System and method with diffusion-based outlier synthesis for anomaly detection

Abstract

A computer-implemented system and method relate to anomaly detection. A first source image is obtained from a first set of images of the in-distribution image dataset and a second source image is obtained from a second set of images of the in-distribution image dataset. A diffusion model generates a modified image using the first source image and the second source image. A non-anomalous tag is automatically generated for the first source image. A non-anomalous tag is also generated for the second source image. An exception tag is generated for the modified image. A training data set is created. The training dataset includes at least a first source image having a non-anomalous tag, a second source image having a non-anomalous tag, and a modified image having an anomalous tag. A machine learning model is trained or fine-tuned using the training data set. The machine learning model is configured to perform a task of anomaly detection.

Description

Technical Field

[0001] The present disclosure relates generally to computer vision, and more particularly to digital image processing and anomaly detection. Background Art

[0002] In general, anomaly detection involves identifying samples that systematically deviate from the distribution of normal samples. Anomaly detection is valuable in multiple fields such as financial security, network security, healthcare applications, etc. Since anomalies are rare and may have critical consequences, there are multiple challenges to collecting anomaly examples and learning their distribution.

[0003] For example, the lack of sufficient labeled anomalies poses challenges in both anomaly detection model training and evaluation. Due to the lost training signal from labeled anomalies, the anomaly detection model may have compromised performance in detecting key anomalies. Uncovering the true performance of anomaly detection models requires a test set consisting of labeled normal data and labeled anomaly data. The absence of sufficient abnormal data in the test set may lead to biased estimates of the model's performance. Moreover, the evaluation of anomaly detection models without access to abnormal samples is underdeveloped. In addition, the evaluation of anomaly detection models may be a bottleneck in the development of new advanced anomaly detection models. Summary of the invention

[0004] The following is an overview of certain embodiments described in detail below. The described aspects are presented only to provide the reader with a brief overview of these specific embodiments, and the description of these aspects is not intended to limit the scope of the present disclosure. In fact, the present disclosure may cover a variety of aspects that may not be explicitly described below.

[0005] According to example aspects, a computer-implemented method relates to anomaly detection. The method includes: receiving a first source image from a first image set of a distribution image data set and receiving a second source image from a second image set of the distribution image data set. The method includes: generating a modified image using the first source image and the second source image via a diffusion model. The method includes: automatically generating a non-anomalous label for the first source image and a non-anomalous label for the second source image. The method includes: automatically generating an anomaly label for the modified image. The method includes: creating a training data set, which includes at least a first source image with a non-anomalous label, a second source image with a non-anomalous label, and a modified image with an anomaly label. The method includes: using the training data set to train or fine-tune a machine learning model. The machine learning model is configured to perform the task of anomaly detection.

[0006] According to an example aspect, a system includes one or more processors and at least one non-transitory computer-readable medium in data communication with the one or more processors. The at least one non-transitory computer-readable medium has computer-readable data including instructions stored thereon, and the instructions are configured to cause the one or more processors to perform a method when executed by the one or more processors. The method includes: receiving a first source image from a first image set of a distributed image data set and receiving a second source image from a second image set of the distributed image data set. The method includes: generating a modified image using the first source image and the second source image via a diffusion model. The method includes: automatically generating a non-anomalous label for the first source image and a non-anomalous label for the second source image. The method includes: automatically generating an anomaly label for the modified image. The method includes: creating a training data set, which includes at least a first source image with a non-anomalous label, a second source image with a non-anomalous label, and a modified image with an anomaly label. The method includes: using the training data set to train or fine-tune a machine learning model. The machine learning model is configured to perform a task of anomaly detection.

[0007] According to an example aspect, a non-transitory computer-readable medium has computer-readable data including instructions stored thereon. The computer-readable data is executable by one or more processors to perform a method. The method includes: receiving a first source image from a first image set of a distributed image data set and receiving a second source image from a second image set of the distributed image data set. The method includes: generating a modified image using the first source image and the second source image via a diffusion model. The method includes: automatically generating a non-anomalous label for the first source image and a non-anomalous label for the second source image. The method includes: automatically generating an anomaly label for the modified image. The method includes: creating a training data set, which includes at least a first source image with a non-anomalous label, a second source image with a non-anomalous label, and a modified image with an anomaly label. The method includes: using the training data set to train or fine-tune a machine learning model. The machine learning model is configured to perform a task of anomaly detection.

[0008] These and other features, aspects and advantages of the present invention are discussed in the following detailed description with reference to the accompanying drawings, wherein like characters indicate like or similar parts throughout the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 is a flow chart of an example of a system with diffusion-based outlier synthesis for anomaly detection according to an example embodiment of the present disclosure.

[0010] Figure 2is a diagram of an example of a system having an anomaly detection system with diffusion-based outlier synthesis according to an example embodiment of the present disclosure.

[0011] Figure 3 is a diagram of an example of a diffusion-based process with a text guide according to an example embodiment of the present disclosure.

[0012] Figure 4 is a diagram of an example of a sequence of synthesized out-of-distribution images with increased editing strength according to an example embodiment of the present disclosure.

[0013] Figure 5 is a diagram of another example of a sequence of synthesized out-of-distribution images with increased editing strength according to an example embodiment of the present disclosure.

[0014] Figure 6 is a diagram of an example of a diffusion-based process with an image wizard according to an example embodiment of the present disclosure.

[0015] Figure 7 is a diagram of another example of a set of synthesized out-of-distribution images according to an example embodiment of the present disclosure.

[0016] Figure 8 is a diagram of an example of a control system including at least a trained anomaly detection model and / or an anomaly detection system according to an example embodiment of the present disclosure.

[0017] Fig. 9 According to an exemplary embodiment of the present disclosure, the present invention relates to mobile machine technology. Figure 8 Diagram of the control system.

[0018] Fig.10 According to the exemplary embodiment of the present disclosure, the manufacturing technology Figure 8 Diagram of the control system.

[0019] Fig.11 According to an exemplary embodiment of the present disclosure, the present invention relates to a security technology Figure 8 Diagram of the control system.

[0020] Fig.12 The present invention relates to imaging technology according to an exemplary embodiment of the present disclosure. Figure 8 Diagram of the control system. DETAILED DESCRIPTION

[0021] The embodiments described herein, which have been shown and described as examples, and their many advantages will be understood from the above description, and it should be apparent that various changes may be made in the form, construction, and arrangement of components without departing from the disclosed subject matter or without sacrificing one or more of its advantages. Indeed, the described forms of these embodiments are merely illustrative. These embodiments are susceptible to various modifications and alternative forms, and the appended claims are intended to cover and include such changes, and are not limited to the specific forms disclosed, but rather cover all modifications, equivalents, and alternatives that fall within the spirit and scope of the present disclosure.

