 25 determines that the image of the input data shows a normal scene. However, if the reconstruction error is greater than or equal to the predetermined threshold, the anomaly detection unit 25 determines that the image of the input data contains abnormal objects that were not present in the sample input data. In other words, the anomaly detection unit 25 determines that the image of the input data shows an abnormal scene. In this example, the anomaly detection unit 25 functions as a block that uses the characteristics of a trained AI model to determine anomalies in images taken in pedestrian-only areas, i.e., the presence or absence of objects such as vehicles.

[0057] The abnormality detection unit 25 may calculate the reconstruction error for a portion of the input data or the reconstructed data, or it may include all of the data. Furthermore, the parameters used by the abnormality detection unit 25 when determining the presence or absence of abnormal objects in the input data are not limited to the above-mentioned reconstruction error, as long as they reflect the difference between corresponding pixel values ​​in the input data and the reconstructed data.

[0058] Furthermore, the abnormality detection device 20 may also include a notification unit having a display capable of showing information or a speaker capable of providing audio notifications. If the abnormality detection unit 25 determines that the image of the input data shows a normal scene, it controls the notification unit to either not send a notification or to send a notification that the image of the input data shows a normal scene. On the other hand, if the abnormality detection unit 25 determines that the image of the input data shows an abnormal scene, it controls the notification unit to either send a notification or to send a notification that the image of the input data shows an abnormal scene.

[0059] [Explanation of the process] Figure 8 is a flowchart showing a typical example of the abnormality detection device 20's processing, and this flowchart explains the processing of the abnormality detection device 20. Details of each process are as described above, so their explanation is omitted here.

[0060] First, the preprocessing unit 21 of the abnormality detection device 20 acquires input data and performs preprocessing on that input data (step S21: preprocessing step). The data compression unit 22 compresses the preprocessed input data to generate compressed data (step S22: compressed data generation step).

[0061] Next, the compactification unit 23 maps a portion of the compressed data onto a compact manifold (step S23: mapping step). The reconstruction unit 24 uses the compressed data to generate reconstructed data in which the input data has been reconstructed (step S24: reconstruction step).

[0062] The processes described in steps S21-S24 are performed in common during both the learning phase and the inference phase. In the inference phase, the process described in step S25 is also performed.

[0063] The abnormality detection unit 25 determines whether or not there are abnormalities in the image of the input data by comparing the input data that has been preprocessed in step S21 with the reconstructed data generated in step S24 (step S25: reconstruction step).

[0064] [Explanation of effects] In recent years, there has been a growing demand for systems that can automatically identify abnormal events captured in images taken by surveillance cameras and other devices. Specific examples of abnormal events include vehicles entering pedestrian-only areas, objects falling onto highways, and incorrect work being carried out in manufacturing facilities.

[0065] When a system uses an AI model to automatically determine whether an event captured in an image is normal or abnormal, it is necessary to input a large number of images containing normal or abnormal events into the AI ​​model during training. However, if various types of abnormal events are conceivable, it becomes difficult to comprehensively prepare data of images containing abnormal events. For example, in the specific example of abnormal events mentioned above, it is not only cars that may enter a pedestrian zone, but also other types of vehicles such as bicycles and objects such as skateboards. Also, when an object falls onto a highway, various types of objects may fall. Thus, it is considered that there are practical difficulties in training an AI model to determine abnormal events.

[0066] For the reasons stated above, it is practically preferable to input a large number of images containing normal phenomena into the AI ​​model during training. When an image to be judged is input to this trained AI model, if the AI ​​model can accurately reproduce the image, it is assumed that the image contains normal phenomena. If the AI ​​model cannot accurately reproduce the image, it is assumed that the image contains abnormal phenomena.

