Ai-driven autoencoder system for public infrastructure

Abstract

The present invention relates to an artificial intelligence-based autoencoder system for detecting public facilities, comprising: a data collection unit that receives data measured from an ultrasonic sensor placed at a key point of a public facility; and a detection unit that analyzes the time-series data of the received ultrasonic sensor to determine whether there is an abnormality in the facility. This project (result) is the result of the Phase 3 LINC 3.0 Leading Specialized College Development Project for Industry-Academic-Research Cooperation, funded by the Ministry of Education and the National Research Foundation of Korea.

Description

Technology Field

[0001] The present invention relates to an artificial intelligence-based autoencoder system for detecting whether a public facility is in a normal state or an abnormal state, and specifically, to a system capable of detecting the state of a facility in real time by analyzing data from an ultrasonic sensor. Background Technology

[0002] Public and hazardous facilities serve as infrastructure that plays a pivotal role in social safety and economic stability, making their stable operation and maintenance a critical task. Major facilities such as bridges, tunnels, dams, and power plants are highly susceptible to structural damage or abnormal conditions due to aging and environmental factors; failure to detect and prevent these issues in a timely manner can lead to large-scale accidents or catastrophic losses. Therefore, the development of technologies that enable real-time monitoring of these facilities, early detection of abnormal conditions, and predictive maintenance is essential.

[0003] Most existing maintenance of public and hazardous facilities relies on direct, periodic human inspections and reactive maintenance methods. These traditional methods have clear limitations. They often fail to detect abnormal conditions in advance, making it difficult to prevent structural damage or sudden accidents; furthermore, despite regular inspections, they cannot monitor the condition of facilities in real time. Additionally, as modern facilities often possess large-scale and complex structures, it is difficult to comprehensively inspect all structural changes using conventional methods.

[0004] Recent advancements in sensor technology and artificial intelligence (AI) are presenting new approaches to overcome these limitations. Among these, ultrasonic sensors are powerful tools capable of precisely diagnosing the condition of structures, collecting data on vibration, stress, and deformation in real time. Organized in a time-series format, this data plays a crucial role in detecting abnormal conditions within facilities and predicting long-term structural changes. Data based on ultrasonic sensors can be utilized as a core technology for anomaly detection, as it allows for the continuous tracking of facility status and the analysis of various environmental factors and physical stresses.

[0005] An autoencoder is a type of artificial neural network capable of unsupervised learning.

[0006] In an autoencoder, the encoder takes input values ​​and performs an encoding process that transforms them into a lower dimension, while the decoder takes the encoded result and restores it to resemble the original input values. Therefore, an autoencoder can perform dimensionality reduction to transform the original data into a dense form called a latent representation, or function as a generative model capable of generating data that is very similar to the original data.

[0007] In anomaly detection, it is utilized in a way that determines an abnormal case if the result restored by an autoencoder trained on normal data differs significantly from the original data.

[0008] Generally, when performing anomaly detection, the majority of cases utilize only data of the same format (structured or unstructured of the same type). This is because heterogeneous data differs completely in input format and learning method, making it very challenging to fuse them to create anomaly detection algorithms. Prior art literature

[0009] Republic of Korea Registered Patent No. 10-2510060 Publication Date: March 14, 2023 The problem to be solved

[0010] The present invention relates to an artificial intelligence-based autoencoder system for detecting risks in public facilities, and specifically aims to provide a system capable of detecting abnormal conditions of facilities in real time by analyzing data from an ultrasonic sensor. means of solving the problem

[0011] An artificial intelligence-based autoencoder system for detecting public facilities according to one embodiment for solving the above-mentioned problem includes: a data collection unit that receives data measured from an ultrasonic sensor placed at a key point of a public facility; and a detection unit that analyzes the time-series data of the received ultrasonic sensor to determine whether there is an abnormality in the facility.

[0012] At this time, the detection unit can detect at least one of a spike, flattening, and a sharp drop by analyzing the time series data of the ultrasonic sensor using an LSTM (Long Short-Term Memory)-based autoencoder.

