System and Method for Forecasting Vessel FGSS failure based on LSTM AE

Abstract

The present invention relates to an LSTM-AE-based FGSS failure prediction system and method for ships that utilizes ship FGSS time series data to detect abnormal signals and applies them to an anomaly detection technique to identify the time and cause of failure. The system comprises: a data collection unit that collects sensor data output from the ship's FGSS equipment; a data preprocessing unit that removes data with errors occurring during the data reception process and performs data normalization; a model learning unit that learns the definition of the normal state of sensor data installed in the FGSS equipment and the range of change due to noise through an LSTM-AE algorithm; and an equipment status information providing unit that provides information about the state of the equipment.

Description

Technology Field

[0001] The present invention relates to the management of FGSS for ships, and specifically to an LSTM AE-based FGSS failure prediction system and method for ships that utilizes ship FGSS time series data to detect abnormal signals and applies them to an anomaly detection technique to enable the identification of the time and cause of the failure. Background Technology

[0002] Following the announcement of revised environmental regulations by the International Maritime Organization (IMO), the sulfur content of ship fuel has been tightened from the existing 3.5% to 0.5%, and interest in and the need for ships using LNG as fuel are increasing.

[0003] Accordingly, the demand for Fuel Gas Supply Systems (FGSS), which use LNG as propulsion fuel for ships, is also increasing.

[0004] The FGSS (Fuel Gas Supply System) is a fuel supply system for LNG vessels that prepares the LNG (Liquefied Natural Gas) and BOG (Boil-off Gas) inside the tank for engine fuel supply.

[0005] In other words, it is a system that supplies gaseous fuel, such as LNG (Liquefied Natural Gas), to ship engines.

[0006] Because the Fuel Gas Supply System (FGSS) involves the interconnection of various equipment and systems, a failure is critical.

[0007] However, like most marine equipment currently, these FGSS also require maintenance and repair activities to be performed on-site due to long-term voyages. Since maintenance and repair of marine equipment at sea is significantly limited compared to that on land, it incurs a lot of time and cost, and the causes of failure are not properly identified.

[0008] Furthermore, most ship maintenance techniques rely on corrective maintenance, which repairs equipment after a breakdown occurs, and preventive techniques such as planned maintenance, which involves periodically replacing equipment regardless of whether a breakdown has occurred.

[0009] As a result, maintenance is performed regardless of the actual level of defects, leading to high costs due to frequent interruptions and parts replacements.

[0010] In particular, although there are attempts to diagnose the current status of FGSS equipment through statistical approaches to detect failures, these methods often fail to diagnose the condition because the operating environment and characteristics of the vessel vary in most cases. Consequently, there are limitations to diagnosis, or estimations are possible with very low precision, making them unsuitable for application on ships.

[0011] Therefore, the development of new technologies is required to enable maintenance through predictive maintenance of ships in order to improve efficiency, minimize operational restrictions caused by breakdowns, and ensure smooth ship operations. Prior art literature

[0012] Republic of Korea Published Patent No. 10-2021-0143106 Republic of Korea Published Patent No. 10-2022-0167008 Republic of Korea Published Patent No. 10-2022-0064098 The problem to be solved

[0013] The present invention aims to solve the problems of conventional ship FGSS management technology by providing a ship FGSS failure prediction system and method based on LSTM AE that utilizes ship FGSS time series data to detect abnormal signals and applies them to an anomaly detection technique to identify the time and cause of the failure.

[0014] The present invention aims to provide an LSTM-AE-based shipboard FGSS fault prediction system and method that uses an LSTM-AE algorithm to define the range of normal data to diagnose the current state of the FGSS equipment, and includes noise signals in the learning process to correct noise and loss caused by the ship's operating environment of the lidar, thereby enabling additional estimation of the state in a noisy environment.