[0067] However, such trained AI models may be able to accurately reproduce images even if abnormal events are captured in the image being judged. In detail, when the AI ​​model compresses the image data, it is thought that in the latent space of the compressed data, latent variables representing abnormal events exist at positions separated from the latent variables representing normal events. However, when this data representing abnormal events is reconstructed, the original image being judged may be reproduced with high accuracy (for example, with a small reconstruction error). For example, if the image being judged captures an event where a car has entered a pedestrian zone, the AI ​​model may be able to accurately reproduce the car in the reconstructed image. This can also be described as the AI ​​model having excessive expressive power.

[0068] The problem of AI models having excessive expressive power arose because the latent space is infinitely vast, and it was difficult for the AI ​​model to control the behavior of abnormal event data in areas where data of normal events did not appear. Because the AI ​​model reproduced abnormal events with high accuracy, the difference between the original image and the reconstructed image became small, potentially leading to missed detections of abnormal events.

[0069] In contrast, the anomaly detection device 20 includes a compactification unit 23 as a block to suppress the AI's excessive expressive capabilities. The compactification unit 23 maps a portion of the compressed data onto a compact manifold, thereby converting points with extreme values ​​in the latent space into points on the compact manifold. That is, the compactification unit 23 suppresses the existence of points representing data for abnormal events in the latent space in regions separated from regions where points representing data for normal events exist. As a result, the reconstructed image suppresses the reproduction of abnormal events that were present in the image being judged. This increases the difference between the image being judged and the reconstructed image. By comparing these two images, abnormal events can be detected.

[0070] Figures 9A and 9B show the image abnormality score when the target video is judged using the related technology and when the target video is judged using the abnormality judgment device 20, respectively. In Figures 9A and 9B, the horizontal axis represents the time in the video, and the vertical axis represents the abnormality score, which is the reconstruction error. Section A1 in Figure 9A and section A2 in Figure 9B represent the sections in the video where abnormal events occurred, respectively.

[0071] In the related technology, the anomaly detection device 20 is not provided with a compaction unit 23, and learning is performed on the anomaly detection device 20 by executing the unsupervised learning shown in Embodiment 2. Figure 9A shows the inference results using the anomaly detection device 20. Here, in Figure 9A, even if the anomaly score of interval A1 is compared with the anomaly score of intervals other than interval A1, no significant difference can be found between the two. Therefore, if the threshold for reconstruction error to determine whether something is normal or abnormal is increased, all scenes in the video will be judged as normal, while if the threshold for reconstruction error is decreased, not only abnormal scenes but also scenes that are actually normal will be judged as abnormal.

[0072] In contrast, Figure 9B shows that when comparing the anomaly score of interval A2 with that of intervals other than A2, the anomaly score of interval A2 tends to be higher than that of intervals other than A2. Therefore, by setting the reconstruction error threshold to an appropriate value, it is possible to distinguish and determine between normal scenes and abnormal scenes. In the example in Figure 9B, by setting the reconstruction error threshold to around 0.294, the situation in which normal scenes are judged as abnormal is suppressed, and it becomes possible to judge abnormal scenes as abnormal.

[0073] Furthermore, the abnormality detection device 20 may include an abnormality detection unit 25 that determines abnormalities in the input data by comparing the pre-processed input data with the reconstructed data generated by the reconstruction unit 24. This allows the abnormality detection device 20 to determine abnormalities in the input data on its own.

[0074] Furthermore, the compactification unit 23 may map only a portion, not all, of the compressed data onto a compact manifold. This allows the anomaly detection device 20 to further suppress the occurrence of abnormal events in the reconstructed data.

[0075] Furthermore, the compactification unit 23 may use a manifold homeomorphic to a circle as the compact manifold. As another example, the compactification unit 23 may use a manifold homeomorphic to a closed disk as the compact manifold. This allows the anomaly detection device 20 to perform characteristic processing using a relatively simple definition, thus facilitating user configuration of the anomaly detection device 20.

[0076] In the embodiments described above, this disclosure has been described as a hardware configuration, but this disclosure is not limited thereto. This disclosure can also be realized by having a processor in a computer execute a computer program to perform the processing of each device in the inspection management device or inspection management system described in the embodiments described above.