[0013] At this time, it is preferable for the sensing unit to analyze the time series data of the ultrasonic sensor by applying dynamic loss weighing to a contrast learning technique.

[0014] Meanwhile, the above-mentioned LSTM (Long Short-Term Memory)-based autoencoder includes an encoder that learns the temporal dependency of the data by compressing the time-series data of the ultrasonic sensor into a latent space, and expresses the compressed output vector as a 'latent vector summarizing high-dimensional information of the data'; and a decoder that restores the time-series data of the ultrasonic sensor, which is the input data, based on the latent vector, wherein the decoder restores the input data based on the latent vector, and is comparatively trained in a direction that reduces reconstructive errors occurring during the restoration process.

[0015] The above comparative learning is,

[0016] It is desirable that this be performed by a contrast learning loss function defined as. Effects of the invention

[0017] According to the artificial intelligence-based autoencoder system for detecting public facilities according to the present invention,

[0018] First, in order to analyze data collected through ultrasonic sensors and detect and predict abnormal conditions early, a loss function with contrast learning techniques and dynamic margins is applied to an LSTM-based autoencoder, so that various abnormal patterns such as flattening, spikes, and sharp declines can be effectively detected in time series data.

[0019] Second, by utilizing data augmentation and dynamic loss weights, the model's generalization performance and learning efficiency can be improved.

[0020] Third, by introducing a dynamic threshold calculation technique, stable performance can be maintained even in various environments and conditions. Since performance improvements in F1 Score and ROC AUC were confirmed compared to conventional methods, it can be seen that this can contribute to the automation of facility status detection and the enhancement of safety. Brief explanation of the drawing

[0021] FIG. 1 is a block diagram of an artificial intelligence-based autoencoder system for detecting public facilities according to the present invention. Figure 2 is a Precision-Recall curve graph of the proposed method and baseline. Figure 3 is the ROC AUC graph of the proposed method and baseline. Figure 4 is a comparison table of restoration loss according to experimental results. Specific details for implementing the invention

[0022] Preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Note that in the accompanying drawings, identical components are indicated by the same reference numerals whenever possible. Furthermore, detailed descriptions of known functions and configurations that may obscure the essence of the invention will be omitted. For the same reason, some components in the accompanying drawings may be exaggerated, omitted, or schematically depicted.

[0023] Before describing the invention, the reasons for the proposed invention and the structure and effects constituting the invention will be explained.

[0024] Anomaly detection technology utilizing ultrasonic sensor data has several significant limitations.

[0025] First, ultrasound data possesses high-dimensional time-series characteristics and includes various anomalous patterns such as transient amplitude variation, frequency distortion, and abnormal long-term trends. Advanced time-series data analysis models are required to precisely analyze and interpret these complex patterns.

[0026] Second, since sensor data is highly sensitive to the external environment and noise, data augmentation techniques are essential to effectively preprocess the data and train models robust against noise.

[0027] Third, existing anomaly detection methods based on fixed thresholds are highly likely to experience performance degradation when environmental conditions change, and a dynamic and flexible threshold setting method capable of adapting to environmental changes is required.

[0028] The present invention includes an AI-based anomaly detection and prediction model utilizing ultrasonic sensor data. The configuration of the present invention performs anomaly detection by integrating a contrast learning technique for performance improvement and a loss function with a dynamic margin through a Long Short-Term Memory (LSTM)-based autoencoder.

[0029] In addition, data augmentation and dynamic thresholding techniques can improve the model's generalization performance and maintain stable performance in various environments.

[0030] The core components of the present invention are as follows. First, contrast learning and a dynamic margin loss function are introduced to an AI model designed to effectively detect and predict various abnormal patterns based on ultrasonic sensor data.

[0031] This invention secures the practical applicability of the model through data augmentation and dynamic threshold setting techniques. This enables the practical possibility of self-diagnosis and predictive maintenance for public and hazardous facilities.