[0022] Anomaly detection often involves the identification of outliers. However, this task is often performed assuming that only a small unlabeled set of inlier data is available. In addition, in order to tune and validate these anomaly detection models, a representative set of validation data is often required, whereby the validation data includes labeled normal samples and labeled anomaly samples. When anomaly detection is deployed in new applications and the distribution of outliers is still unknown, such a data set with validation data may be difficult to obtain. The AD system 100 solves this technical problem by providing a technical solution related to the generation of these synthetic outliers and the training, fine-tuning and / or validation of anomaly detection models using at least synthetic outliers.

[0023] Figure 1 is a flow chart of an example of an anomaly detection (AD) system 100 with diffusion-based outlier synthesis for anomaly detection according to an example embodiment. The AD system 100 utilizes diffusion-based outlier synthesis for near anomaly detection use cases. The AD system 100 uses diffusion-based generation techniques for outlier exposure 116. The AD system 100 is built on diffusion-based generation with an asymmetric reverse process (Asyrp). Asyrp refers to the process of discovering a semantic latent space ("h-space") in a frozen pre-trained diffusion model. In addition, the AD system 100 is built on diffusion-based generation with DiffStyle. Diffstyle refers to blending intermediate features of a content image in h-space into an inverted style image (i.e., x T ) during the generation process.

[0024] In addition, if Figure 1 As shown in , the AD system 100 generates synthetic outliers (e.g., generated / synthesized / modified images) to use them for training anomaly detection models 114 and / or evaluating anomaly detection models 114. The AD system 100 is also configured to generate synthetic anomalies for near anomaly detection use cases (such as outlier exposure 116, model validation 118, etc.).

[0025] Furthermore, the AD system 100 is configured to operate in a plurality of use cases that are associated with outlier detection settings. Figure 1 As shown in , the AD system 100 is configured to create and use a data set 112 having synthetic anomalies (or synthetic outliers) with respect to at least one anomaly detection model 114 as a validation data set for model validation 118. The AD system 100 is configured to use synthetic anomalies in domains where such validation data sets do not exist. As another example, the AD system 100 is configured to use synthetic anomalies (or synthetic outliers) of at least one anomaly detection model 114 as a training data set for outlier exposure 116.

[0026] As discussed above, the AD system 100 is configured to use the dataset 112 for the anomaly detection model 114 for validation 118. The dataset 112 includes synthetic anomalies (e.g., generated / modified / synthesized images) together with their corresponding source images of the benign intra-class dataset 102. In addition, the AD system 100 is configured to generate normal labels or non-anomalous labels for the source images of the benign intra-class dataset 102. The AD system 100 is configured to generate an anomaly label for each of the generated / modified / synthesized images. The AD system 100 can be configured to generate an anomaly label indicating the degree of modification to the source image.

[0027] The AD system 100 assumes that a small set of benign in-class examples from the benign in-class dataset 102 is available for training, but no validation set exists. The AD system 100 then executes a diffusion-based outlier generation method 110 via at least one pre-trained diffusion model 104 together with an image / text wizard 106 to create a dataset 112. The dataset 112 can be used in the same / similar manner as the real validation dataset will be used to select model hyperparameters, thereby evaluating between candidate model architectures and estimating performance on a test dataset. In order for the synthetic validation set to be effective, then: (i) the design choices made using the synthetic validation set should be similar to the design choices made using the real validation set; and (ii) evaluation metrics such as the area under the receiver operating characteristic curve (AUROC) should be similar between the synthetic and real validation sets.

[0028] The AD system 100 may include an experimental setup in which a synthetic validation set is evaluated. In this experimental setup, the AD system 100 uses (i) a battery of anomaly detectors (or anomaly detection models 114) of varying quality and (ii) a selection of anomaly detection tasks with a labeled validation set to create a true benchmark test of anomaly detectors ("real validation"). The AD system 100 then benchmarks the anomaly detection model 114 on a test set with synthetic anomalies (e.g., generated / synthesized / modified images). These are then denoted as "synthetic validation" or "synthetic benchmarks". Empirical studies have found that "synthetic benchmarks" closely reflect "real validation" in terms of both relative rankings of anomaly detection methods and absolute AUROC, thereby encouraging the use of synthetic anomalies for validation of anomaly detectors in newly deployed environments.

[0029] In addition to using synthetic outliers for validation 118, the AD system 100 is also configured to use synthetic outliers (e.g., generated / modified / synthesized images) to improve the performance of the anomaly detection model 114 through outlier exposure 116. For outlier exposure 116, the AD system 100 is configured to use a pre-trained visual transformer model. The AD system 100 is configured to fine-tune the visual transformer model on a proxy classification task and use the distance in the trained embedding space as anomaly detection.

[0030] The proxy classification task for fine-tuning involves a binary classification layer added to the visual transformer model, and the anomaly detection model 114 is trained on benign in-class examples and synthetic outlier examples. In addition to the proxy classification task, the AD system 100 also includes a regression-based task. The AD system 100 uses a varying editing strength α to generate a variety of synthetic outliers with text guides.

[0031] When fine-tuning the anomaly detection model 114, the AD system 100 is configured to use a regression task to predict α. After fine-tuning, the AD system 100 removes the prediction head of the visual transformer model. The data set 112 is then converted into the embedding space of the visual transformer model (i.e., the last layer before the prediction layer) and used as a feature library for anomaly detection. At test time, the L2 distance to the closest k examples in the feature library is used as the anomaly score.

[0032] Figure 2is a diagram of an example of a system 200 having an AD system 100 with diffusion-based outlier synthesis according to an example embodiment. The system 200 includes at least a processing system 210 having at least one processing device. For example, the processing system 210 includes at least an electronic processor, a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), a microprocessor, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), any suitable processing technology, or any number and combination of the above. The processing system 210 is operable to provide functionality as described herein.

