AI-Based Predictive Maintenance System for Water Treatment Process Facilities

Abstract

The present invention relates to a predictive maintenance system for predicting abnormalities in a microfilter and an ion exchange resin of a water treatment process facility, comprising: a data collection unit (10) that collects flow rate, pressure, differential pressure, resistivity, and water quality data in real time from a sensor; a data loading unit (20) that stores the collected time series data in an AI analysis database; a data cleaning unit (30) that performs missing value correction, spike removal, unit normalization, and time structure rearrangement on the stored data to make it suitable for learning; an AI analysis unit (40) that trains an LSTM-Autoencoder-based deep learning model using the cleaned normal time series data as input and calculates a restoration error for real-time input data; and an anomaly detection unit (50) that determines an equipment abnormality event when the restoration error or error index exceeds a preset threshold. The present invention provides a predictive maintenance system comprising a digital display unit (60) that visualizes the judgment result of the anomaly detection unit on a dashboard, and is characterized by predicting the performance degradation or replacement time of the microfilter and ion exchange resin through the restoration error calculated by the AI ​​analysis unit (40) and the threshold-based analysis of the anomaly detection unit (50).

Description

Technology Field

[0001] The present invention relates to a predictive maintenance system for water treatment process equipment comprising a data collection unit, a data loading unit, a data refinement unit, an LSTM-Autoencoder-based AI analysis unit, an anomaly detection unit that performs threshold-based status determination, and a visualization unit, in order to perform real-time status monitoring of microfilters and mixed-bed ion exchange resins (MB Polishers) used in water treatment processes, as well as AI-based anomaly detection and replacement cycle prediction. Background Technology

[0002] As illustrated in Fig. 1, in a conventional water treatment process (e.g., an ultrapure water manufacturing process), the microfilter at the beginning of the process, which is a pretreatment filter, and the mixed-bed ion exchange resin (MB Polisher) at the end perform the key water purification role, and if filter performance or ion removal ability decreases, the overall product quality may rapidly decline.

[0003] In existing field operations, workers have relied on manual equipment inspections at regular intervals or visual inspection of data changes; however, instantaneous data spikes and rapid performance deterioration within a short period are difficult to detect in a timely manner based solely on human experience.

[0004] In particular, microfilters have an average replacement cycle of 30 days and mixed-bed ion exchange resins have an average replacement cycle of 10 days, but the replacement time may be significantly advanced or delayed depending on actual process conditions (flow rate, increase in differential pressure, decrease in resistivity, etc.), which may lead to increased unnecessary maintenance costs or the possibility of quality defects.

[0005] To address these issues, it is necessary to integrate the real-time collection of quantitative sensor data, AI-based anomaly detection models, the application of optimized thresholds for each facility, digital twin-based visualization, and the incorporation of predictive maintenance results into operational scenarios.

[0006] However, existing water treatment processes have been unable to implement these functions due to limitations such as analog-based sensors, issues with equipment PLC integration, and the lack of data-based analysis for anomaly detection.

[0007] With the recent advancement of digital twin and AI-based manufacturing innovation technologies, LSTM-Autoencoder-based anomaly detection techniques based on time-series sensor data have begun to be utilized for equipment predictive maintenance, and methods that model normal patterns by learning process data such as filter pressure, differential pressure, flow rate, TOC, and resistivity, and detect anomalies using reconstruction error, are being put into practical use.

[0009] [Prior Art Literature]

[0010] [Patent Literature]

[0011] 1. Korean Patent Publication No. 10-2024-0015248 (February 5, 2024) The problem to be solved

[0012] The present invention relates to a predictive maintenance system configured to collect flow rate, pressure, differential pressure, resistivity, and water quality information in real time from sensors attached to microfilters and mixed-bed ion exchange resins in a water treatment manufacturing process, learn normal patterns in an LSTM-Autoencoder-based AI analysis unit through a data loading and purification process, and then transmit the results of a state judgment from an anomaly detection unit using restoration errors and thresholds derived from the learned model to a visualization unit, thereby enabling a field operator to check in advance whether there are abnormalities in the filters and ion exchange resins and when to replace them. means of solving the problem

[0013] The technical problem of the present invention as described above is achieved by the following means.