[0015] The present invention aims to provide an LSTM AE-based FGSS failure prediction system and method for ships that can enhance the failure prediction characteristics of an FGSS system using an LSTM AE (Auto-Encoder) by collecting actual data from a Vesselink and considering the time point of the ship data to extract characteristic factors (outliers, which are characteristic data) that may be the cause of failure of the ship FGSS.

[0016] The present invention aims to provide an LSTM AE-based FGSS failure prediction system and method for ships that can analyze the causal relationship of system failures through outlier data and identify the cause of failure data occurrence, thereby predicting signs of abnormality in ship equipment and systems in advance to protect equipment and prevent abnormal operation.

[0017] The present invention aims to provide an LSTM AE-based FGSS fault prediction system and method for ships that predicts abnormal signs in ship equipment and systems in advance, provides a specific alarm and fault diagnosis system regarding ship failures, and enables efficient finding of maintenance timing.

[0018] Other objectives of the present invention are not limited to those mentioned above, and other unmentioned objectives will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0019] The LSTM AE-based ship FGSS fault prediction system according to the present invention for achieving the above-mentioned purpose is characterized by comprising: a data collection unit that collects sensor data output from the ship's FGSS equipment; a data preprocessing unit that removes data with errors occurring during the data reception process and performs data normalization; a model learning unit that learns the definition of the normal state of sensor data installed in the FGSS equipment and the range of change due to noise through an LSTM-AE algorithm; and an equipment status information providing unit that provides information about the state of the equipment.

[0020] Here, the AE (AutoEncoder) that performs the LSTM-AE algorithm in the model training section is composed of an encoder and a decoder, with the number of nodes being the same for input and output. It is characterized by generating a reconstruction error through a process of receiving data in the input layer, compressing information as the number decreases as it goes to the hidden layer, and restoring it back to the output layer, and using the degree of the reconstruction error that occurs at this time as an anomaly score to compare with a threshold to determine whether the data is normal or abnormal.

[0021] And the AE (AutoEncoder) does not require labels for normal and abnormal data, and learning proceeds in a direction similar to the input and output data. In the encoder stage, features of the input data are extracted and compressed through a hidden layer and sent to a latent layer. In the latent layer, the most distinctive data features are prioritized for reconstruction. In the decoder stage, the data is retrieved from the latent layer and restored to the original data.

[0022] Furthermore, data loss occurs during the process of compressing data in the encoder layer and sending it to the latent layer, resulting in a difference between the restored data and the original data. Additionally, after training is complete, the newly input data is characterized by the loss of features with lower priority compared to the original data, and the data being restored in a direction similar to the existing training data.

[0023] In addition, if the TIME_STAMP value in the FGSS data extracted from the data preprocessing unit is NaN or there are duplicate values, it is determined that an error occurred during the process of receiving data from the Vessel link and removed, and downsampling is performed to reduce the size of the data set.

[0024] In addition, to make the values ​​of each attribute of the data within a certain range, the Min-Max Scaler function is used to normalize all data so that they have values ​​between 0 and 1, and correlation analysis is performed between ship FGSS data features before training to sub-group features with high Pearson correlation coefficients (pearson correlation coefficient 0.6 or higher).

[0025] Furthermore, it is characterized by reducing training time and resources by reducing dimensionality through feature grouping, and enabling all features to be trained equally without placing training weight on large-scale features.

[0026] In addition, the model training unit is characterized by changing the shape of the input data to 3 dimensions to use the LSTM-AE model, and performing a sliding window of the data to learn the temporal dependency, which is a characteristic of time series data.

[0027] In addition, to verify the results of an LSTM-AE model that performs anomaly detection by compressing normal data from an encoder to a latent layer and then restoring it through a decoder to compare the difference with the original, the data restoration value is characterized by defining an anomaly score through the reconstruction probability of the trained model.

[0028] In addition, to identify anomaly data, a point that is 99% or higher in the cumulative probability distribution of a loss function defined as an anomaly score is designated as a threshold, and data points with an anomaly score higher than the threshold are determined to be anomaly data.