[0033] The system 200 includes a memory system 220, which is operatively connected to the processing system 210. In an example embodiment, the memory system 220 includes: at least one non-transitory computer-readable storage medium, which is configured to store various data and provide access to various data to enable at least the processing system 210 to perform operations and functions as disclosed herein. In an example embodiment, the memory system 220 includes a single memory device or multiple memory devices. The memory system 220 may include electrical, electronic, magnetic, optical, semiconductor, electromagnetic or any suitable storage technology that can operate with the AD system 100. For example, in an example embodiment, the memory system 220 may include a random access memory (RAM), a read-only memory (ROM), a flash memory, a disk drive, a memory card, an optical storage device, a magnetic storage device, a memory module, any suitable type of memory device, or any number and combination of the above. With respect to the processing system 210 and / or other components of the AD system 100, the memory system 220 is local, remote, or a combination thereof (e.g., partially local and partially remote). For example, the memory system 220 may include at least a cloud-based storage system (eg, a cloud-based database system) that is remote from the processing system 210 and / or other components of the AD system 100 .

[0034] The memory system 220 includes at least the AD system 100, machine learning data 230, and other related data 240 stored thereon. More specifically, the AD system 100 includes computer-readable data having instructions that, when executed by the processing system 210, are configured to provide at least diffusion-based and image-based / text-based outlier synthesis for anomaly detection as described in the present disclosure. The computer-readable data may include instructions, codes, routines, various related data, any software technology, or any number and combination of the above. The AD system 100 includes a machine learning model that includes at least a pre-trained diffusion model 104, a pre-trained visual language model 108, and an anomaly detection model 114. The pre-trained diffusion model 104 is configured to perform at least a reverse diffusion process. For example, the pre-trained diffusion model 104 includes at least one denoised diffusion implicit model (DDIM). The pre-trained diffusion model 104 includes at least an encoder (e.g., Figure 3 The encoder 304 or Figure 6 The pre-trained diffusion model 104 includes at least a decoder (e.g., Figure 3 The decoder 310 or Figure 6 The decoder 616 of the image decoding network. Moreover, as an example, the pre-trained visual language model 108 includes: a contrastive language image pre-training (CLIP), which includes at least an image encoder (denoted as E I ) and the text encoder (denoted as E T ). The anomaly detection model 114 is a machine learning model that is configured to perform anomaly detection. The anomaly detection model 114 is configured to generate prediction data indicating anomaly detection or non-anomaly detection in response to an input, such as sensor data (e.g., a digital image). For example, the anomaly detection model 114 may include a classifier (e.g., a binary classifier with classes of anomaly and non-anomaly) or a regression model. The anomaly detection model 114 may include an autoencoder-based anomaly detection model, a self-supervised anomaly detection model, a kernel-based anomaly detection model, any other anomaly detection model, or any number and combination of the above.

[0035] Moreover, the machine learning data 230 includes sensor data, source image data, various loss data, various weight data, various parameter data, and any related data that enables the AD system 100 to learn and / or perform functions as described in the present disclosure. The machine learning data 230 includes the benign in-class data set 102, the data set 112, synthetic anomalies (e.g., generated / modified / synthesized images), and various data related to outlier exposure 116 and / or model validation 118. Meanwhile, other related data 240 provides various data that enables the AD system 100 to perform functions as discussed herein (e.g., operating system, etc.).

[0036] The system 200 may include at least one sensor system 250. The sensor system 250 includes one or more sensors. For example, the sensor system 250 includes an image sensor, a camera, a radar sensor, a light detection and ranging (LIDAR) sensor, a thermal sensor, an ultrasonic sensor, an infrared sensor, a motion sensor, an audio sensor, an inertial measurement unit (IMU), any suitable sensor, or any number and combination of the above. The sensor system 250 is operable to communicate with one or more other components of the system 200 (e.g., the processing system 210 and the memory system 220). For example, the sensor system 250 may provide sensor data, which is then used by the processing system 210 to generate image data (e.g., a digital image) based on the sensor data. In this regard, the processing system 210 is configured to obtain sensor data directly or indirectly from one or more sensors of the sensor system 250. The sensor system 250 is local, remote, or a combination thereof (e.g., partially local and partially remote). Upon receiving sensor data, the processing system 210 is configured to process the sensor data and / or generate image data (e.g., a digital image) in conjunction with the AD system 100, machine learning data 230, other relevant data 240, any suitable software and / or hardware components, or any number and combination of the above.

[0037] In addition, the system 200 may include at least one other component. Figure 2As shown in , the memory system 220 is also configured to store other relevant data 240, which is related to the operation of one or more components (e.g., sensor system 250, I / O device 260 and other functional modules 270) of the AD system 100. In addition, the system 200 is configured to include: one or more I / O devices 260 (e.g., display device, keyboard device, speaker device, etc.), which are related to the AD system 100. Moreover, the system 200 includes other functional modules 270 that assist the operation of the AD system 100 or contribute to the operation of the AD system 100, such as any appropriate hardware, software or combination thereof. For example, the other functional modules 270 include communication technologies that enable the components of the system 200 to communicate with each other as described herein. At this point, the system 200 is operable to perform the functions of the AD system 100. The system 200 is operable to train, verify, test, adopt and / or deploy an anomaly detection model 114 with diffusion-based outlier synthesis for anomaly detection, as described herein.

[0038] Figure 3 is a flowchart associated with a process 300 for generating a modified image (or generated image xgen) according to an example embodiment. The AD system 100 uses some components of the Asyrp process to create a text-guided generation, but the AD system 100 differs from the original Asyrp process in at least two ways. First, the AD system 100 uses a non-domain-specific text guide. Second, the AD system 100 defines the editing strength of each exception. In contrast, the original Asyrp process is evaluated on a well-defined domain and assumes that the source and target text are known (e.g., modifying "face" to "smiley face"). In contrast, the AD system 100 provides a more general solution that does not assume a specific domain and does not require domain-specific text as input.