[0014] (1) A predictive maintenance system for predicting abnormalities in microfilters and mixed-bed ion exchange resins in a water treatment process,

[0015] A data collection unit that collects flow rate, pressure, differential pressure, resistivity, and water quality data from sensors in real time;

[0016] A data loading unit that stores the above-mentioned collected time series data in an AI analysis database;

[0017] A data cleaning unit that performs missing value correction, spike removal, unit normalization, and time structure reordering on the above-mentioned stored data to make it suitable for learning;

[0018] An AI analysis unit that trains an LSTM-Autoencoder-based deep learning model using refined normal time series data as input and calculates the restoration error for real-time input data;

[0019] An anomaly detection unit that determines an equipment abnormality event when the above restoration error or error index exceeds a preset threshold;

[0020] and includes a digital display unit that visualizes the judgment result of the above-mentioned anomaly detection unit on a dashboard,

[0021] A predictive maintenance system characterized by predicting the performance degradation or replacement time of a microfilter and a mixed-bed ion exchange resin through a restoration error calculated by the AI ​​analysis unit and a threshold-based analysis by the anomaly detection unit.

[0023] (2) In the above (1),

[0024] The above-mentioned data collection unit collects the pressure, differential pressure, and flow rate of the filter through pressure sensors and flow meters installed before and after the microfilter, and collects water quality data including resistivity and total organic carbon through resistivity meters and water quality sensors installed after the mixed-bed ion exchange resin, characterized by a predictive maintenance system.

[0026] (3) In the above (1),

[0027] The above data cleaning unit is characterized by performing interpolation to maintain constant time intervals of time series data, checking timestamp continuity, detecting and removing spikes, and automatically identifying normal sections in a predictive maintenance system.

[0029] (4) In the above (1),

[0030] The above AI analysis unit further includes a composite loss function application unit, wherein the above composite loss function application unit is characterized by simultaneously considering reconstruction error and remaining life prediction error to restore normal patterns well and accurately predict remaining life.

[0032] (5) In the above (4),

[0033] A predictive maintenance system characterized by a composite loss function represented by the following equation (1).

[0034] Equation (1):

[0035] Here, represents the proportion of reconstruction error, and represents the proportion of the remaining lifespan prediction error, and is reconstruction loss, represents the predicted lifespan loss.

[0037] (6) In the above (1),

[0038] A predictive maintenance system characterized in that the above AI analysis unit further includes a health index analysis unit, wherein the above health index analysis unit additionally applies a monotonic constraint to cause the health index (HI), which indicates the state of the microfilter and ion exchange resin, to change only in a monotonic decrease over time.

[0040] (7) In the above (6),

[0041] The above health index analysis unit calculates the health index of the microfilter from the following formula (2) A predictive maintenance system characterized by calculating ).

[0042] Equation (2):

[0043] Here, , represents differential pressure degradation and flow rate degradation, respectively, and are expressed by the following equations (2-1) and (2-2), respectively.

[0044] Equation (2-1): ( represents the minimum differential pressure in the normal section, and indicates a replacement recommendation or permissible limit differential pressure.)

[0045] Equation (2-2): ( represents the design standard or normal flow rate, and represents the minimum allowable flow rate under abnormal conditions.)

[0046] , is a weight that reflects the importance of the two signals, differential pressure and flow rate, respectively.

[0048] (8) In the above (6),

[0049] The above health index analysis unit [describes] the health index of the mixed-bed ion exchange resin from the following formula (3). A predictive maintenance system characterized by calculating ).

[0050] Equation (3)

[0051] Here, , These represent resistivity degradation and water quality degradation, respectively, and are expressed by the following equations (3-1) and (3-2), respectively.

[0052] Equation (3-1): ( represents the resistivity in a very clean state (maximum performance), and indicates the minimum allowable resistivity (replace if this is below).

[0053] Equation (3-2): ( is a water quality indicator such as TOC (lower is better), and represents the ideal minimum value, and indicates the maximum allowable value (anything above this is defective).

[0054] , These are weights that reflect the importance of the two signals, resistivity and water quality, respectively, and their relative size can be determined depending on the type of ion exchange resin.

[0056] (9) In the above (1),

[0057] A predictive maintenance system characterized by the threshold of the above-mentioned anomaly detection unit being set through the statistical distribution of normal data, cumulative error analysis, consultation with field experts, or ROC-based optimization techniques.