[0077] Figure 10 is a block diagram showing an example of the hardware configuration of an information processing device in which the processing of each embodiment described above is performed. Referring to Figure 10, this information processing device 90 includes a signal processing circuit 91, a processor 92, and a memory 93.

[0078] The signal processing circuit 91 is a circuit for processing signals in accordance with the control of the processor 92. The signal processing circuit 91 may also include a communication circuit that performs at least one of the following: receiving a signal from a transmitting device or transmitting a signal to a receiving device.

[0079] The processor 92 is connected to the memory 93 and performs the processing of the device described in the above embodiment by reading and executing a computer program from the memory 93. As an example of the processor 92, one of the following may be used: CPU (Central Processing Unit), MPU (Micro Processing Unit), FPGA (Field-Programmable Gate Array), DSP (Demand-Side Platform), or ASIC (Application Specific Integrated Circuit), or multiple of these may be used in parallel.

[0080] Memory 93 is composed of volatile memory, non-volatile memory, or a combination thereof. Memory 93 is not limited to one unit, but may be provided in multiple units. Volatile memory may be RAM (Random Access Memory) such as DRAM (Dynamic Random Access Memory) or SRAM (Static Random Access Memory). Non-volatile memory may be ROM (Read Only Memory) such as PROM (Programmable Random Only Memory) or EPROM (Erasable Programmable Read Only Memory), flash memory, or SSD (Solid State Drive).

[0081] Memory 93 is used to store one or more instructions. Here, one or more instructions are stored in memory 93 as a program. The processor 92 can perform the processing described in the above embodiment by reading and executing these programs from memory 93.

[0082] Furthermore, the memory 93 may include not only memory located outside the processor 92, but also memory built into the processor 92. The memory 93 may also include storage located separately from the processors that make up the processor 92. In this case, the processor 92 can access the memory 93 via an I / O (Input / Output) interface.

[0083] As described above, the one or more processors in each of the above embodiments execute one or more programs that include a set of instructions for causing a computer to perform the algorithm described with reference to the drawings. This process enables the information processing described in each embodiment.

[0084] The program, when loaded into a computer, includes a set of instructions or software code for causing the computer to perform one or more of the functions described in the embodiments. The program may be stored on a non-temporary computer-readable medium or a physical storage medium. Examples, but not limited to, include random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technologies, CD-ROM, digital versatile disk (DVD), Blu-ray® disc or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices. The program may be transmitted over a temporary computer-readable medium or a communication medium. Examples, but not limited to, include temporary computer-readable medium or a communication medium that includes electrical, optical, acoustic or other forms of propagating signals.

[0085] Some or all of the embodiments described above may also be described as follows, but are not limited to. Furthermore, some or all of the elements (e.g., configuration and function) described in Appendices 2 to 7 that are dependent on Appendice 1 may also be dependent on Appendices 8 and 9 in the same way as in Appendices 2 to 7. Thus, some or all of the elements described in any appendice may be applied to various hardware, software, recording means, systems, and methods for recording software. (Note 1) A data compression unit generates compressed data by reducing the dimensionality of the input data, A compactification unit that maps at least a portion of the compressed data onto a compact manifold, The system includes a reconstruction unit that generates reconstructed data from which the input data has been reconstructed using the compressed data, which includes the data mapped onto the compact manifold. Data processing device. (Note 2) The system further comprises a determination unit that determines an abnormality in the input data by comparing the input data with the reconstructed data. The data processing device described in Appendix 1. (Note 3) The determination unit calculates the difference between the pixel value of a predetermined pixel in the input data image and the pixel value of the pixel corresponding to the predetermined pixel in the reconstructed data image, compares the sum of the differences obtained by adding them to each pixel in all or part of the input data image with a predetermined threshold, and determines that there is an abnormality in the input data image if the sum is greater than or equal to the predetermined threshold. The data processing device described in Appendix 2. (Note 4) The system further includes a notification unit that performs notification according to whether or not there is an abnormality in the input data. A data processing device as described in Appendix 2 or 3. (Note 5) The compactification unit maps a portion of the compressed data onto the compact manifold. A data processing device as described in any one of the items 1 to 4 of the appendix. (Note 6) The compactification unit uses a manifold homeomorphic to the circumference as the compact manifold. A data processing device as described in any one of the items 1 to 5 of the appendix. (Note 7) The compactification section uses a manifold that is homeomorphic to a closed disk as the compact manifold. A data processing device as described in any one of the items 1 to 5 of the appendix. (Note 8) By reducing the dimensionality of the input data, compressed data is generated. At least a portion of the compressed data is mapped onto a compact manifold, Using the compressed data which includes the data mapped onto the compact manifold, the input data is reconstructed to generate reconstructed data. The data processing methods performed by computers. (Note 9) By reducing the dimensionality of the input data, compressed data is generated. At least a portion of the compressed data is mapped onto a compact manifold, Using the compressed data which includes the data mapped onto the compact manifold, the input data is reconstructed to generate reconstructed data. A program that causes a computer to perform a task.