[0039] First, the AD system 100 is configured to extract a meaningful direction for ΔT, so that the source text does not need to be used. The AD system 100 is configured to extract a meaningful direction for ΔT by using an image encoder E I and the source image x source To extract the meaningful direction for ΔT. Secondly, the AD system 100 uses the auxiliary out-of-domain text y aux , while Asyrp uses the target text. The AD system 100 defines the directional loss L′ dir Specifically, the AD system 100 calculates the directional loss data via equation 3: ΔT′=E T (y aux )-E l (x source ) [1] ΔI=E I (xtarget )-E l (x source ) [2]

[0040] The AD system 100 follows a training process for an h-space extractor for Asyrp, which is defined to extract Δh - the direction in h-space that induces the desired change. Once Δh is found, the AD system 100 applies Asyrp during the backdiffusion process, but uses the linear properties of h-space and defines an editing strength α, which indicates the level of how strongly the image is modified. Considering the backdiffusion process D that generates the image x from the latent code z as expressed by x=D(z), the AD system 100 adds the h-space term Δh to generate the generated / modified / synthesized image (i.e., xgen, such as x0). The h-space term Δh may be referred to herein as additional embedded data 308. The AD system 100 generates the modified image via Equation 4: x gen =D(z+aΔh) [4]

[0041] and, Figure 3 Multiple iterations of process 300 involving input data 302 and corresponding output data 314 are illustrated. More specifically, AD system 100 includes multiple iterations of process 300, wherein x The input data 302 from T to x1 are generated from x t Output data 314 to x0. In this example, x0 refers to the generated / modified / synthesized image, and x T refers to the latent code of the source image generated via the forward diffusion process of the diffusion model 104. Also, in this example, x T It can be a noisy image including Gaussian noise. In addition, Figure 3 The process 300 is shown to involve at least an encoder 304 of the diffusion model 104 and a decoder 310 of the diffusion model 104. In this regard, for example, the encoder 304 receives input data 302 and generates embedded data 306 by encoding the input data 302. As an example iteration, the AD system 100 is configured to: (i) obtain x T as input to encoder 304; (ii) via encoder 304, the input (e.g., x T ) is encoded based on x T(iii) generating updated embedded data 312 by combining embedded data 306 with additional embedded data 308 (Δh); and (iv) decoding updated embedded data 312 via decoder 310 to generate a generated / synthesized / modified image (e.g., x t As another example, in the next iteration, the AD system 100 is configured to: (i) obtain x t as input to encoder 304; (ii) via encoder 304, the input (e.g., x t ) is encoded based on x t (iii) generating updated embedded data 312 by combining embedded data 306 with additional embedded data 308; and (iv) decoding updated embedded data 312 via decoder 310 to generate a generated / synthesized / modified image (e.g., x t-1 ). The AD system 100 is configured to continue the process 300 for multiple iterations until the generated / modified / synthesized image (x0) is generated as output data 314. In this example, for example, the dataset 112 includes at least: (i) source images from the benign intra-class dataset 102 along with corresponding normal / non-anomalous labels; and (ii) each generated / modified / synthesized image (x0) along with a corresponding abnormal label.

[0042] Figure 4 is a diagram of an example of a sequence 400 of synthetic out-of-distribution images with increasing editing strength α∈[0,1] according to an example embodiment. More specifically, the sequence 400 includes a source image 402 followed by a set of synthetic out-of-distribution images. In this non-limiting example, the source image 402 is an image of a Tennessee warbler, which is a type of bird.

[0043] The set of synthetic out-of-distribution images includes multiple generated images. Each generated image xgen is represented as Figure 3 Each generated image includes a source image 402, which has been guided by Dolphin to assist out-of-domain text (i.e., y aux = dolphin). For example, the set of synthetic out-of-distribution images includes generated image 404 (i.e., x0 with the lowest α value from the set) and generated image 414 (i.e., x0 with the highest α value from the set). Also, as a non-limiting example, the set of synthetic out-of-distribution images includes a plurality of generated images that are between generated image 404 and generated image 414, such as generated image 406, generated image 408, generated image 410, generated image 412, and generated image 413. Figure 4, which exhibit increasing alpha values ​​in this order. In this regard, the generated image 404 (i.e., x0) has an alpha value lower than the generated image 406 (i.e., x0). The generated image 406 (i.e., x0) has an alpha value lower than the generated image 408 (i.e., x0). The generated image 408 (i.e., x0) has an alpha value lower than the generated image 410 (i.e., x0). The generated image 410 (i.e., x0) has an alpha value lower than the generated image 412 (i.e., x0). The generated image 412 (i.e., x0) has an alpha value lower than the generated image 414 (i.e., x0). The lower alpha represents a smaller degree of modification in the direction of the text guide (e.g., "Dolphin") relative to the source image (e.g., "Tennessee Warbler" image).

[0044] As described above, as an example, the AD system 100 is configured to create a dataset 112, which may include at least the source image 402 and each of the generated / synthesized / modified images of the sequence 400 (e.g., image 404, image 406, image 408, image 410, image 412, image 414, etc.). In addition, the AD system 100 is also configured to generate an abnormal label (or various classes of abnormal labels) for each of the generated images of the sequence 400. That is, the AD system 100 is configured to create a dataset 112, which includes: (i) the source image 402 together with a non-abnormal label for the source image 402; and (ii) each of the generated / synthesized / modified images of the sequence 400 together with a non-abnormal label for Figure 4 4. The data set includes a corresponding anomaly label for each of the generated images of the sequence 400 shown in FIG. This data set may be used when training, fine-tuning, or validating the anomaly detection model 114 for a particular application.

[0045] Figure 5 is a diagram of another example of a sequence 500 of synthetic out-of-distribution images with increasing editing strength α∈[0,1] according to an example embodiment. More specifically, the sequence 500 includes a source image 502 followed by a set of synthetic out-of-distribution images. In this non-limiting example, the source image 502 is an image of a gladiolus, which is a type of flower.

[0046] The set of synthetic out-of-distribution images includes multiple generated images. Each generated image x gen is represented as Figure 3 Each generated image includes a source image 502, which has been guided by Dolphin to assist out-of-domain text (i.e., y aux= dolphin). For example, the set of synthetic out-of-distribution images includes generated image 504 (i.e., x0 with the lowest α value from the set) and generated image 514 (i.e., x0 with the highest α value from the set). Also, as a non-limiting example, the set of synthetic out-of-distribution images includes a plurality of generated images that are between generated image 504 and generated image 514, such as generated image 506, generated image 508, generated image 510, generated image 512, and generated image 514. Figure 4 , which exhibit increasing alpha values ​​in this order. In this regard, generated image 504 (i.e., x0) has a lower alpha value than generated image 506 (i.e., x0). Generated image 506 (i.e., x0) has a lower alpha value than generated image 508 (i.e., x0). Generated image 508 (i.e., x0) has a lower alpha value than generated image 510 (i.e., x0). Generated image 510 (i.e., x0) has a lower alpha value than generated image 512 (i.e., x0). Generated image 512 (i.e., x0) has a lower alpha value than generated image 514 (i.e., x0). Lower alpha represents a smaller degree of modification in the direction of the text guide (e.g., "Dolphin") relative to the source image (e.g., "Gladiolus" image).