[0059] (10) In the above (1),

[0060] A predictive maintenance system characterized by the above-described digital display visually displaying real-time equipment status, abnormal events, restoration error trends, and replacement cycle prediction values ​​on a process flow diagram-based 2D or 3D interface. Effects of the invention

[0062] According to the present invention, it is possible to finely detect real-time state changes of microfilters and mixed-bed ion exchange resins, which are core equipment of a water treatment process, and there is an advantage of being able to determine the equipment status based on numerical values ​​by replacing the existing inspection method based on operator experience.

[0063] In particular, by detecting equipment performance degradation early, such as increased differential pressure, decreased flow rate, and reduced resistivity, excessive contamination or deterioration of the ion exchange resin can be predicted in advance, and the appropriate replacement time can be determined, thereby reducing unnecessary replacement costs.

[0064] In addition, by learning normal time series patterns based on an LSTM-Autoencoder model, it is possible to distinguish instantaneous outliers, sensor noise, or measurement errors in the data, and by combining this with threshold-based state judgment, the accuracy of anomaly detection is significantly improved.

[0065] The results of the AI ​​analysis are visualized in real-time on a dashboard, allowing operators to comprehensively check process status, equipment status, and performance degradation trends, and prevent production stoppages by identifying potential equipment failures in advance.

[0066] Therefore, the present invention contributes to improving the productivity of water treatment manufacturing plants, enhancing equipment maintenance efficiency, reducing quality defect rates, and improving energy and resource efficiency based on ESG. Brief explanation of the drawing

[0067] Figure 1 is a process flow diagram of an ultrapure water manufacturing process including a pretreatment filter (microfilter) and a mixed-bed ion exchange resin in a conventional water treatment process. Figure 2 is a configuration diagram of a water treatment manufacturing facility predictive maintenance system using AI according to the present invention. Figure 3 is a configuration diagram of an AI analysis unit according to the present invention. Figure 4 shows examples of equipment measurement information graphs and predictive maintenance outlier graphs using the predictive maintenance system according to the present invention. Specific details for implementing the invention

[0068] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described in detail below. However, the present invention is not limited to the embodiments disclosed below but can be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims.

[0069] In describing the embodiments of the present invention, specific descriptions of known functions or configurations will be omitted if it is determined that such detailed descriptions could unnecessarily obscure the essence of the invention. Furthermore, the terms described below are defined in consideration of their functions in the embodiments of the present invention, and these definitions may vary depending on the intentions or practices of the user or operator. Therefore, such definitions should be based on the content throughout this specification.

[0070] Hereinafter, the contents of the present invention will be explained in detail step by step with reference to FIGS. 2 to 4.

[0071] The water treatment process facility predictive maintenance system (100) using AI according to the present invention includes a data collection unit (10), a data loading unit (20), a data refinement unit (30), an AI analysis unit (40), an anomaly detection unit (50), and a digital display unit (60).

[0072] The data collection unit (10) performs the function of collecting flow rate, pressure, differential pressure, resistivity, and water quality data in real time from sensors installed in the microfilter and mixed-bed ion exchange resin equipment of the water treatment process (e.g., ultrapure water manufacturing process) and transmitting them to a subsequent data analysis structure.

[0073] The above data collection unit (10) collects basic data capable of determining whether the filter is contaminated through pressure sensors, differential pressure sensors, and flow meters installed at the front and rear ends of the microfilter. At the rear end of the mixed-bed ion exchange resin (MB Polisher), resistivity and water quality (TOC, etc.) data are measured to monitor changes in ion removal performance.

[0074] The collection cycle can preferably be set at 1-minute intervals, and the original data is transmitted as is, including the continuity of data timestamps, sensor spike detection results, etc.

[0075] Additionally, the data collection unit (10) can be linked with the equipment PLC or gateway, and the collected data is immediately transferred to the data loading unit (20).

[0076] The data loading unit (20) performs the function of loading original time series data transmitted from the data collection unit (10) into an AI analysis DB or cloud data storage (21), and is designed to allow the input sequence to be configured into a certain length (Time Window) for LSTM-Autoencoder training.

[0077] The above data loading unit (20) sequentially stores data such as flow rate, pressure, differential pressure, and resistivity collected from sensors, and provides a long-term time series data structure for learning an analysis model.

[0078] At this time, when loading in the data loading unit (20), the data structure can be sorted based on equipment classification, sensor type, unit, and collection time, and data integrity is verified, and damaged data is passed to the data cleaning unit (30).

[0079] The data cleaning unit (30) performs data cleansing, such as defining the data to be trained, correcting missing values, removing spikes, and converting units, and structures the data into an input tensor suitable for model training.