[0029] A ship FGSS fault prediction method based on LSTM AE according to the present invention for achieving other purposes is characterized by comprising: a data collection step for collecting sensor data output from FGSS equipment of a ship; a data preprocessing step for removing data that has errors occurring during the data reception process and performing data normalization; a model learning step for learning the definition of the normal state of sensor data installed in the FGSS equipment and the range of change due to noise through an LSTM-AE algorithm; and an equipment status information providing step for providing information about the state of the equipment.

[0030] Here, during the model training phase, data is received from the input layer, information is compressed as it moves toward the hidden layer with a decreasing number, and then restored to the output layer to generate a reconstruction error. The degree of this reconstruction error is used as an anomaly score and compared with a threshold to determine whether the data is normal or abnormal.

[0031] In addition, it is characterized by not requiring labels for normal and abnormal data, learning proceeds in a direction where the input and output data are similar, the encoder stage extracts features from the input data and compresses them through a hidden layer to send to a latent layer, the latent layer prioritizes the most distinctive data features for reconstruction, and the decoder stage retrieves data from the latent layer and restores it to the original data.

[0032] Furthermore, data loss occurs during the process of compressing data in the encoder layer and sending it to the latent layer, resulting in a difference between the restored data and the original data. Additionally, after training is complete, the newly input data is characterized by the loss of features with lower priority compared to the original data, and the data being restored in a direction similar to the existing training data.

[0033] In addition, in the data preprocessing step, if the TIME_STAMP value in the extracted FGSS data is NaN or there are duplicate values, it is determined that an error occurred during the process of receiving data from the Vessel link and removed; downsampling is performed to reduce the size of the dataset; normalization is performed using the Min-Max Scaler function to ensure that all data have values ​​between 0 and 1 so that the values ​​of each attribute of the data are within a certain range; and correlation analysis is performed between the vessel FGSS data features before training, and sub-grouping is performed on features with high Pearson correlation coefficients (pearson correlation coefficient 0.6 or higher).

[0034] In addition, to use the LSTM-AE model during the model training phase, the shape of the input data is changed to 3 dimensions, and a sliding window of the data is performed to learn the temporal dependency, which is a characteristic of time series data.

[0035] In addition, to verify the results of an LSTM-AE model that performs anomaly detection by compressing normal data from an encoder to a latent layer and then restoring it through a decoder to compare the difference with the original, the data restoration value is characterized by defining an anomaly score through the reconstruction probability of the trained model.

[0036] In addition, to identify anomaly data, a point that is 99% or higher in the cumulative probability distribution of a loss function defined as an anomaly score is designated as a threshold, and data points with an anomaly score higher than the threshold are determined to be anomaly data. Effects of the invention

[0037] The shipboard FGSS fault prediction system and method based on LSTM AE according to the present invention, as described above, have the following effects.

[0038] First, abnormal signals are detected by utilizing ship FGSS time series data, and this is applied to anomaly detection techniques to enable the identification of the time and cause of failure.

[0039] Second, the LSTM-AE algorithm is used to define the range of normal data to diagnose the current state of the FGSS equipment, and noise signals are included in the learning process to correct for noise and loss caused by the ship's operating environment of the lidar, thereby enabling additional estimation of the state in a noisy environment.

[0040] Third, after collecting actual data from Vesselink and considering the time of the vessel data, characteristic factors (outliers, which are characteristic data) that may be the cause of failure in the vessel FGSS are extracted to enhance the failure prediction characteristics of the FGSS system using LSTM AE (Auto-Encoder).

[0041] Fourth, through outlier data, the causal relationship of system failures can be analyzed and the cause of failure data can be identified, thereby enabling the prediction of abnormal signs in ship equipment and systems in advance to protect equipment and effectively prevent abnormal operation.