[0047] As described above, as an example, the AD system 100 is configured to create a dataset 112, which may include at least the source image 502 and each of the generated images of the sequence 500 (e.g., image 504, image 506, image 508, image 510, image 512, image 514, etc.). In addition, the AD system 100 is also configured to generate an abnormal label for each of the generated images of the sequence 500 (or various abnormal labels for different degrees of modification of the source image). That is, the AD system 100 is configured to create a dataset 112, which includes: (i) the source image 502 together with a non-abnormal label for the source image 502; and (ii) each of the generated images (xgen or x0) of the sequence 500 together with a non-abnormal label for Figure 5 500. The dataset may be used when training, fine-tuning, or validating the anomaly detection model 114 for a particular application.

[0048] Figure 6 6 is a diagram of an example of a process 600 including image-guided back diffusion according to an example embodiment. The AD system 100 uses a DiffStyle process to create an image-guided generation. In the DiffStyle process, a first image I(1) and a second image I(2) are transformed into a latent space by a forward diffusion process, thereby generating a first latent code and the second hidden code The first hidden code With the first h-spatial component h (1) . Second hidden code With a second h-spatial component h (2) .

[0049] Next, the AD system 100 generates a first h-spatial component h (1) and the second h-spatial component h (2) The AD system 100 performs linear interpolation using Equation 5 to generate h gen ,h gen May be referred to as "additional embedded data". h gen =(1-γ)h (1) +γh (2) [5]

[0050] AD system 100 uses the first hidden code To perform Asyrp, where h-space (i.e., the innermost component of the UNet) is replaced by h during process 600 gen Alternatively, process 600 includes a back diffusion process. AD system 100 is configured to transform the first image I (1) is considered as a “style image” and the second image I (2) is regarded as a “content image”. Moreover, in Equation 5, γ represents the relative strength of the content image during the style transfer of the style image. For example, the AD system 100 used γ=0.7γ as a default value in some experiments.

[0051] When generating synthetic outliers for anomaly detection, the AD system 100 assumes that no information about the outlier distribution is known. Therefore, the AD system 100 only performs DiffStyle with known elements, such as benign in-class examples with a benign in-class dataset 102 (e.g., a public dataset such as Tiny-ImageNet) for outlier exposure 116.

[0052] And, as an example, Figure 6 The process 600 is illustrated, involving: (i) multiple iterations of input data 602 and corresponding output data 618; and (ii) multiple iterations of input data 604 and corresponding output data 620. In addition, the AD system 100 includes multiple iterations of the process 600, in which: (i) from arrive The input data 602 are generated from arrive Output data 618 of , where x0 refers to the generated / modified / synthesized image for dataset 112 . Figure 6 The process 600 is shown to involve at least an encoder 606 of the diffusion model 104 and a decoder 310 of the diffusion model 104. The encoder 606 receives input data 602 and generates embedded data 608 by encoding the input data 602. The encoder 606 also receives input data 604 and generates embedded data 610 by encoding the input data 604. As an example iteration, the AD system 100 is configured to: (a) via the encoder 606, Encoding is performed to generate embedded data 608 (ie, h (1) ); (b) via encoder 606, by Encoding is performed to generate embedded data 610 (ie, h (2) ); (c) by embedding data 608 (i.e., h (1) ) and embedded data 610 (ie, h (2) ) are combined to generate additional embedded data 612 (ie, h gen ); (d) by embedding the data 608 (ie, h (1) ) and additional embedded data 612 (ie, h gen ) are combined to generate updated embedded data 614; (e) by decoding the updated embedded data 614 via a decoder 616 to generate As output data 618, represents a generated / synthesized / modified version of the image; and (f) by decoding the embedded data 610 (ie, h (2) ) is decoded to generate As output data 620, represents a generated / synthesized / modified version of the image. As another example, for example, in the next iteration, the AD system 100 is configured to: (a) via the encoder 606, by Encoding is performed to generate embedded data 608 (ie, h (1) ); (b) via encoder 606, by Encoding is performed to generate embedded data 610 (ie, h (2) ); (c) by embedding data 608 (i.e., h (1) ) and embedded data 610 (ie, h (2) ) are combined to generate additional embedded data 612 (ie, h gen ); (d) by embedding the data 608 (ie, h (1) ) and additional embedded data 612 (ie, h gen) are combined to generate updated embedded data 614; (e) by decoding the updated embedded data 614 via a decoder 616 to generate As output data 618, represents a generated / synthesized / modified version of the image; and (f) by decoding the embedded data 610 (ie, h (2) ) is decoded to generate As output data 620, Represents a generated / synthesized / modified version of the image. The AD system 100 is configured to continue the process 300 for multiple iterations until a generated / modified / synthesized image (x0) is generated as output data 314. In this example, for example, the dataset 112 includes at least: (i) source images from the benign intra-class dataset 102 along with corresponding normal / non-anomalous labels; and (ii) each generated / modified / synthesized image (x0) along with a corresponding abnormal label.

[0053] Figure 7 7 is a diagram of another example of a data set 700 including source images and synthetic out-of-distribution images according to an example embodiment. More specifically, the data set 700 includes a plurality of source images including a set 702 of “first images” (e.g., “first source images”) shown in a first row and a set 704 of “second images” (e.g., “second source images”) shown in a first column. Figure 7 In , the set of first source images 702 includes a first source image 706, a first source image 708, a first source image 710, a first source image 712, and a first source image 714. Figure 7 , the set 704 of second source images includes a second source image 716, a second source image 718, a second source image 720, a second source image 722, and a second source image 724. Each first image represents a style image I (1) At the same time, each second image represents the content image I (2) The generated images are shown in the region between the first row and the first column, so that the set of generated images includes columns 2 to 6 as defined by rows 2 to 6. In this non-limiting example, for example, there are Figure 7 Each generated image (i.e., x0) is based on Figure 6 The process 600 shown in FIG. 600, using the image wizard, via h gen , as expressed in Equation 5. In this regard, as a non-limiting example, the generated image 726 (ie, x0) is generated using the first source image 708 and the second source image 718 via Figure 6Also, as another non-limiting example, the generated image 728 (ie, x0) is generated using the first source image 712 and the second source image 724 via Figure 6 The image is generated by process 600 with an image wizard. Figure 7 The other generated images in are created from their corresponding first source images from the set 702 arranged in the same column and their corresponding second images from the set 704 arranged in the same row. As mentioned previously, the combined set 700 (e.g., Figure 7 All source images and all generated images shown in ) are examples of training data that may be used when training anomaly detection model 114 for a given application.