[0080] The above data cleaning unit (30) preferably performs interpolation to ensure consistency of time intervals for the collected time series data, and classifies outlier candidates through a spike detection algorithm to remove them from the training data or use them for separate outlier labeling.

[0081] In addition, only normal data is extracted and used to construct the training dataset for the LSTM-Autoencoder, which is a key element of the unsupervised learning process for learning normal measurement patterns.

[0082] The AI ​​analysis unit (40) preferably uses an LSTM-Autoencoder-based model to learn the characteristics of the normal equipment state and calculates the restoration error of the input data to derive a basis value for determining whether there is an abnormality.

[0083] The AI ​​analysis unit (40) according to the present invention further includes a regression model that directly predicts the remaining lifespan of a filter using a latent expression vector extracted from an input time series.

[0084] More specifically, the LSTM_Autoencoder receives time series data consisting of flow rate, pressure before and after the filter, differential pressure, resistivity, and water quality information, generates a latent vector that summarizes the facility status, and the latent vector is passed to a decoder to reconstruct the time series data, and at the same time, is passed to a remaining life prediction module composed of a separate fully connected neural network to calculate the remaining replacement days or remaining throughput of the filter.

[0085] During the training process, model parameters are optimized to achieve a balanced improvement in reconstruction performance and lifespan prediction performance by simultaneously considering a loss function that minimizes the reconstruction error between the input data and the reconstructed data, and a loss function that minimizes the difference between the predicted remaining lifespan and the remaining lifespan labels calculated from the actual replacement history.

[0086] In an embodiment of the present invention, the AI ​​analysis unit (40) includes a composite loss function application unit (41) as shown in FIG. 3.

[0087] The above composite loss function application unit (41) applies a composite loss function, preferably expressed by the following equation (1), by simultaneously considering the reconstruction error and the remaining lifespan prediction error in order to restore the normal pattern well and accurately predict the remaining lifespan.

[0088] Equation (1):

[0089] Here, represents the proportion of reconstruction error, and represents the proportion of the remaining lifespan prediction error, and It is calculated as the Mean Squared Error (MSE) as the reconstruction loss, and It can be calculated using the mean squared error or mean absolute error as the life prediction loss.

[0090] The data input to the AI ​​analysis unit (40) consists of multivariate time-series sensor data acquired in real time from the microfilter and mixed-bed ion exchange resin equipment of the water treatment process. The input data has the form of time-series blocks arranged continuously based on a certain time interval, and includes multiple sensor values ​​indicating the operating status of the filter or mixed-bed ion exchange resin at each time point.

[0091] Specifically, the input data at each point in time consists of the fluid flow rate measured by the flow sensor, the pressure values ​​and differential pressure measured by pressure sensors before and after the filter, the resistivity value measured after the mixed-bed ion exchange resin, and water quality information including Total Organic Carbon (TOC). These sensor values ​​are aligned at equal time intervals and grouped into time windows of a fixed length to form a single input sequence.

[0092] Such input sequences are also used as normal training data for learning time series patterns collected in a normal state, and new time series data flowing in in real time is applied to the trained model to become a subject of analysis for determining the equipment status. Therefore, the data input to the AI ​​analysis unit (40) is not a value at a single point in time, but has a structure that includes continuous changes over time, and is configured to be transmitted to the neural network model while maintaining these time series characteristics.

[0093] In the present invention, the AI ​​analysis unit (40) restores the normal operation pattern of the equipment through an LSTM-Autoencoder model based on input time series data and outputs various analysis results necessary to evaluate the condition of the equipment by comparing it with actual input values.

[0094] To this end, the AI ​​analysis unit (40) generates a result value that restores the input sequence, which is a value that is reconstructed by the model based on the sensor value at each point in time being judged as a "normal pattern." The restored time series is output in the same format as the input data and can be said to be the result of estimating the expected value during normal operation.

[0095] The AI ​​analysis unit (40) calculates a restoration error based on the difference between the input value and the restoration value. The above restoration error quantitatively represents the deviation from the normal state of the equipment and can be calculated in various forms, such as an error index or integrated error value summarizing the entire sequence, as well as errors at individual time points. This error index serves as a key indicator that quantitatively shows the possibility of performance degradation of the equipment.

[0096] In addition, the AI ​​analysis unit (40) outputs anomaly detection analysis results including the magnitude, trend, and rate of change of the restoration error to provide analysis information necessary for the subsequent anomaly detection unit (50). For example, pattern-based analysis results can also be produced, such as whether the restoration error gradually increases over time or whether a sudden anomaly appears at a specific point in time.