[0032] FIG. 1 is a block diagram of an artificial intelligence-based autoencoder system for detecting public facilities according to an embodiment of the present invention. FIG. 1 is an implementation of the main configuration of the present invention in the form of a system. Referring to FIG. 1, the artificial intelligence-based autoencoder system (1000) according to the present invention includes a data collection unit (100) and a detection unit (200).

[0033] The data collection unit (100) is a hardware device that receives measurement data from an ultrasonic sensor placed at a key point of a public facility and includes a physical interface connected to the ultrasonic sensor via wired or wireless connection. Considering that the ultrasonic sensor data handled in the present invention is time-series data, it is preferable that the data collection unit (100) be configured with hardware capable of minimizing noise and data contamination.

[0034] The detection unit (200) analyzes the received ultrasonic sensor data to determine whether the facility is normal or abnormal.

[0035] According to one embodiment, the detection unit (200) uses an LSTM (Long Short-Term Memory)-based autoencoder to analyze time-series data of an ultrasonic sensor and detects at least one of abnormal data such as a spike, a flatline, and a drop.

[0036] Spikes, flattening, and sharp drops are merely examples of abnormal data for detecting anomalies, and the type of abnormal data to be detected may be appropriately changed depending on the type and characteristics of the public facilities to be detected.

[0037] According to one embodiment, the sensing unit (200) can analyze time series data of an ultrasonic sensor by applying dynamic loss weighing to a contrastive learning technique.

[0038] The sensing unit (200) includes an encoder (210) and a decoder (220). The sensing unit (200) can be implemented in software on a general-purpose computing device, for example, in a cloud server of a data center, and depending on the scale of implementation, it may also be implemented on a personal information communication terminal such as a personal PC or a smartphone. The sensing unit (200) is configured using an LSTM (Long Short-Term Memory)-based autoencoder, and the encoder (210) and decoder (220) are logical components forming the autoencoder and are not specifically limited as hardware constituting the device.

[0039] The encoder (210) learns the temporal dependency of the data by compressing the time series data into a latent space, and expresses the compressed output vector as a 'latent vector summarizing high-dimensional information of the data'.

[0040] The decoder (220) restores the input data, which is time series data (time series data of the ultrasonic sensor), based on the latent vector.

[0041] At this time, the decoder (220) is contrast-learned in a way that restores the input data based on a latent vector, while reducing the reconstructive error that occurs during the restoration process, and the contrast-learning is the following formula

[0042] It is desirable that this be performed by a contrast learning loss function defined as.

[0043] Here is a latent vector, and are the latent vectors corresponding to the positive and negative pairs, respectively, is a pair of labels (1 for positive, 0 for negative), is cosine similarity, represents the threshold value.

[0044] Next, the present invention will be described in more detail based on experimental data and results.

[0045] Experimental data

[0046] The data used for verification of the present invention is time-series data collected on the condition of public and hazardous facilities through an ultrasonic sensor. The ultrasonic sensor data is time-series data, and will be briefly referred to as 'time-series data' for explanation.

[0047] Time-series data was continuously collected through ultrasonic sensors attached to the facility and includes key features capable of distinguishing between normal and abnormal states. As ultrasonic sensor data is a signal reflecting the structural characteristics of the facility, it responds sensitively to vibrations, stress, or changes in the external environment, thereby accurately indicating the condition of the facility.

[0048] Each sample of the time series data was extracted from an interest interval of 4,000-5,000 indices where an abnormal state is likely to occur. The extracted data was divided into time series data of fixed length (100) and a sliding window was applied.

[0049] The sliding window is designed to maintain the temporal continuity of the data by creating overlapping segments of the shape. Each sample can improve the stability and efficiency of model training by adjusting the range of values ​​through normalization during the preprocessing stage.