[0054] Figure 8 According to the exemplary embodiment, Figure 1-2 800 of an example of an AD system 100 and / or a trained anomaly detection model 114. Figure 8 800, which includes the AD system 100. In this example, the system 800 includes at least a sensor system 810, a control system 820, and an actuator system 830. The system 800 is configured so that the control system 820 controls the actuator system 830 based on the sensor data from the sensor system 810. More specifically, the sensor system 810 includes one or more sensors and / or corresponding devices to generate sensor data. For example, the sensor system 810 includes an image sensor, a camera, a radar sensor, a light detection and ranging (LIDAR) sensor, a thermal sensor, an ultrasonic sensor, an infrared sensor, a motion sensor, a satellite-based navigation sensor (e.g., a global positioning system (GPS) sensor), an optical sensor, an audio sensor, any suitable sensor, or any number and combination of the above. When obtaining detection from the environment, the sensor system 810 is operable to communicate with the control system 820 via an input / output (I / O) system 870 and / or other functional modules 850, which includes communication technology.

[0055] The control system 820 is configured to obtain sensor data directly or indirectly from one or more sensors of the sensor system 810. In this regard, the sensor data may include sensor data from a single sensor or sensor fusion data from multiple sensors. Upon receiving an input including at least sensor data, the control system 820 is operable to process the sensor data via a processing system 840. In this regard, the processing system 840 includes at least one processor. For example, the processing system 840 includes an electronic processor, a CPU, a GPU, a TPU, a microprocessor, an FPGA, an ASIC, a processing circuit, any suitable processing technology, or any number and combination of the above. When processing at least the sensor data, the processing system 840 is configured to extract, generate and / or obtain appropriate input data (e.g., digital image data) for the AD system 100. In addition, the processing system 840 is operable to generate output data (e.g., predictive data indicating abnormal sensor data or non-abnormal / normal sensor data) via the AD system 100 based on communication with the memory system 860. Additionally, the processing system 840 may be operable to provide actuator control data to the actuator system 830 based on the output data (eg, prediction data from a trained anomaly detection model).

[0056] The memory system 860 is a computer or electronic storage system that is configured to store various data and provide access to various data to implement at least operations and functions as disclosed herein. The memory system 860 includes a single device or multiple devices. The memory system 860 includes electrical, electronic, magnetic, optical, semiconductor, electromagnetic, any suitable memory technology, or any combination thereof. For example, the memory system 860 may include RAM, ROM, flash memory, disk drive, memory card, optical storage device, magnetic storage device, memory module, any suitable type of memory device, or any number and combination of the above. In an example embodiment, with respect to the control system 820 and / or the processing system 840, the memory system 860 is local, remote, or a combination thereof (e.g., partially local and partially remote). For example, the memory system 860 may include at least a cloud-based storage system (e.g., a cloud-based database system), which is remote from other components of the processing system 840 and / or the control system 820.

[0057] The memory system 860 includes at least the AD system 100, which is executed via the processing system 840. The AD system 100 is configured to receive or obtain input data, which includes at least one digital image (e.g., sensor data or a digital image based on sensor data). In addition, via the processing system 840, the AD system 100 is configured to generate output data (e.g., prediction data indicating abnormality or non-abnormality) based on the input data (e.g., digital image). In this case, the AD system 100 uses at least a trained anomaly detection model 114, which has been trained using a data set 112, and the data set 112 includes source images and generated / synthesized / modified images. In addition, the memory system 860 includes: a computer vision application 880, which includes computer-readable data, which includes instructions that use the trained anomaly detection model 114 and / or the output data of the AD system 100 to provide multiple computer vision services for the control system 820. The computer vision application 880 works with the AD system 100 to provide a plurality of computer vision services (e.g., identifying normal / non-abnormal products and / or defective / abnormal products in manufacturing) to the control system 820 so that the control system 820 can control the actuator system 830 according to the computer vision services. The memory system 860 is also configured to store other related data 890 related to the operation of one or more components of the system 800 (e.g., the sensor system 810, the actuator system 830, etc.).

[0058] In addition, if Figure 8 As shown in , system 800 includes other components that contribute to the operation of control system 820 with respect to sensor system 810 and actuator system 830. Furthermore, as Figure 8 As shown in , the control system 820 includes: an I / O system 870, which includes one or more interfaces for one or more I / O devices related to the system 800. For example, the I / O system 870 provides at least one interface facing the sensor system 810 and at least one interface facing the actuator system 830. Moreover, the control system 820 is configured to provide other functional modules 850 that assist the operation of the system 800 and / or contribute to the operation of the system 800, such as any appropriate hardware technology, software technology, or any combination thereof. For example, the other functional modules 850 include an operating system and communication technology that enable the components of the system 800 to communicate with each other as described herein. At least using Figure 8 In the configuration discussed in the examples, system 800 is applicable in various technologies.

[0059] Fig. 9 According to example embodiments, mobile machine technology Figure 8As a non-limiting example, the mobile machine technology 900 includes at least a partially autonomous vehicle or robot. Fig. 9 In the embodiment of the present invention, the mobile machine technology 900 is at least a partially autonomous vehicle that includes a sensor system 810. The sensor system 810 includes an optical sensor, an image sensor, a video sensor, an ultrasonic sensor, a position sensor (e.g., a GPS sensor), a radar sensor, a LIDAR sensor, any suitable sensor, or any number and combination of the above. One or more of the sensors can be integrated with respect to the vehicle. The sensor system 810 is configured to provide sensor data to a control system 820.

[0060] The control system 820 is configured to obtain or generate image data based on sensor data or sensor fusion data from the sensor system 810. In addition, the control system 820 is configured to pre-process the sensor data to provide input data (e.g., digital image data) in a suitable form to the AD system 100. The AD system 100 is advantageously configured to provide predictive data related to anomalies with respect to autonomous driving.

[0061] In addition, the control system 820 is configured to generate actuator control data based on the computer vision application 880, which is based on at least the output data (e.g., prediction data, etc.) of the AD system 100. In this regard, the control system 820 is configured to generate actuator control data that allows safer and more accurate control of the actuator system 830 of the vehicle through the prediction data (e.g., anomaly detection and / or non-anomaly detection) provided by the AD system 100. The actuator system 830 may include a braking system, a propulsion system, an engine, a transmission system, a steering system, or any number and combination of vehicle actuators. The actuator system 830 is configured to control the vehicle so that the vehicle follows the rules of the road and avoids collisions based on at least the output data (e.g., prediction data) generated by the AD system 100 in response to receiving one or more digital images based on sensor data.