[0097] Consequently, the output data of the AI ​​analysis unit (40) has a complex form that includes not only simply restored time series values, but also restoration errors and error indices, abnormal pattern analysis results, etc., which are directly used to determine the equipment status. This output data is transmitted to the abnormality detection unit (50) and used to finally determine whether there is an abnormality in the equipment based on whether a threshold value is exceeded.

[0098] According to one embodiment of the present invention, in order to predict the replacement cycle of the microfilter and the mixed-bed ion exchange resin more practically and intuitively, the AI ​​analysis unit (40) may be configured not only to calculate the remaining useful life (RUL) as a single continuous value, but also to predict it in a stepwise form divided into several classes so that it can be directly utilized in the maintenance decision-making process of the equipment (hereinafter referred to as the stepwise RUL prediction module).

[0099] More specifically, the step-type RUL prediction module proposed in the present invention classifies and outputs the remaining life value into several stages, such as a “7 days or more” stage where the condition of the equipment is good and there is sufficient time until replacement, a “3 days or more but less than 7 days” stage where preparation for replacement is required, a “0 days or more but less than 3 days” stage indicating an imminent replacement, and a “0 days or less” stage where the equipment performance has already exceeded the normal allowable range.

[0100] This step-by-step RUL prediction method reflects the characteristics of filter equipment, where degradation progresses gradually, providing clear management signals that operators can immediately utilize when establishing equipment replacement plans. It also possesses structural advantages that allow for seamless integration with subsequent equipment management logic, such as warning importance, alarm intensity, and visual emphasis methods on digital displays.

[0101] Furthermore, this multi-stage prediction structure goes beyond simply predicting "how much remains" and clearly distinguishes and presents which management stage the equipment has currently passed, thereby significantly improving the reliability and practical applicability of equipment condition assessment.

[0102] In addition, according to another embodiment of the present invention, the AI ​​analysis unit (40) further includes a health index analysis unit (43). The health index analysis unit (43) may additionally apply a monotonic constraint so that the Health Index (HI), which indicates the condition of the equipment, changes only in a certain direction (monotonically decreasing or monotonically increasing) over time, in order to more accurately reflect the physical characteristics of the equipment, such as filters and ion exchange resins, in which the AI ​​model can more accurately reflect the physical characteristics of the equipment.

[0103] That is, since the health index of the filter has a relatively high value immediately after replacement and gradually decreases as operating time passes or usage accumulates, the health index analysis unit (43) imposes restrictions on the learning process so that unnatural and irregular fluctuations, such as the HI increasing and decreasing unreasonably over time, do not occur.

[0104] By applying such monotonous constraints, the health index output by the health index analysis unit (43) maintains a physically consistent pattern with the actual deterioration process of the equipment, and as a result, the stability of the remaining lifespan prediction, the reliability of the replacement time prediction, and the accuracy of establishing a long-term trend-based preventive maintenance strategy are significantly improved.

[0105] In addition, the health index analysis unit (43) according to the present invention, when combined with an anomaly detection result based on restoration error, exerts synergy to ensure that the health index has monotony, thereby providing an advantageous effect of constructing a sophisticated state determination logic that simultaneously reflects the degree of equipment deterioration and the intensity of abnormal signs.

[0106] In a preferred embodiment of the present invention, the health index of the microfilter ( ) can be obtained from the following equation (2).

[0107] Equation (2)

[0108] Here, , represents differential pressure degradation and flow rate degradation, respectively, and are expressed by the following equations (2-1) and (2-2), respectively.

[0109] Equation (2-1): ( represents the minimum differential pressure in the normal section, and indicates a replacement recommendation or permissible limit differential pressure.)

[0110] Equation (2-2): ( represents the design standard or normal flow rate, and represents the minimum allowable flow rate under abnormal conditions.)

[0111] , These are weights that reflect the importance of the two signals, differential pressure and flow rate, respectively, and their relative magnitudes can be determined depending on the type of microfilter.

[0112] In addition, as a preferred embodiment of the present invention, the health index of the mixed-bed ion exchange resin ( ) can be obtained from the following equation (3).

[0113] Equation (3)

[0114] Here, , These represent resistivity degradation and water quality degradation, respectively, and are expressed by the following equations (3-1) and (3-2), respectively.