[0042] Fifth, by predicting signs of abnormalities in ship equipment and systems in advance, a specific warning and fault diagnosis system regarding ship failures is established to efficiently determine maintenance timing. Brief explanation of the drawing

[0043] FIG. 1 is a configuration diagram of an LSTM AE-based FGSS fault prediction system for ships according to the present invention. Figure 2 is a configuration diagram of an LSTM-AutoEncoder. FIG. 3 is a flowchart illustrating a shipboard FGSS fault prediction method based on LSTM AE according to the present invention. Figure 4 is a flowchart of the pretreatment process according to the present invention. Figure 5 is a configuration diagram showing an example of downsampling after classification of ship FGSS data. Figure 6 is a configuration diagram showing an example of correlation analysis of ship FGSS data. FIG. 7 is a flowchart illustrating the learning process of a model according to the present invention. Figure 8 is the cumulative probability distribution of the loss function Figure 9 is a graph of abnormal data detection in subgroup 1. Specific details for implementing the invention

[0044] Hereinafter, preferred embodiments of the LSTM AE-based FGSS fault prediction system and method for ships according to the present invention will be described in detail as follows.

[0045] The features and advantages of the LSTM AE-based FGSS fault prediction system and method for ships according to the present invention will become apparent through the detailed description of each embodiment below.

[0046] Figure 1 is a configuration diagram of a ship FGSS fault prediction system based on LSTM AE according to the present invention.

[0047] The terms used in this disclosure have been selected to be as widely used and general as possible, taking into account their functions within this disclosure; however, these terms may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been selected at the applicant's discretion, and in such cases, their meanings will be described in detail in the relevant description of the invention. Therefore, terms used in this disclosure should be defined not merely by their names, but based on their meanings and the overall content of this disclosure.

[0048] When a part of a specification is described as "including" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as "...part" or "module" as used in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware or software, or as a combination of hardware and software.

[0049] The LSTM AE-based ship FGSS failure prediction system and method according to the present invention utilize ship FGSS time series data to detect abnormal signals and apply them to an anomaly detection technique to enable the identification of the time and cause of the failure.

[0050] To this end, the present invention may include a configuration that uses an LSTM-AE algorithm to define a range of normal data to diagnose the current state of the FGSS equipment, and includes a noise signal in the learning process to correct noise and loss caused by the operating environment of the LiDAR on the ship, thereby enabling additional estimation of the state in a noisy environment.

[0051] The present invention may include a configuration that can enhance the failure prediction characteristics of an FGSS system using an LSTM AE (Auto-Encoder) by collecting actual data from a Vesselink and considering the time of the vessel data to extract characteristic factors (outliers, which are characteristic data) that may be the cause of failure of the vessel FGSS.

[0052] As shown in FIG. 1, the shipboard FGSS fault prediction system based on LSTM AE according to the present invention includes a data collection unit (10) that collects sensor data output from the shipboard FGSS equipment, a data preprocessing unit (20) that removes data that has errors occurring during the data reception process and performs data normalization, a model learning unit (30) that learns the definition of the normal state of the sensor data installed in the FGSS equipment and the range of change due to noise through an LSTM-AE algorithm, and an equipment status information providing unit (40) that provides information about the state of the equipment.

[0053] The shipboard FGSS fault prediction system based on LSTM AE according to the present invention, having such a configuration, can analyze the causal relationship of system failure through outlier data and identify the cause of failure data occurrence, thereby predicting signs of abnormality in ship equipment and systems in advance to protect equipment and prevent abnormal operation.

[0054] In particular, it enables the prediction of abnormal signs in ship equipment and systems in advance, establishing a specific warning and fault diagnosis system for ship failures, and efficiently determining maintenance timing.

[0055] In the present invention, 'abnormal data' is defined as follows.

[0056] In a ship system, all data points that deviate significantly from the expected or normal value range, and data that does not follow the expected pattern, are defined as 'abnormal data'.

[0057] As such, unusual data that does not match expected patterns or behaviors can be a cause of failure.

[0058] There are various types of abnormal data on ships. Outliers refer to data points that deviate significantly from the normal distribution of values ​​for specific parameters and are caused by sensor malfunctions, equipment failures, or external environmental factors.