[0086] Each drawing referenced in the embodiments described above is merely illustrative to illustrate one or more embodiments. Each drawing may be associated with one or more other embodiments rather than with only one specific embodiment. As those skilled in the art will understand, various features or steps described with reference to any one drawing can be combined with features or steps shown in one or more other drawings, for example, to create embodiments not explicitly shown or described. Not all features or steps shown in any one drawing to illustrate an exemplary embodiment are necessarily required, and some features or steps may be omitted. The order of steps shown in any of the drawings may be changed as appropriate.

[0087] Although the present disclosure has been described above, it is not limited to the embodiments described above. Various modifications to the structure and details of the present disclosure are possible, as can be understood by those skilled in the art within the scope of the disclosure. Furthermore, each embodiment can be combined with other embodiments as appropriate.

[0088] This application claims priority based on Japanese Patent Application No. 2023-031025, filed on 1 March 2023, and incorporates all of its disclosures herein. [Explanation of Symbols]

[0089] 10 Data Processing Devices 11 Data compression section 12 Compaction section 13 Reconstruction part 20 Abnormality determination device 21 Preprocessing unit 22 Data compression unit 23 Compactification section 24 Reconfiguration section 25 Abnormality determination section

Claims

1. A data compression means that generates compressed data by reducing the dimensionality of the input data, A compactification means for mapping a portion of the compressed data onto a compact manifold, The system comprises reconstruction means for generating reconstructed data in which the input data is reconstructed using the compressed data which includes the data mapped onto the compact manifold. Data processing device.

2. The system further comprises a determination means for determining an anomaly in the input data by comparing the input data with the reconstructed data. The data processing device according to claim 1.

3. The determination means calculates the difference between the pixel value of a predetermined pixel in the input data image and the pixel value of the pixel corresponding to the predetermined pixel in the reconstructed data image, compares the sum of the differences obtained by adding them to each pixel in all or part of the input data image with a predetermined threshold, and determines that there is an abnormality in the input data image if the sum is greater than or equal to the predetermined threshold. The data processing device according to claim 2.

4. The system further comprises notification means that performs notification according to whether or not there is an abnormality in the input data. The data processing apparatus according to claim 2 or 3.

5. The compactification means uses a manifold homeomorphic to the circumference as the compact manifold. The data processing apparatus according to any one of claims 1 to 3.

6. The compactification means uses a manifold homeomorphic to a closed disk as the compact manifold. The data processing apparatus according to any one of claims 1 to 3.

7. By reducing the dimensionality of the input data, compressed data is generated. A portion of the compressed data is mapped onto a compact manifold, Using the compressed data which includes the data mapped onto the compact manifold, the input data is reconstructed to generate reconstructed data. The data processing methods performed by computers.

8. By reducing the dimensionality of the input data, compressed data is generated. A portion of the compressed data is mapped onto a compact manifold, Using the compressed data which includes the data mapped onto the compact manifold, the input data is reconstructed to generate reconstructed data. A program that causes a computer to perform a task.

Images