[0115] Equation (3-1): ( represents the resistivity in a very clean state (maximum performance), and indicates the minimum allowable resistivity (replace if this is below).

[0116] Equation (3-2): ( is a water quality indicator such as TOC (lower is better), and represents the ideal minimum value, and indicates the maximum allowable value (anything above this is defective).

[0117] , These are weights that reflect the importance of the two signals, resistivity and water quality, respectively, and their relative size can be determined depending on the type of ion exchange resin.

[0118] The anomaly detection unit (50) determines in real time whether there is an anomaly in the equipment using the restoration error and indicator value input from the AI ​​analysis unit (40), performs a threshold-based status analysis, and then transmits an event signal to the digital display unit (60).

[0119] In the present invention, the digital display unit (60) includes any display means that visualizes the judgment result of the anomaly detection unit (50) on a dashboard using a predetermined GUI or digital twin, etc.

[0120] The data input to the anomaly detection unit (50) consists of analysis results calculated in real time by the AI ​​analysis unit (40), and these data are used as key indicators necessary to determine how much the equipment deviates from its normal operation pattern.

[0121] First, the anomaly detection unit (50) receives as input the restoration error value obtained by comparing the time series data restored by the AI ​​analysis unit (40) with the actual sensor value. The restoration error may be calculated for each point in time, or may be provided as a total error or average error in the form of a summary of the entire time series section, and is an indicator that most directly reflects signs of abnormality in equipment performance.

[0122] In addition, the AI ​​analysis unit (40) can further calculate an error index or error score, an error fluctuation trend value, and an error increase rate over time to more clearly express the condition of the equipment based on the restoration error, and these indicators are also provided as inputs to the anomaly detection unit (50). These indicators are useful not only for judging anomalies at a simple point-by-point level but also for judging the flow of performance degradation of the equipment.

[0123] In addition, a threshold value set by reflecting the characteristics of each piece of equipment and past failure patterns is also transmitted to or stored in the anomaly detection unit (50), and this is used as a judgment criterion to determine at what level the restoration error or error index should be considered an abnormal state.

[0124] Accordingly, the input data of the anomaly detection unit (50) is not a single value, but can be composed of a set of various types of analysis results including the magnitude of the restoration error, the temporal change pattern of the error, the standard threshold for each facility, and the error distribution characteristics. This input data is used to precisely determine whether the current state of the facility is within a normal range, or whether there are signs of an anomaly or a possibility of failure.

[0125] The anomaly detection unit (50) according to the present invention determines whether there is an anomaly in the equipment based on the result of comparing and analyzing the input restoration error and error index with the set threshold, and generates various forms of output data to transmit the result of this determination to the digital display unit (60) or the upper system.

[0126] The most basic output is a judgment result indicating whether an "anomaly has occurred" at a specific point in time or within a specific time window. This can be a binary result that designates an abnormal state if the restoration error exceeds a threshold and a normal state if it is below the threshold.

[0127] In addition, the anomaly detection unit (50) can generate an alert level to quantitatively indicate not only whether a simple anomaly has occurred but also the degree of the anomaly. For example, if the threshold is slightly exceeded, a level 1 alert is issued; if a moderate level of exceedance occurs, a level 2 alert is issued; and if a severe level of exceedance occurs, a level 3 alert is issued. By outputting alert levels classified by level, the facility manager can immediately recognize the severity of the situation.

[0128] In addition, the anomaly detection unit (50) can determine whether the restoration error is on an increasing trend, a temporary spike, or continuously rising, and output a state change pattern. The result of this pattern analysis provides important information for distinguishing whether the equipment's performance is fundamentally degrading or if it is a temporary phenomenon caused by a specific environmental change, going beyond simple immediate anomaly detection.

[0129] In addition, the anomaly detection unit (50) outputs standardized status data (e.g., JSON or structured message) including necessary status codes, error indices, abnormal section information per sensor, time of occurrence information, and equipment ID information to visually represent the equipment status on the digital display unit (60). This output data is designed to reflect the real-time status of the equipment on a dashboard and to be linked with an alarm system or maintenance management system to induce immediate action by the operator.

[0130] As such, the data output from the anomaly detection unit (50) consists of various information that can be immediately utilized by the equipment operator, such as quantitative indicators of the probability of failure, stages of anomaly occurrence, and trends in equipment status changes, beyond simple fault detection results, and this information serves as key judgment data for the predictive maintenance system.