[0059] Deviation of data from the expected range in data trends leads to emerging issues or potential faults, which can manifest as a sudden increase or decrease in data, unexpected changes in patterns, or abnormal correlations between parameters.

[0060] In addition, if data inconsistent with the data itself or the data points is detected, it may be caused by problems with sensor measurements, unexpected signal changes, or illogical readings that disregard physical principles.

[0061] The present invention extracts such abnormal data to predict failures and detects them early before they develop into serious problems, thereby minimizing ship downtime, repair costs, and risk factors.

[0062] This improves the overall performance and efficiency of the ship system and can lead to reduced fuel consumption, optimized engine operation, and extended equipment lifespan.

[0063] The anomaly detection model is explained as follows.

[0064] There are various models for detecting anomaly data, including statistical analysis, regression analysis, machine learning, and deep learning; among these, machine learning and deep learning models are being researched most actively.

[0065] Depending on the presence or absence of data labels, it can be divided into supervised, semi-supervised, and unsupervised learning.

[0066] Data extracted from ships exhibits data class imbalance.

[0067] Ship system data consists of lists focused on the performance and efficiency of the system, and it is realistically very difficult to secure high-quality data capable of diagnosis and prediction by sensitively reflecting failure phenomena.

[0068] The ship FGSS fault diagnosis algorithm according to the present invention uses an unsupervised learning-based model to reflect the difficulty in collecting abnormal data due to the characteristics of ships.

[0069] In unsupervised learning, data labels do not exist.

[0070] Therefore, a threshold is required to distinguish between normal and abnormal data for anomaly detection. If the anomaly score is higher than the threshold, it is determined to be abnormal data, and if it is lower, it is determined to be normal data.

[0071]

[0072] The model used in this invention is LSTM-AE The anomaly score is defined using the negative log likelihood of the observations.

[0073] represents the mean and covariance of the distribution of reconstruction scores calculated through the model, and represents the estimated outlier score value.

[0074] A high Anomaly Score indicates that the reconstruction from the LSTM-AE model was not successful, meaning that data points exceeding the range of learned input data values ​​have been restored.

[0075] Here, the threshold is determined as the point where the cumulative probability density is 99% or higher when training is performed with normal data.

[0076] The LSTM-AE is explained as follows.

[0077] Figure 2 is a configuration diagram of an LSTM-AutoEncoder.

[0078] Since abnormal data is found in smaller numbers compared to normal data in most vessels, this invention uses an Auto-Encoder (AE) technique that utilizes only normal data during the learning process.

[0079] An AutoEncoder consists of an encoder and a decoder, and the number of nodes is the same for input and output.

[0080] Data is received from the input layer, and as it moves to the hidden layer, the number of data points decreases while compressing the information. This process of restoring the data to the output layer generates a reconstruction error; the degree of this reconstruction error is used as an anomaly score and compared with a threshold to determine whether the data is normal or abnormal.

[0081] AE (AutoEncoder) does not require labels for normal and abnormal data, and learning proceeds in a direction where the input and output data are similar.

[0082] In the encoder stage, features are extracted from the input data, and the data is compressed and sent to the latent layer via the hidden layer. In the latent layer, the most distinctive data features are prioritized for reconstruction.

[0083] In the decoder stage, data is retrieved from the latent layer and restored to its original form. Typically, data loss occurs during the process of compressing data in the encoder layer and sending it to the latent layer, which results in a difference between the restored data and the original data.

[0084] After training is complete, the newly input data loses features with lower priority compared to the original data and is restored in a direction similar to the existing training data.

[0085] In addition, since ship system data extracts multidimensional time series data, an LSTM-AutoEncoder algorithm with an applied LSTM is used to learn the dependency on time.

[0086] In the present invention, data collection involves collecting feature data of the ship system and collecting real-time data at regular intervals from the Vessel Link.

[0087] Extract data from Vessel Link and save it as a CSV file.