[0050] Anomaly detection techniques

[0051] The present invention is designed based on an LSTM autoencoder by combining a contrastive learning technique and dynamic loss weighting, and aims to provide an unsupervised learning-based anomaly detection system that operates stably even in unlabeled environments. The detailed configuration of the present invention will be described below.

[0052] The sensing unit (200) of the present invention has a configuration including an LSTM autoencoder of an encoder-decoder structure.

[0053] The encoder effectively learns the temporal dependencies of the data by compressing time-series data into a latent space, and the compressed output vector is represented as a latent vector that summarizes the high-dimensional information of the data.

[0054] The decoder is designed to reconstruct input data based on these latent vectors and is trained to minimize reconstruction errors that occur during the restoration process of normal data.

[0055] In unlabeled environments, reconstruction error serves as a key criterion for distinguishing between normal and abnormal data. Normal data exhibits low reconstruction error because it matches learned patterns, whereas abnormal data generates high reconstruction error due to differences from learned normal patterns. This characteristic is utilized as a suitable metric for unsupervised learning-based anomaly detection, and the formula defined below can be used.

[0056]

[0057] is input data, is restored data, represents the number of data points.

[0058] During the reconstruction process, abnormal data generates a higher reconstruction error compared to normal data, thereby enabling the effective detection of abnormal conditions. However, relying solely on reconstruction errors may result in the failure to effectively detect subtle abnormal patterns in the data or cases where the boundary with normal data is unclear. To address this, the present invention discloses a configuration for contrast learning.

[0059] Contrast learning learns the similarities and differences between data to improve the model's latent spatial representation and more clearly distinguish anomalies. Positive and negative pairs are formed, and positive pairs are trained to maximize similarity while negative pairs are trained to maximize difference. The contrast learning loss function can be defined as follows.

[0060]

[0061] Here is a latent vector, and are the latent vectors corresponding to the positive and negative pairs, respectively, is a pair of labels (1 for positive, 0 for negative), is cosine similarity, represents the threshold value.

[0062] This invention introduces dynamic loss weights to balance the reconstruction loss function and the contrast learning loss function during the learning process. This is designed considering that the learning focus differs between the early and later stages of learning.

[0063] In the initial training phase, it is important for the model to learn the structural features of the input data based on normal data and enhance its reconstruction capabilities. In this phase, reconstruction loss serves as a key indicator of learning, helping the model learn normal state patterns and accurately reproduce the temporal dependencies of the data. Therefore, the weight of the reconstruction loss function is initially set high.

[0064] As training progresses, the model sufficiently learns the ability to restore normal data, and the focus shifts to distinguishing abnormal states. At this point, the weight of the contrast learning loss function gradually increases. The contrast learning loss focuses on clearly distinguishing normal and abnormal data in the latent space and helps the model optimize the relationships between latent vectors. Through this, the model can distinguish between normal and abnormal states more effectively.

[0065] The final loss function is expressed as a weighted sum between two losses, as shown in the formula below.

[0066]

[0067] It decreases linearly as learning progresses, starting at an initial value of 1 and gradually reaching 0. This method focuses on learning the restoration of normal data in the early stages and strengthens the ability to distinguish abnormal states in the later stages.

[0068] In addition, the present invention introduces a data augmentation technique to efficiently process input data and enhance the learning stability of the model. Data augmentation helps the model adapt to various data patterns by generating new training samples through the modification of existing data. This allows the model to mitigate the problem of overfitting and learn a wider data distribution even in a limited dataset environment.

[0069] Fine noise was added to the data to verify the invention. This addition of noise reflects the uncertainty that may occur in the actual operating environment of the sensor data, enabling the model disclosed in the present invention (i.e., the detection unit (200)) to learn input data under various conditions. Through this, the model acquires the ability to reliably detect abnormal states even in data containing noise.

[0070] In addition, time axis transformation was applied to improve performance regarding temporal variations in the data. This distorts or transforms the time-series characteristics of the input data, inducing the model to learn temporal dependencies in various ways. Time axis transformation can help the model respond more sensitively to temporal changes in the input data by shifting the data forward or backward or adjusting the sampling interval.