[0131] As shown in FIG. 4, the digital display unit (60) visually displays the operating status of the equipment, abnormal events, performance degradation trends, etc., on a dashboard based on status data transmitted from the abnormal detection unit (50).

[0132] The data input to the digital display unit (60) may preferably include various forms of real-time information necessary to reproduce the actual operating state of the water treatment process equipment in a virtual space and to visually provide the abnormality detection results to the user.

[0133] To this end, the digital display unit (60) receives the equipment status judgment result transmitted from the anomaly detection unit (50) as input. The data includes analysis details such as whether an anomaly has occurred in the equipment, the grade of the anomaly warning, the time of occurrence of the anomaly, and the error index or the magnitude of the restoration error that served as the basis for the anomaly judgment, and serves as key data that comprehensively expresses the equipment status.

[0134] The digital display unit (60) utilizes raw sensor data as input, which is transmitted in real time from the data collection unit (10) or retrieved after being stored in the data loading unit (20). This sensor data includes values ​​that directly reflect equipment performance, such as flow rate, pressure, differential pressure, resistivity, and water quality (e.g., TOC), and is visualized on the monitoring screen in the form of real-time measurement graphs, gauges, color changes, etc.

[0135] The results of the analysis of restoration error and performance degradation trends calculated by the AI ​​analysis unit (40) are also provided as input to the digital display unit (60). This data is used to visually represent the degree of deviation from the normal operation pattern of the equipment, and through this, the operator can intuitively understand the long-term performance degradation trend of the equipment.

[0136] In addition, since the digital display unit (60) is preferably linked to a virtual model configured based on the physical layout structure of the equipment and the process flow diagram, metadata such as equipment ID, equipment location, connection pipeline information, and process step is also used as input. Through this metadata, a one-to-one correspondence between the actual equipment and the virtual equipment becomes possible, and it becomes possible to intuitively display where an abnormality occurred during the entire process.

[0137] In this way, the data input to the digital display unit (60) is not limited to simple sensor values ​​or abnormal results, but is composed of a complex real-time data set including dynamic changes of the equipment, prediction information of the analysis unit, process configuration information, etc., and this serves as a core basis for reproducing the process in a virtual environment.

[0138] The digital display unit (60) generates and outputs visualized information based on input equipment status information, sensor data, and anomaly detection results so that the operator can intuitively understand the status of the process and equipment. This output data has functions beyond simple screen display and is configured to directly support the operator's decision-making.

[0139] To this end, the digital display unit (60) outputs real-time status information indicating the current status of each piece of equipment on a virtual process screen. This indicates whether the equipment is operating normally, is in an abnormal warning stage, or is in a faulty state. In the case of a warning stage, it is intuitively displayed in the form of colors, flashing, icons, etc., so that the operator can immediately recognize the problem situation.

[0140] The digital display unit (60) displays changes in input sensor data as a time graph or outputs measurement data visualization information in the form of a gauge or progress bar that visualizes the range of measurement values. Through this, the operator can check at a glance whether the flow rate, pressure, differential pressure, resistivity, and TOC values ​​are within the normal range.

[0141] The digital display unit (60) outputs abnormal event notification information based on the warning grade and restoration error-based judgment results transmitted by the abnormal detection unit (50). The notification is expressed as a pop-up on the screen, a message box, or a visual highlighting effect of an equipment object, and may also be provided in the form of a sound or a push alarm linked to an external system if necessary.

[0142] The digital display unit (60) can output a visualization of the predictive maintenance analysis results, including the predicted replacement cycle value of the equipment, the performance degradation trend, and the remaining lifespan information of the equipment. This is a visual representation of the long-term data-based analysis results of the AI ​​analysis unit in the form of a visual graph or index, and is used by the equipment manager to directly establish replacement timing or maintenance plans.

[0143] In addition, the digital display unit (60) highlights the location of an abnormality in a specific piece of equipment on the overall process flow diagram and outputs process unit visualization information that allows the operator to immediately know which section of the process path the problem occurred in. Through this, the operator can quickly locate the problematic equipment and conduct an on-site inspection.

[0144] Ultimately, the output data of the digital display unit (60) consists of data that provides not just a simple screen display, but comprehensive and visual information necessary for equipment operation, such as real-time operating status, warnings of abnormal occurrences, changes in equipment performance, prediction of replacement timing, and location information based on the process flow table. This output data performs an important function of directly supporting the monitoring and judgment of process operation by processing the results of the abnormality detection unit (50) and the AI ​​analysis unit (40) into a user-friendly form.