[0088] The LSTM AE-based ship FGSS fault prediction method according to the present invention is described in detail as follows.

[0089] FIG. 3 is a flowchart illustrating a ship FGSS fault prediction method based on LSTM AE according to the present invention.

[0090] The LSTM AE-based ship FGSS fault prediction method according to the present invention comprises, as shown in FIG. 3, a step (S501) of inputting various sensor data output from the ship's FGSS equipment, and

[0091] It includes a step (S502) of collecting changes in output values ​​and noise of the sensor being operated over time, a step (S503) of performing learning through an LSTM-AE algorithm, and a step (S504) of learning the definition of the normal state of the sensor data installed in the FGSS equipment and the range of change due to noise to provide information about the state of the equipment.

[0092] The data preprocessing process can be explained in detail as follows.

[0093] Figure 4 is a flowchart of the pretreatment process according to the present invention.

[0094] In one embodiment according to the present invention, a total of 6 features are extracted from the ship FGSS data.

[0095] Features of unextracted data are stored in the Cloud to be used as data for researching new anomaly detection models.

[0096] If the TIME_STAMP value in the extracted FGSS data is NaN or contains duplicate values, it is advisable to remove it because an error occurred during the process of receiving data from the Vessel link.

[0097] In the case of ship FGSS data, since there is almost no fluctuation with a data collection cycle (10 seconds), it is downsampled and reduced to 1-minute intervals.

[0098] This reduces the size of the dataset, lowering the multiplicity of learning and enabling the model to learn patterns that are more generalizable than learning specific parts.

[0099] Figure 5 is a configuration diagram showing an example of downsampling after classification of ship FGSS data.

[0100] Learning models rely on optimization techniques, but since optimization can be difficult if there are data samples with significantly different scales, normalization is performed using the Min-Max Scaler function to ensure that all data have values ​​between 0 and 1 in order to make the values ​​of each attribute of the data within a certain range.

[0101] In addition, correlation analysis is performed between ship FGSS data features before training, and sub-grouping is performed on features with high Pearson correlation coefficients (pearson correlation coefficient 0.6 or higher).

[0102] This reduces dimensionality by grouping features, thereby reducing training time and resources, and allows all features to be trained equally without giving training weight to large-scale features.

[0103] Figure 6 is a schematic diagram showing an example of correlation analysis of ship FGSS data, and Table 1 shows the FGSS data sub-grouped according to the Pearson correlation coefficient.

[0104]

[0105] Figure 7 is a flowchart illustrating the learning process of a model according to the present invention.

[0106] Through the data preprocessing process, the sub-groups are classified into three, and training is performed on each group individually.

[0107] In this invention, since an LSTM-AE model is used that applies LSTM, which is a method of predicting the next state using information from the previous state at each time step, the shape of the input data is changed to three dimensions.

[0108] In addition, a sliding window is performed on the data to learn the temporal dependency, which is a characteristic of time series data.

[0109] If you perform a sliding window, you can learn the local temporal dependency of each window, and since the data was swung through the window, it can help extract data features.

[0110] After finishing the parameter settings, train the LSTM-AE model with the data.

[0111] In order to verify the results of an LSTM-AE model that performs anomaly detection by compressing normal data from an encoder to a latent layer and then restoring it through a decoder to compare the difference with the original, the present invention defines the data restoration value as an anomaly score through the reconstruction probability of the learned model.

[0112] In one embodiment, to identify anomaly data, a threshold of 0.213 is set as the point where the cumulative probability distribution of the loss function defined as the anomaly score is 99% or higher, and data points with an anomaly score higher than the threshold are determined to be anomaly data.

[0113] The specification of the threshold is not limited to this.

[0114] Figure 8 shows the distribution of the loss function of Sub-Group 1 through the training of the LSTM-AE (Long Short Term Memory AutoEncoder) model.

[0115] Through this, to identify anomaly data, the threshold was set to 0.213, which is the point where the cumulative probability distribution of the loss function is 99% or higher.