[0062] Fig.10 A schematic diagram of a control system 820 is depicted that is configured to control a system 1000 (e.g., a manufacturing machine) of a manufacturing system 1002 (e.g., a portion of a production line), which may include a punch cutter, a cutter, a gun drill, etc. The control system 820 may be configured to control an actuator system 830 that is configured to control the system 1000 (e.g., a manufacturing machine).

[0063] The sensor 810 of the system 1000 (e.g., a manufacturing machine) may be an optical sensor configured to capture one or more properties of the manufactured product 1004. The control system 820 may be configured to determine the state (e.g., abnormal or non-abnormal, defective or non-defective, etc.) of the manufactured product 1004 based on the sensor data (e.g., one or more digital images) captured by the sensor system 810 via the trained anomaly detection model 114. The actuator system 830 may be configured to control the system 1000 (e.g., a manufacturing machine) according to the determined state of the manufactured product 1004 for subsequent manufacturing steps of the manufactured product 1006. The actuator system 830 may be configured to control the function of the system 1000 (e.g., a manufacturing machine) on the subsequent manufactured product 1006 of the system 1000 (e.g., a manufacturing machine) according to the determined state of the manufactured product 1004.

[0064] Fig.11 is a security technology according to an example embodiment Figure 8 As a non-limiting example, the security technology 1100 includes at least a monitoring system, a control access system, a supervision system, or any suitable type of security device. For example, as an example, Fig.11 In connection with the security technology 1100, the security technology 1100 is configured to physically control the locked and unlocked states of the lock of the door 1102 and display the enhanced image / video on the display technology 1104. The security technology 1100 may also trigger an alarm and / or provide an electronic notification to other communication devices / technologies. In this example, the sensor system 810 includes at least an image sensor configured to provide image / video data. The sensor system 810 may also include other sensors, such as motion sensors, infrared sensors, etc.

[0065] The control system 820 is configured to obtain image / video data from the sensor system 610. In addition, the control system 820 is configured to generate actuator control data that allows for safer and more accurate control of the actuator system 830 by using the output data (e.g., predictive data indicating anomaly detection or non-anomaly detection) generated by the AD system 100. The control system 820 is configured to display the output data of the AD system 100, or any data related to the computer vision application 880, or any number and combination of the above, on the display technology 1104.

[0066] Fig.12 is an imaging technique according to an example embodiment Figure 8 Diagram of the control system. Fig.128 is a diagram of a system 800 for an imaging technique 1200 according to an example embodiment. As non-limiting examples, the imaging technique 1200 includes a magnetic resonance imaging (MRI) device, an x-ray imaging device, an ultrasound device, a medical imaging device, any suitable type of imaging device, or any number and combination of the above. Fig.12 , the sensor system 810 includes at least one image sensor. The control system 820 is configured to obtain image data from the sensor system 810. The control system 820 is also configured to generate prediction data indicating anomaly detection or non-anomaly detection based on the sensor data (e.g., digital images) at least via the trained anomaly detection model 114. In addition, the control system 820 is configured to display any relevant data (e.g., sensor data, prediction data, output data of the AD system 100, relevant data from the computer vision application 880, or any number and combination of the above) on the display 1202.

[0067] As described above, the embodiments are advantageous and provide multiple advantages. For example, the AD system 100 is configured to estimate the performance of the anomaly detection model without access to any real anomaly samples. Moreover, the AD system 100 is configured to generate synthetic anomaly samples by interpolating real normal images using the pre-trained diffusion model 104. In this regard, the generation of synthetic anomaly samples does not require model training, because the diffusion model 104 is pre-trained, and the generation of synthetic anomaly samples does not require access to real anomaly samples. The AD system 100 constructs a synthetic test set with real normal images and synthetic anomaly samples. Extensive experiments show that the rankings of various anomaly detection models 114 trained on the synthetic test set disclosed herein match their rankings on the real test set. Moreover, the synthetic test set achieves anomaly-free hyperparameter tuning of the anomaly detection model 114.

[0068] Moreover, the AD system 100 is configured to utilize the semantic space in the diffusion model to perform untrained style transfer and outlier synthesis. The AD system 100 demonstrates that even when using only inlier data as input to the diffusion process, a collection of near-distributed outliers that closely estimate the performance of a true validation data set can be created. The AD system 100 is configured to generate synthetic outliers (e.g., generated / modified / synthesized images). The AD system 100 is configured to provide an accurate estimate of the validation performance and is thus useful when selecting between candidate anomaly detection models. In this regard, the AD system 100 has been evaluated across several model types and data sets to demonstrate its generality.

[0069] Additionally, by controlling the generation of synthetic outliers and generating the dataset 112, the AD system 100 overcomes a large barrier in the deployment of anomaly detection methods, where validation techniques are often lacking in the absence of labeled anomalies, such as in industrial settings, medical settings, etc. Moreover, given new environments (e.g., a new plant or a new medical device), the AD system 100 is advantageous in providing the labeled dataset 112 with synthetic outliers in many environments where there is no existing labeled dataset that can be used to validate that a given anomaly detection model will detect the anomaly of interest.

[0070] That is, the above description is intended to be illustrative rather than restrictive, and is provided in the context of a specific application and its requirements. It can be appreciated by those skilled in the art from the above description that the present invention can be implemented in a variety of forms, and various embodiments can be implemented individually or in combination. Therefore, although the embodiments of the present invention have been described in conjunction with their specific examples, without departing from the spirit and scope of the described embodiments, the general principles defined herein can be applied to other embodiments and applications, and the true scope of the embodiments of the present invention and / or methods is not limited to the embodiments shown and described, because for technical practitioners, when studying the drawings, the specification and the appended claims, various modifications will become apparent. Additionally or alternatively, components and functions can be separated or combined differently from the various described embodiments, and different terms can be used to describe components and functions. These and other variations, modifications, additions and improvements can fall within the scope of the present disclosure as defined in the appended claims.