[0146] The foregoing description is merely an illustrative explanation of the technical concept of the present invention, and those skilled in the art to which the present invention pertains will be able to make various modifications and variations within the scope of the essential characteristics of the present invention. Accordingly, the embodiments disclosed in the present invention are intended to explain, not limit, the technical concept of the present invention, and the scope of the technical concept of the present invention is not limited by these embodiments. The scope of protection of the present invention shall be interpreted by the claims below, and all technical concepts within an equivalent scope shall be interpreted as being included within the scope of rights of the present invention. Explanation of the symbols

[0147] 10: Data Collection Department 20: Data Loading Section 30: Data Cleansing Department 40: AI Analysis Department 50: Anomaly Detection Unit 60: Digital display unit

Claims

Claim 1 A predictive maintenance system for predicting abnormalities in microfilters and mixed-bed ion exchange resins of water treatment process equipment comprises: a data collection unit (10) that collects flow rate, pressure, differential pressure, resistivity, and water quality data in real time from sensors; a data loading unit (20) that stores the collected time-series data in an AI analysis database; a data purification unit (30) that performs missing value correction, spike removal, unit normalization, and time structure rearrangement on the stored data to make it suitable for learning; an AI analysis unit (40) that trains an LSTM-Autoencoder-based deep learning model using the purified normal time-series data as input and calculates a restoration error for real-time input data; an anomaly detection unit (50) that determines an equipment abnormality event when the restoration error or error index exceeds a preset threshold; and a digital display unit (60) that visualizes the determination result of the anomaly detection unit, wherein the performance degradation or replacement time of the microfilter and ion exchange resin is determined through the restoration error calculated by the AI ​​analysis unit (40) and the threshold-based analysis of the anomaly detection unit (50). A predictive maintenance system characterized in that the AI ​​analysis unit (40) includes a composite loss function application unit (41), wherein the composite loss function application unit (41) simultaneously considers reconstruction error and remaining lifespan prediction error to accurately predict remaining lifespan and restore normal patterns well. Claim 2 A predictive maintenance system according to claim 1, wherein the data collection unit (10) collects the pressure, differential pressure, and flow rate of the filter through pressure sensors and flow meters installed at the front and rear ends of the microfilter, and collects water quality data including resistivity and total organic carbon through resistivity meters and water quality sensors installed at the rear end of the mixed-bed ion exchange resin. Claim 3 In claim 1, the data refinement unit (30) is characterized by performing interpolation to maintain a constant time interval of time series data, checking timestamp continuity, detecting and removing spikes, and automatically identifying normal sections, in a predictive maintenance system. Claim 4 delete Claim 5 A predictive maintenance system characterized in that, in the first paragraph, the composite loss function is represented by the following equation (1). Equation (1): Here, represents the proportion of reconstruction error, and represents the proportion of the remaining lifespan prediction error, and is reconstruction loss, represents the predicted lifespan loss. Claim 6 A predictive maintenance system according to claim 1, wherein the AI ​​analysis unit further includes a health index analysis unit, and the health index analysis unit additionally applies a monotonic constraint such that the health index (HI), which indicates the state of the microfilter and the ion exchange resin, changes only in a monotonic decrease over time. Claim 7 In Clause 6, the above health index analysis unit [determines] the health index of the microfilter (from the following formula (2) A predictive maintenance system characterized by calculating ). Equation (2): Here, , represents differential pressure degradation and flow rate degradation, respectively, and are expressed by the following equations (2-1) and (2-2), respectively. Equation (2-1): ( represents the minimum differential pressure in the normal section, and represents the replacement recommendation or allowable limit differential pressure.) Equation (2-2): ( represents the design standard or normal flow rate, and represents the minimum allowable flow rate under abnormal conditions.) , is a weight that reflects the importance of the two signals, differential pressure and flow rate, respectively. Claim 8 delete Claim 9 A predictive maintenance system according to claim 1, characterized in that the threshold of the anomaly detection unit is set through the statistical distribution of normal data, cumulative error analysis, consultation with field experts, or ROC-based optimization techniques. Claim 10 A predictive maintenance system according to claim 1, characterized in that the digital display unit visually displays real-time equipment status, abnormal events, restoration error trends, and replacement cycle prediction values ​​on a process flow diagram-based 2D or 3D interface.

Images