[0116] It is determined to be abnormal if the reconstruction error calculated using the test dataset exceeds the corresponding threshold.

[0117] Figure 9 is a graph of abnormal data detection in subgroup 1.

[0118] It can be seen that data randomly placed into the test set exceeds the threshold indicated by the red line, and more than 133 data points are detected starting from 2022-10-04 4:20. This means that failures can be predicted by detecting abnormal data.

[0119] The LSTM-AE-based ship FGSS failure prediction system and method according to the present invention described above utilize ship FGSS time series data to detect abnormal signals and apply them to an anomaly detection technique to enable the identification of the failure time and cause. It uses an LSTM-AE algorithm to define the range of normal data to diagnose the current state of the FGSS equipment, and includes noise signals in the learning process to correct for noise and loss caused by the ship's operating environment of the LiDAR, thereby enabling additional estimation of the state in a noisy environment.

[0120] As explained above, it will be understood that the present invention is implemented in a modified form without departing from the essential characteristics of the invention.

[0121] Therefore, the described embodiments should be considered in an illustrative rather than a limiting sense, and the scope of the invention is defined by the claims rather than the foregoing description, and all variations within the equivalent scope should be interpreted as being included in the invention. Explanation of the symbols

[0122] 10. Data Collection Department 20. Data Preprocessing Section 30. Model learning section 40. Equipment Status Information Provision Section