[0071] The data augmentation technique applied in this invention goes beyond simply increasing the amount of data to enable the model to adapt to a wider variety of situations and comprehensively understand potential changes in the data. In particular, by improving generalization performance, it provides a foundation for the model to operate stably on new data without relying on training data. Given the characteristics of sensor data, the aforementioned data augmentation technique can play a crucial role in ensuring stable anomaly detection in various environments.

[0072] Experiment setup

[0073] To verify the performance of the present invention, we will describe a test experiment combining synthetic data and actual ultrasound data.

[0074] The purpose of the experiment is to quantitatively evaluate the ability of a model (i.e., a detection unit (200)) to distinguish between normal data and abnormal data.

[0075] The test data was constructed based on synthetic data and actual ultrasound data.

[0076] Synthetic data was generated by creating a total of 5,000 samples based on sinusoidal time-series data, with each sample consisting of 100 time-series points. Anomaly data was generated by inserting structural anomaly patterns such as flatlines, spikes, and drops. These anomaly patterns reflect abnormal conditions that may occur in actual sensor environments.

[0077] The actual ultrasound data consists of time-series data collected during the process of monitoring the condition of public and hazardous facilities, and includes both normal and abnormal states. In particular, the actual dataset is divided into v1, v2, and v3, which were collected under different environmental conditions, weather, and operational situations. Each dataset was collected on different dates, and environmental factors at the time of collection influenced the data characteristics, resulting in structural differences between the datasets.

[0078] These characteristics play an important role in evaluating how effectively a model can adapt to variability between data.

[0079] The test data consisted of a total of 1,000 samples, combining 500 ultrasound samples and 500 synthetic samples. This configuration was designed to comprehensively evaluate the performance of the model by simultaneously reflecting the controlled characteristics of the synthetic data and the realism of the actual data.

[0080] The generalization performance and adaptability between datasets of the models were analyzed by performing cross-tests on models trained on each dataset (v1, v2, v3) on other datasets, reflecting the differences between the datasets.

[0081] Through test experiments, abnormal conditions were detected based on reconstruction errors. The present invention is designed so that normal data exhibits low reconstruction errors and abnormal data exhibits high reconstruction errors; therefore, it provides flexibility in response to data characteristics and environmental changes through a dynamic threshold, enabling stable anomaly detection even in various situations.

[0082] Model performance was evaluated quantitatively using key metrics such as the F1 score, ROC AUC, and Precision-Recall Curve (PRC). The F1 score was used to assess the harmony between precision and recall, while the ROC AUC measured the overall ability to separate normal and abnormal data. The PRC was used as a useful metric to evaluate model performance in data-imbalanced environments.

[0083] In each test scenario, the difference between the model's output value and the actual label was analyzed to evaluate how clearly the model disclosed in this invention can distinguish between normal and abnormal data. An experimental design combining synthetic and real data provides an environment for verifying the model's anomaly detection performance more deeply, and allows for the systematic evaluation of the model's generalization performance and stability under various data conditions.

[0084] Experimental results

[0085] In an experiment to verify the present invention, the performance of the proposed contrast-learning-based LSTM autoencoder and a non-learning-based LSTM autoencoder were compared using the v1, v2, and v3 datasets, and the generalization ability of the model was verified. The experiment was conducted using a cross-test method in which the model was trained on one dataset and then tested on two other datasets. This setup evaluated the model's variability between data and its adaptability to various environments.

[0086] Figure 2 is a Precision-Recall curve graph of the proposed method and baseline, and Figure 3 is an ROC AUC graph of the proposed method and baseline.

[0087] In Figure 2, (a), (b), and (c) are graphs of the model trained with v1, v2, and v3 data using the proposed method, and (d), (e), and (f) are results of the baseline model trained with the same data.

[0088] In Figure 3, (a), (b), and (c) are ROC-AUC graphs of the proposed method trained with v1, v2, and v3 data, and (d), (e), and (f) are results of training a baseline model using the same data.