[0071] That is, the above description is intended to be illustrative rather than restrictive, and is provided in the context of a specific application and its requirements. It can be appreciated by those skilled in the art from the above description that the present invention can be implemented in a variety of forms, and various embodiments can be implemented individually or in combination. Therefore, although the embodiments of the present invention have been described in conjunction with their specific examples, without departing from the spirit and scope of the described embodiments, the general principles defined herein can be applied to other embodiments and applications, and the true scope of the embodiments of the present invention and / or methods is not limited to the embodiments shown and described, because for technical practitioners, when studying the drawings, the specification and the appended claims, various modifications will become apparent. Additionally or alternatively, components and functions can be separated or combined differently from the various described embodiments, and different terms can be used to describe components and functions. These and other variations, modifications, additions and improvements can fall within the scope of the present disclosure as defined in the appended claims.

Claims

1. A computer-implemented method for anomaly detection, the computer-implemented method comprising: receiving a first source image from a first set of images in the distributed image dataset and receiving a second source image from a second set of images in the distributed image dataset; generating a modified image using the first source image and the second source image via a diffusion model; automatically generating a non-anomalous label for the first source image and a non-anomalous label for the second source image; Automatically generating anomaly labels for the modified images; creating a training data set comprising at least a first source image with a non-anomalous label, a second source image with a non-anomalous label, and a modified image with an anomaly label; as well as The training dataset is used to train or fine-tune a machine learning model, wherein the machine learning model is configured to perform the task of anomaly detection.

2. The computer-implemented method of claim 1, wherein: The machine learning model comprises a classifier or a regression model; and The diffusion model is pre-trained and frozen when the modified image is generated.

3. The computer-implemented method of claim 1 , further comprising: receiving a third source image from the first set of images and a fourth source image from the second set of images; generating another modified image using the third source image and the fourth source image via the diffusion model; automatically generating a non-anomalous label for the third source image and a non-anomalous label for the fourth source image; automatically generating an anomaly label for the another modified image; as well as An additional data set is created that includes at least the third source image with the non-anomalous label, the fourth source image with the non-anomalous label, and the another modified image with the anomalous label.

4. The computer-implemented method of claim 3, further comprising: Using the additional dataset to validate the machine learning model; as well as The machine learning model is selected from among other machine learning models for testing and adoption.

5. The computer-implemented method of claim 1, wherein the modified image includes the style of the first image and the content of the second image.

6. The computer-implemented method of claim 1, wherein the diffusion model generates the modified image by performing linear interpolation on at least a first latent code of the first source image and a second latent code of the second source image.

7. The computer-implemented method of claim 1 , further comprising: employing the machine learning model to generate prediction data in response to receiving current sensor data, the prediction data indicating whether the current sensor data is abnormal data or non-anomalous data; and The predicted data is used to control the actuator.

8. A system for anomaly detection, the system comprising: one or more processors; at least one non-transitory computer-readable medium in data communication with the one or more processors, the at least one non-transitory computer-readable medium having computer-readable data including instructions stored thereon, the instructions being configured, when executed by the one or more processors, to cause the one or more processors to perform a method comprising: receiving a first source image from a first set of images in the distributed image dataset and receiving a second source image from a second set of images in the distributed image dataset; generating a modified image using the first source image and the second source image via a diffusion model; automatically generating a non-anomalous label for the first source image and a non-anomalous label for the second source image; Automatically generating anomaly labels for the modified images; creating a training data set comprising at least a first source image with a non-anomalous label, a second source image with a non-anomalous label, and a modified image with an anomaly label; and The training dataset is used to train or fine-tune a machine learning model, wherein the machine learning model is configured to perform the task of anomaly detection.

9. The system of claim 8, wherein: The machine learning model comprises a classifier or a regression model; and The diffusion model is pre-trained and frozen when the modified image is generated.

10. The system of claim 8, wherein the method further comprises: receiving a third source image from the first set of images and a fourth source image from the second set of images; generating another modified image using the third source image and the fourth source image via the diffusion model; automatically generating a non-anomalous label for the third source image and a non-anomalous label for the fourth source image; automatically generating an anomaly label for the another modified image; as well as An additional data set is created that includes at least the third source image with the non-anomalous label, the fourth source image with the non-anomalous label, and the another modified image with the anomalous label.

11. The system of claim 10, wherein the method further comprises: Using the additional dataset to validate the machine learning model; as well as The machine learning model is selected from among other machine learning models for testing and adoption.

12. The system of claim 8, wherein the modified image includes the style of the first image and the content of the second image.

13. The system of claim 12, wherein the diffusion model generates the modified image by performing linear interpolation on at least a first latent code of the first source image and a second latent code of the second source image.

14. The system of claim 8, wherein the method further comprises: employing the machine learning model to generate prediction data in response to receiving current sensor data, the prediction data indicating whether the current sensor data is abnormal data or non-anomalous data; and The predicted data is used to control the actuator.

15. A non-transitory computer readable medium having computer readable data including instructions stored thereon, the computer readable data executable by one or more processors to perform a method comprising: receiving a first source image from a first set of images in the distributed image dataset and receiving a second source image from a second set of images in the distributed image dataset; generating a modified image using the first source image and the second source image via a diffusion model; automatically generating a non-anomalous label for the first source image and a non-anomalous label for the second source image; Automatically generating anomaly labels for the modified images; creating a training data set comprising at least a first source image with a non-anomalous label, a second source image with a non-anomalous label, and a modified image with an anomaly label; as well as The training dataset is used to train or fine-tune a machine learning model, wherein the machine learning model is configured to perform the task of anomaly detection.

16. The non-transitory computer readable medium of claim 15, wherein: The machine learning model comprises a classifier or a regression model; and The diffusion model is pre-trained and frozen when the modified image is generated.

17. The non-transitory computer readable medium of claim 15, further comprising: receiving a third source image from the first set of images and a fourth source image from the second set of images; generating another modified image using the third source image and the fourth source image via the diffusion model; automatically generating a non-anomalous label for the third source image and a non-anomalous label for the fourth source image; automatically generating an anomaly label for the another modified image; as well as An additional data set is created that includes at least the third source image with the non-anomalous label, the fourth source image with the non-anomalous label, and the another modified image with the anomalous label.

18. The non-transitory computer readable medium of claim 17, using the additional data set to validate the machine learning model; and The machine learning model is selected from among other machine learning models for testing and adoption.

19. The non-transitory computer readable medium of claim 15, wherein: The modified image includes the style of the first image and the content of the second image; and The diffusion model generates a modified image by performing linear interpolation on at least a first latent code of the first source image and a second latent code of the second source image.

20. The non-transitory computer readable medium of claim 15, further comprising: employing the machine learning model to generate prediction data in response to receiving current sensor data, the prediction data indicating whether the current sensor data is abnormal data or non-anomalous data; and The predicted data is used to control the actuator.

Images