Claims

Claim 1 An LSTM AE-based shipboard FGSS fault prediction system comprising: a data collection unit for collecting sensor data output from FGSS equipment of a ship; a data preprocessing unit for removing data that has errors occurring during the data reception process and performing data normalization; a model learning unit for learning the definition of the normal state of sensor data installed in the FGSS equipment and the range of change due to noise through an LSTM-AE algorithm; and an equipment status information providing unit for providing information on the state of the equipment; wherein the model learning unit changes the shape of the input data to three dimensions to use the LSTM-AE model and performs a sliding window of the data to learn the temporal dependency, which is a characteristic of time-series data. Claim 2 A shipboard FGSS fault prediction system based on LSTM AE according to claim 1, wherein the AE (AutoEncoder) performing the LSTM-AE algorithm in the model learning unit is composed of an encoder and a decoder, with the number of nodes being the same at the input and output, and generates a reconstruction error by receiving data in the input layer, compressing information as the number decreases as it goes to the hidden layer, and restoring it back to the output layer, and uses the degree of the reconstruction error generated at this time as an anomaly score to determine whether the data is normal or abnormal by comparing it with a threshold. Claim 3 A shipboard FGSS fault prediction system based on LSTM AE according to claim 2, characterized in that the AE (AutoEncoder) does not require labels for normal and abnormal data, learning proceeds in a direction where the input data and output data are similar, the encoder stage extracts features of the input data and compresses them through a hidden layer to send to a latent layer, the latent layer prioritizes the most distinctive data features for reconstruction, and the decoder stage retrieves data from the latent layer and restores it to the original data. Claim 4 A shipboard FGSS fault prediction system based on LSTM AE according to claim 3, characterized in that data loss occurs during the process of compressing data in the encoder layer and sending data to the latent layer, resulting in a difference between the restored data and the original data, and after training is completed, the newly input data is restored in a direction similar to the existing training data, with data of features having lower priority than the original data disappearing. Claim 5 An LSTM AE-based FGSS fault prediction system for ships according to claim 1, characterized in that if the TIME_STAMP value in the FGSS data extracted from the data preprocessing unit is NaN or there are duplicate values, it is determined that an error occurred during the process of receiving data from the Vessel link and removed, and downsampling is performed to reduce the size of the data set. Claim 6 An LSTM AE-based ship FGSS fault prediction system according to claim 5, characterized by normalizing all data to have values ​​between 0 and 1 using a Min-Max Scaler function to make the values ​​of each attribute of the data within a certain range, and performing correlation analysis between ship FGSS data features prior to training to sub-group features with a Pearson correlation coefficient of 0.6 or higher. Claim 7 A shipboard FGSS fault prediction system based on LSTM AE according to claim 6, characterized by reducing training time and resources by reducing dimensionality through feature grouping, and enabling all features to be trained equally without placing training weight on large-scale features. Claim 8 delete Claim 9 An LSTM-AE-based shipboard FGSS fault prediction system according to claim 1, characterized by defining an anomaly score for the data restoration value through the reconstruction probability of the learned model to verify the results of an LSTM-AE model that performs anomaly detection by compressing normal data from an encoder to a latent layer, restoring it through a decoder, and comparing the difference with the original. Claim 10 A shipboard FGSS fault prediction system based on LSTM AE according to claim 9, characterized in that, to identify anomaly data, a point where the cumulative probability distribution of a loss function defined as an anomaly score is 99% or higher is designated as a threshold, and data points with an anomaly score higher than the threshold are determined to be anomaly data. Claim 11 A ship FGSS fault prediction method based on LSTM AE, comprising: a data collection step for collecting sensor data output from FGSS equipment of a ship; a data preprocessing step for removing data that has errors occurring during the data reception process and performing data normalization; a model training step for learning the definition of the normal state of sensor data installed in the FGSS equipment and the range of variation due to noise through an LSTM-AE algorithm; and an equipment status information provision step for providing information about the state of the equipment; wherein, in the model training step, the shape of the input data is changed to three dimensions to use the LSTM-AE model, and a sliding window of the data is performed to learn the temporal dependency, which is a characteristic of time series data. Claim 12 A shipboard FGSS fault prediction method based on LSTM AE according to claim 11, characterized in that, during the model training phase, data is received from an input layer, information is compressed as it moves toward a hidden layer and restored to an output layer to generate a reconstruction error, and the degree of the reconstruction error generated at this time is used as an anomaly score and compared with a threshold to determine whether the data is normal or abnormal. Claim 13 A ship FGSS fault prediction method based on LSTM AE according to claim 12, characterized in that it does not require labels for normal and abnormal data, learning proceeds in a direction similar to the input and output data, in the encoder stage extracts features of the input data and compresses them through a hidden layer to send to a latent layer, in the latent layer prioritizes the most distinctive data features for reconstruction, and in the decoder stage retrieves data from the latent layer and restores it to the original data. Claim 14 A ship FGSS fault prediction method based on LSTM AE according to claim 13, characterized in that data loss occurs during the process of compressing data in the encoder layer and sending data to the latent layer, resulting in a difference between the restored data and the original data, and after training is completed, the newly input data is restored in a direction similar to the existing training data, with data of features having lower priority than the original data disappearing. Claim 15 A ship FGSS fault prediction method based on LSTM AE according to claim 11, characterized in that, in the data preprocessing step, if the TIME_STAMP value in the extracted FGSS data is NaN or duplicate values ​​exist, it is determined that an error occurred during the process of receiving data from the Vessel link and removed; downsampling is performed to reduce the size of the dataset; normalization is performed using a Min-Max Scaler function to ensure that all data have values ​​between 0 and 1 to make the values ​​of each attribute of the data within a certain range; and correlation analysis is performed between ship FGSS data features before training, and sub-grouping is performed on features with a Pearson correlation coefficient of 0.6 or higher. Claim 16 delete Claim 17 A shipboard FGSS fault prediction method based on LSTM AE according to claim 11, characterized by defining the data restoration value as an Anomaly Score through the reconstruction probability of the learned model to verify the results of an LSTM-AE model that performs anomaly detection by compressing normal data from an encoder to a latent layer, restoring it through a decoder, and comparing the difference with the original. Claim 18 A shipboard FGSS failure prediction method based on LSTM AE according to claim 17, characterized in that, to identify anomaly data, a point where the cumulative probability distribution of a loss function defined as an anomaly score is 99% or higher is designated as a threshold, and data points with an anomaly score higher than the threshold are determined to be anomaly data.

Images