[0089] Figure 4 is a comparison table of restoration loss according to experimental results.

[0090] As a result of measuring the restoration loss, the proposed contrast learning-based model recorded superior performance compared to the existing model under all test conditions. As shown in Figure 4, the proposed method exhibits a stable restoration loss value of 0.0022 on average. This indicates robust adaptability to variability between data. On the other hand, the baseline model showed a high loss value of 0.0165 on average. Furthermore, it can be seen that the loss value of the existing model showed significant variability depending on the data, and performance degradation was pronounced under some conditions.

[0091] The proposed model recorded an average ROC AUC value of 0.95 or higher under all test conditions, demonstrating high discrimination ability to effectively distinguish between normal and abnormal data. In PRC analysis as well, it maintained a balance between high precision and recall, exhibiting stable detection performance even in situations of data imbalance. This indicates that the proposed method induced the learning of the boundary between normal and abnormal data in the latent space relative to the baseline.

[0092] As explained in the experimental results, the model proposed in this invention demonstrated consistent and superior performance compared to existing models in key evaluation metrics such as restoration loss, ROC AUC, and PRC. These results support the effectiveness of the contrast learning technique and dynamic loss weights, which induce the proposed model to clearly learn the potential boundary between normal and abnormal data. Furthermore, it can be seen that dynamic thresholding and data augmentation techniques further enhance model performance in various environments and provide robust detection capabilities against uncertainty and noise that may occur in actual sensor environments.

[0093] This invention presents a new approach that overcomes the various limitations of existing anomaly detection technologies and makes a substantial contribution to real-time monitoring and predictive maintenance systems for public and hazardous facilities. Future research is expected to enable the application of more comprehensive models by fusing various types of sensor data in addition to ultrasonic sensors, thereby contributing to enhancing the stability and reliability of public infrastructure.

[0094] The embodiments of the invention disclosed in this specification and drawings are provided merely as specific examples to facilitate the explanation of the technical content of the invention and to aid in understanding the invention, and are not intended to limit the scope of the invention. It is obvious to those skilled in the art that other variations based on the technical concept of the invention are possible in addition to the embodiments disclosed herein. Explanation of the symbols

[0095] 1000: AI-based autoencoder system for detecting public facilities 100 : Data Collection Unit 200 : Detector

Claims

Claim 1 An artificial intelligence-based autoencoder system comprising: a data collection unit that receives data measured from ultrasonic sensors placed at key points of public facilities; and a detection unit that analyzes time-series data from the received ultrasonic sensors to determine whether there is an abnormality in the facilities. Claim 2 An artificial intelligence-based autoencoder system according to claim 1, wherein the sensing unit analyzes time-series data of the ultrasonic sensor using an LSTM (Long Short-Term Memory)-based autoencoder to detect at least one of abnormal data such as a spike, a flatline, and a drop. Claim 3 An artificial intelligence-based autoencoder system according to claim 2, wherein the sensing unit analyzes time series data of the ultrasonic sensor by applying dynamic loss weighing to a contrastive learning technique. Claim 4 In claim 3, the LSTM (Long Short-Term Memory)-based autoencoder comprises: an encoder that learns the temporal dependency of the data by compressing the time-series data of the ultrasonic sensor into a latent space, and expresses the compressed output vector as a 'latent vector summarizing high-dimensional information of the data'; and a decoder that restores the time-series data of the ultrasonic sensor, which is the input data, based on the latent vector, wherein the decoder restores the input data based on the latent vector, and is contrast-learned in a direction that reduces reconstructive errors occurring during the restoration process, and the contrast-learning The contrast learning loss function defined as (where, where is a latent vector, and are the latent vectors corresponding to the positive and negative pairs, respectively, is a pair of labels (1 for positive, 0 for negative), is cosine similarity, An artificial intelligence-based autoencoder system characterized by being performed by (indicating a threshold value).

Images